UNIVERSITY OF CALIFORNIA

Los Angeles

Patterned Exceptions in Phonology

A dissertation submitted in partial satisfaction of the

requirements for the degree Doctor of Philosophy

in Linguistics

by

Kie Ross Zuraw

2000
© Copyright by

Kie Ross Zuraw

2000
The dissertation of Kie Ross Zuraw is approved

____________________________________
Donka Minkova

____________________________________
Carson Schütze

____________________________________
Bruce Hayes, Committee Co-chair

____________________________________
Donca Steriade, Committee Co-chair

University of California, Los Angeles

2000

ii
 TABLE OF CONTENTS

1. Introduction ................................................................................................................... 1
1.1. Lexical regularities................................................................................................... 1
1.1.1. Regularities within morphemes......................................................................... 1
1.1.1.1. Zimmer’s conundrum................................................................................. 2
1.1.2. Regularities within morphologically complex words ....................................... 5
1.1.3. Regularities across words.................................................................................. 6
1.2. Exceptions to lexical patterns................................................................................... 7
1.2.1. Regularities in a separate system: the Stochastic Constraint Model ................. 8
1.3. Preview of the proposal............................................................................................ 9
1.4. Tagalog................................................................................................................... 11
1.4.1. Phonology sketch ............................................................................................ 12
1.4.2. Notes on the data ............................................................................................. 14
1.5. Appendix: OT basics.............................................................................................. 15
2. The model as applied to nasal substitution ............................................................... 18
2.1. Chapter overview ................................................................................................... 18
2.2. Nasal Substitution .................................................................................................. 19
2.2.1. The phenomenon ............................................................................................. 19
2.2.2. Distribution of exceptions ............................................................................... 22
2.2.3. Productivity of nasal substitution.................................................................... 33
2.3. An experiment........................................................................................................ 36
2.3.1. Introduction ..................................................................................................... 36
2.3.2. Task I: productivity ......................................................................................... 36
2.3.2.1. Results of Task I....................................................................................... 39
2.3.3. Task II: acceptability....................................................................................... 42
2.3.3.1. Results of Task II ..................................................................................... 42
2.4. The grammar .......................................................................................................... 45
2.4.1. Desiderata for an analysis ............................................................................... 45
2.4.2. Paradigm Uniformity....................................................................................... 45
2.4.3. Input-Output Correspondence ......................................................................... 46
2.4.4. Listedness ........................................................................................................ 49
2.4.5. Constraints specific to nasal substitution ........................................................ 53
2.4.6. Summary of constraints................................................................................... 61
2.4.7. Stochastic constraint ranking .......................................................................... 64
2.5. Representations: encoding exceptionality.............................................................. 67
2.5.1. Substitution diacritics ...................................................................................... 68
2.5.2. Underspecification .......................................................................................... 69
2.5.3. Allomorph listing ............................................................................................ 71
2.6. The Learner ............................................................................................................ 72
2.7. The Speaker............................................................................................................ 78
2.7.1. Probability of a candidate’s being optimal...................................................... 78
2.7.2. Generating a listed form .................................................................................. 80

iii
 2.7.3. Generating a novel form.................................................................................. 84
2.8. The Listener ........................................................................................................... 85
2.8.1. Introduction ..................................................................................................... 85
2.8.2. Reconstructing the underlying form................................................................ 85
2.8.3. Acceptability judgments.................................................................................. 98
2.9. Chapter Summary................................................................................................. 101
2.10. Appendix: experimental stimuli......................................................................... 103
2.11. Appendix: Calculating probabilities of rankings ............................................... 105
2.11.1. Pairwise ranking requirements .................................................................... 105
2.11.2. Complex ranking requirements ................................................................... 108
2.12. Appendix: Sample calculation in Mathematica ................................................. 111
3. Simulating the adoption of a new word................................................................... 114
3.1. Chapter overview ................................................................................................. 114
3.2. Assimilated loanwords ......................................................................................... 114
3.3. Model of the speech community .......................................................................... 118
3.4. How the simulation works ................................................................................... 121
3.5. Simulation results................................................................................................. 124
3.6. Chapter summary ................................................................................................. 127
3.7. Appendix: Functions used in the simulation........................................................ 128
4. The model as applied to vowel height alternations ................................................ 129
4.1. Chapter overview ................................................................................................. 129
4.2. Vowel height in Tagalog...................................................................................... 130
4.3. Analysis of vowel lowering/raising ..................................................................... 135
4.4. Aggressive Reduplication .................................................................................... 139
4.4.1. Analysis ......................................................................................................... 146
4.5. Distribution of exceptions in the loanword vocabulary ....................................... 152
4.5.1. Aggressive Reduplication applied to the vowel raising ................................ 156
4.6. Similarity along other dimensions ....................................................................... 162
4.7. Representations .................................................................................................... 168
4.7.1. Separate entries for derivatives?.................................................................... 168
4.7.2. Environment-tagged allomorphs ................................................................... 169
4.8. Modeling raising .................................................................................................. 175
4.9. Learnability .......................................................................................................... 179
4.10. Chapter summary ............................................................................................... 182
4.11. Appendix: statistical significance of influences on raising................................ 183
5. Alternatives to Encoding Lexical Regularities in the Grammar .......................... 185
5.1. A separate module................................................................................................ 185
5.2. Associative memory............................................................................................. 187
5.3. The dual mechanism model ................................................................................. 190
5.3.1. Evidence for a qualitative difference between irregulars and regulars ......... 192
5.3.2. Why are regular pasts not listed? .................................................................. 195

iv
6. Summary .................................................................................................................... 198

References ...................................................................................................................... 200

TABLE OF EXHIBITS

(1) Under- and overspecification......................................................................................... 4

(2) Stress in English words with -ic .................................................................................... 5
(3) English present-past mappings ...................................................................................... 6
(4) Tagalog phoneme inventory ........................................................................................ 12
(5) Examples of Tagalog affixes ....................................................................................... 14
(6) Sample OT tableau ...................................................................................................... 17
(7) Nasal-substituting prefixes with various stems ........................................................... 20
(8) Rates of nasal substitution for entire lexicon............................................................... 23
(9) Rates of substitution for various prefixes .................................................................... 25
(10) Voicing and nasal substitution: observed frequencies............................................... 28
(11) Voicing and nasal substitution: expected frequencies............................................... 28
(12) Place of articulation and nasal substitution: observed frequencies ........................... 30
(13) Place of articulation and nasal substitution: expected frequencies............................ 30
(14) Place of articulation and nasal substitution: (observed-expected)/expected values .. 31
(15) Pairwise differences in rate of substitution................................................................ 32
(16) Differing behavior among derivatives of the same stem........................................... 33
(17) Semantic unpredictability with nasal-substituting affixes......................................... 34
(18) Unpredictable stress/length shifts associated with nasal-substituting affixes ........... 34
(19) Personal characteristics of experiment participants................................................... 36
(20) Sample card for Task I............................................................................................... 37
(21) Sample sentence pair for Task I ................................................................................ 38
(22) Rates of substitution on novel words......................................................................... 40
(23) Overall rates of substitution on novel words, broken down by participant............... 41
(24) Example stimuli for Task II....................................................................................... 42
(25) Acceptability judgments: substituted - unsubstituted; error bars indicate 95%
confidence interval .................................................................................................... 43
(26) Nasal substitution as coalescence .............................................................................. 46
(27) Constraints against coalescence................................................................................. 48
(28) Corr-IO constraints: sample violations...................................................................... 49
(29) USELISTED ................................................................................................................. 50
(30) Violations of USELISTED ........................................................................................... 51
(31) Interaction of a family of USEX%LISTED constraints and Paradigm Uniformity ..... 52
(32) Interaction of a unitary USELISTED constraint and Paradigm Uniformity................. 53
(33) NASSUB ...................................................................................................................... 54
(34) *NC8 ............................................................................................................................ 55
(35) Coalescence within vs. across listed items ................................................................ 56

v
(36) Distribution of consonants in roots of the form C1V(C2)C3V(C4) ............................ 58
(37) *[N, *[n, *[m .............................................................................................................. 59
(38) *[N >> *[n >> *[m ..................................................................................................... 60
(39) Constraints affecting nasal substitution..................................................................... 62
(40) Input-Output Correspondence requires use of listed form ........................................ 62
(41) Coining of novel words, using the ranking in (40).................................................... 63
(42) Hypothetical constraint system.................................................................................. 65
(43) Sample lexical entry for stem-listing approach (cf. (16)).......................................... 67
(44) Partial lexical entries for underspecification approach.............................................. 69
(45) Learning, starting with two equally-ranked constraints ............................................ 72
(46) Mini-lexicon for learning........................................................................................... 74
(47) Sample learning trial.................................................................................................. 74
(48) Ranking values arrived at by Gradual Learning Algorithm ...................................... 76
(49) Availability as a function of listedness...................................................................... 78
(50) Hypothetical tableau .................................................................................................. 79
(51) Ranking requirements for candidate a in (50) to be optimal ..................................... 79
(52) Simple hypothetical tableau....................................................................................... 79
(53) Probability of Ci's outranking Cj in a given utterance ............................................... 80
(54) Four candidates for a listed, substituted word ........................................................... 81
(55) Candidate probabilities if /mampupuntol/ exists ....................................................... 83
(56) P(input|output) for various stem-initial obstruents .................................................... 83
(57) Probabilities of outcomes when no listed form exists ............................................... 84
(58) Choosing the optimal input ....................................................................................... 85
(59) Three possibilities on hearing [mamumuntol]........................................................... 87
(60) Bayesian inversion of probabilities compared by listener......................................... 87
(61) P(/maN/+/RCV/+/puntol/) = 1/((1+e-3+6*Listedness(whole word))*(1+e3-6*Productivity(maN+Rcv))) 89
(62) Idiosyncrasies in maN-Rcv- words ............................................................................. 91
(63) Prior probabilities of /mamumuntol/ and /mampupuntol/ ......................................... 93
(64) Calculating (60) when listener has no listed form..................................................... 94
(65) Prior probability of the output ................................................................................... 94
(66) Final result for (64).................................................................................................... 95
(67) Determining the input given output [mampupuntol]................................................. 95
(68) Probability of listener’s guessing that speaker used a listed word: substituted -
unsubstituted.............................................................................................................. 98
(69) P([mamumuntol]) ...................................................................................................... 99
(70) Predicted acceptability of substituted vs. unsubstituted for novel words................ 100
(71) Predicted and experimental acceptability values (substituted - unsubstituted) ....... 101
(72) Novel stimulus stems............................................................................................... 103
(73) Real-word stimulus stems........................................................................................ 104
(74) Mi - Mj ..................................................................................................................... 105
(75) Arriving at a selection point for a constraint in a given utterance........................... 106
(76) Calculating P(Ci>>Cj) ............................................................................................. 106
(77) P(Ci >> Cj) = P(Mi - Mj > -0.5) = .64 ...................................................................... 108

vi
(78) Pairwise rankings are not independent .................................................................... 109
(79) Possible total rankings of three constraints ............................................................. 109
(80) Substitution rates for Spanish stems, all affixal patterns combined ........................ 115
(81) Listener’s procedure for estimating P(output|inputi) ............................................... 122
(82) Estimating P(inputi|output) ...................................................................................... 123
(83) Simulation results for novel words after 150 “years”.............................................. 124
(84) Nasal substitution in real Spanish loans .................................................................. 125
(85) Nasal substitution in entire Tagalog lexicon ........................................................... 125
(86) Hand-crafted grammar to produce the desired results for /b/-initial stems ............. 126
(87) Simulation results using the handcrafted grammar in (86) ..................................... 126
(88) Deciding whether to pay attention to a speaker....................................................... 128
(89) Prior probabilities of inputs (see §2.8.2) ................................................................. 128
(90) Updating listedness.................................................................................................. 128
(91) Distribution of mid and high vowels ....................................................................... 130
(92) Suffixation-induced alternations.............................................................................. 131
(93) Vowel coalescence .................................................................................................. 131
(94) Transglottal vowels.................................................................................................. 132
(95) Exceptional native words......................................................................................... 133
(96) *NONFINALMID ....................................................................................................... 135
(97) *FINAL[u]................................................................................................................. 135
(98) Tableaux illustrating underspecification analysis.................................................... 137
(99) Similarity enhancement in English.......................................................................... 139
(100) Pseudoreduplicated words in Tagalog................................................................... 140
(101) Over- and underapplication in pseudoreduplicated roots ...................................... 142
(102) Overapplication of nasal substitution .................................................................... 143
(103) REDUP .................................................................................................................... 146
(104) Violations of REDUP for a 3-syllable input............................................................ 147
(105) Factorial typology of REDUP, CORR-IO, and CORR-BR ........................................ 148
(106) Loanword stems with nonfinal mid vowels and final [u]...................................... 152
(107) Alternation in loanword stems............................................................................... 152
(108) Effect of mid vowel in penult on probability of raising ........................................ 153
(109) Vowel harmony as a mechanism for preventing alternation ................................. 155
(110) Effect of matching backness between penult and ultima, given a mid penult....... 155
(111) Effect of proximity ................................................................................................ 156
(112) Aggressive reduplication blocks vowel raising ..................................................... 157
(113) A ranking that prevents correspondence between mismatched vowels ................ 158
(114) Vowel non-raising in CV reduplication................................................................. 160
(115) Effect of onset place of articulation on rate of raising .......................................... 162
(116) Effect of onset manner on rate of raising .............................................................. 163
(117) Effect of onset voicing on rate of raising .............................................................. 164
(118) Effect of onset shape on rate of raising ................................................................. 164
(119) Effect of rhyme shape on rate of raising................................................................ 165
(120) Effect of vowel length on raising .......................................................................... 166
(121) Effect of number of shared properties on raising .................................................. 167

vii
(122) Syncope ................................................................................................................. 170
(123) Suffixal allomorphs—sample partial lexical entries ............................................. 170
(124) MATCHCONTEXT .................................................................................................... 171
(125) Faithful use of suffixal allomorphs........................................................................ 171
(126) Variability for constructed suffixal allomorphs..................................................... 173
(127) Vowel height in a novel word: identical syllables................................................. 176
(128) Vowel height in a novel word: similar syllables ................................................... 176
(129) Vowel height in a novel word: dissimilar syllables............................................... 177
(130) Grammar used in simulation.................................................................................. 181
(131) Rate of raising in novel words with mid penults, using the grammar in (130) ..... 181
(132) Raising and mid vowel in the penult: observed frequencies ................................. 183
(133) Raising and mid vowel in the penult: expected frequencies ................................. 183
(134) Statistical significance of various inhibitors of vowel raising............................... 184

viii
 ACKNOWLEDGMENTS

First I have to thank Bruce Hayes and Donca Steriade, my advisers. They’ve been the

best thing (among many very fine things) about my six years here at UCLA. Just about

everything I know, they taught me, and certainly any good idea I’ve ever had has come

from a conversation with one of them. And without their encouragement to follow my

nose, this would have been a much tamer dissertation, for better or for worse. I hope

Bruce and Donca realize that they’re my role models in all areas of life. I doubt I can ever
live up to the example they’ve set as scholars, teachers, and mentors, but at least I know

what to shoot for.

I asked Carson Schütze to be on my committee because I knew he’d ask hard

questions, and he did. I know I haven’t really answered all of them, but trying to answer

them has clarified my thinking a lot, and I think made this a better work. I asked Donka

Minkova to be on my committee because of its diachronic flavor; thanks to her for giving

me an informed perspective on my claims about lexical change and the role of

interactions in the speech community.

What first attracted me to UCLA was the sense of excitement among the students

and faculty over what they were doing. Sustaining that environment requires time and

effort, and I’d like to thank all of the students and faculty in this department for the

interaction that has been invaluable to me. Particularly to be thanked are Adam Albright,

Dan Albro, Marco Baroni, Roger Billerey, Katherine Crosswhite, Janine Ekulona,

Christina Foreman, Matt Gordon, Bruce Hayes, Sun-Ah Jun, Pat Keating, Ed Keenan,

Robert Kirchner, Peggy MacEachern, Pam Munro, Carson Schütze, Ed Stabler, Donca

Steriade, Siri Tuttle, and Jie Zhang.

ix
 This dissertation draws mainly on data from Tagalog. I was first introduced to the

language in a field methods course at McGill taught by Lisa Travis, with Natividad del

Pilar as language consultant. Tania Azores-Gunter was my Tagalog teacher for two years

at UCLA, and I’m grateful to her for her patience and encouragement. Thanks to all the

Tagalog speakers who shared their knowledge of the language with me, especially Nenita

Pambid-Domingo and Angel Camandang.

Over the years I’ve been fortunate to have many outstanding teachers. I would

have been lucky to study with just one of them; to have so many to thank is a wonder.
From my years at FACE, I should Frank Cottam for introducing me to linguistics and

Iwan Edwards for the lesson that what’s worth doing is worth doing well. At the

University of Illinois Laboratory High School, David Bergandine, Chris Butler, Mort

Castle (who didn’t work at Uni, but taught me while I was there), Sandra Dawson,

Elizabeth Jokusch, Peter Kimball, Rosemary Laughlin, Pat McLaughlin, Rick Murphy,

Frances and John Newman, Bernard Norcott, Al Smith, David Stone, and Joanne

Wheeler provided an atmosphere of constant intellectual stimulation, encouraged

thorough and systematic thought, and generally made us feel that anything was possible.

At McGill, Heather Goad, Myrna Gopnik, and Glynne Piggott helped me begin to

become a linguist by treating me as though I already was one.

Finishing graduate school isn’t the hardest thing in the world, but it takes its toll.

Katherine Crosswhite, Leah Gordon, and Peggy MacEachern provided friendship that I

couldn’t have done without, as did Linda and Phil Ross, my first teachers and most loyal

supporters. And anyone who knows me knows that I would never have made it through

without the aid and comfort of Bryan Zuraw. He gave me constant encouragement,

tolerated my mood swings and mounting paranoia, and, as the pace of work on this

x
document accelerated, took over all my non-dissertation responsibilities and made sure I

had clean clothes every morning and a hot meal every night.

xi
 VITA

March 29, 1973 Born, Montreal, Quebec, Canada

1990-1994 James McGill Entrance Scholarship

McGill University
Montreal, Quebec, Canada

1992 Sarah Rosenfeld Prize in Yiddish

McGill University

1993 Betty Workman Yaffe Prize in Yiddish

McGill University

1994 Undergraduate Research Assistant

Familial Language Impairment Project
McGill University

1994 B. A. with First Class Honours, Linguistics

McGill University

1994-1996 Bourse de maîtrise en recherche

Fonds pour la Formation de Chercheurs et l’Aide à
la Recherche

1994-1997 National Science Foundation Graduate Fellowship

1995-1999 Teaching Assistant, Associate, Fellow

Department of Linguistics
University of California, Los Angeles

1996 M. A., Linguistics

University of California, Los Angeles
Los Angeles, California

1996-1997 Phonetics Laboratory Computer Assistant

Department of Linguistics
University of California, Los Angeles

1997-1998 Teaching Assistant Consultant

Department of Linguistics
University of California, Los Angeles

1998 Instructional Software Programmer

Department of Linguistics
University of California, Los Angeles

1999 Dissertation Year Fellowship

University of California, Los Angeles

xii
 PUBLICATIONS AND PRESENTATIONS

Zuraw, Kie (April 1996). Moving Phonotactics: Variability in Infixation and

Reduplication of Tagalog Loanwords. Paper presented at the Third Annual
Meeting of the Austronesian Formal Linguistics Association, Los Angeles,
California.

Zuraw, Kie (April 1998). Tagalog Nasal Substitution: Allomorphic Emergence of the Un-
marked. Paper presented at the Southwest Workshop on Optimality Theory,
Tucson, Arizona.

Zuraw, Kie (January 1999). Knowledge of Lexical Regularities: Evidence from Tagalog
Nasal Substitution. Paper presented at the Annual Meeting of the Linguistic
Society of America, Los Angeles, California.

Zuraw, Kie (June 1999). Regularities in the Polymorphemic Lexicon. Invited paper
presented at the Workshop on the Lexicon in Phonetics and Phonology,
University of Alberta, Edmonton, Alberta, Canada. To appear in proceedings.

Zuraw, Kie (January 2000). Aggressive Reduplication in Tagalog. Paper presented at the
Annual Meeting of the Linguistic Society of America, Chicago, Illinois.

Zuraw, Kie (February-March 2000). Patterned Exceptions in Phonology. Invited

colloquium presented at the University of California, Los Angeles; the University
of Southern California; and the University of California, Irvine.

Zuraw, Kie (to appear). Regularities in the Polymorphemic Lexicon. University of

Alberta Papers in Experimental and Theoretical Linguistics 8.

xiii
 ABSTRACT OF THE DISSERTATION

Patterned Exceptions in Phonology

by

Kie Ross Zuraw

Doctor of Philosophy in Linguistics

University of California, Los Angeles, 2000

Professor Bruce Hayes, Co-chair

Professor Donca Steriade, Co-chair

Standard Optimality-Theoretic grammars contain only the information necessary

to transform inputs into outputs; regularities among inputs are not accounted for. Using

the example of Tagalog nasal substitution, this dissertation presents a model of how

lexical regularities could be learned, represented in the grammar, used by speakers and

listeners, and perpetuated over time.

Lexical regularities are represented as low-ranking constraints, their rankings

learned through exposure to the lexicon using Boersma’s Gradual Learning Algorithm.

High-ranked constraints ensure the primacy of listed pronunciations; but when a speaker

produces a novel word, these high-ranking constraints are irrelevant and the constraints

that encode lexical regularities take over. The subterranean constraints are stochastically

ranked; speakers’ behavior on novel words probabilistically reflect the lexical

regularities. The listener uses the same grammar to produce well-formedness judgments

for novel words and to reconstruct inputs from an interlocutors’ outputs. The model’s

xiv
well-formedness judgments reproduce the experimental result that although the

productivity of nasal substitution on novel words is low, nasal-substituted novel words

are judged more acceptable than non-substituted words in certain cases.

Bayesian reasoning by the listener favors novel nasal-substituted words—they are

disproportionately likely to become listed. A computer simulation of the speech

community confirms that although nasal substitution is the minority pronunciation for

novel words, a word may eventually enter the lexicon as nasal-substituted.

Tagalog vowel raising under suffixation is close to exceptionless in the native

vocabulary but quite exceptionful among loanwords. A loan stem’s probability of

resisting raising is highly influenced by its degree of internal similarity. I propose that

internal similarity encourages speakers to construe a word as reduplicated, even without

morphosyntactic motivation; raising is blocked because it would disrupt base-reduplicant

identity.

Alternatives to encoding lexical regularities in the grammar are considered. It is

argued that the vowel raising facts are not amenable to an associative memory account.

The qualitative difference between “regulars” and “exceptions” cited by proponents of

the Dual-Mechanism model as evidence for leaving lexical regularities out of the

grammar reduces to a difference between listed words and synthesized words; this

difference can arise through listener reasoning, without a prior qualitative difference.

xv
xvi
 1. Introduction
This dissertation presents a model of how phonological patterns in the lexicon could be

learned and used by speakers and hearers, and perpetuated over time. This chapter

introduces the phenomenon of lexical patterns, discusses why they are problematic in

current phonological thinking, and gives a preview of the model.

1.1. Lexical regularities

I will use the terms lexical regularity and phonological pattern to refer to generalizations

about the phonological properties of the set of words in a language. Regularities can be

observed that apply within morphemes, within morphologically complex words, and

across sets of words.

1.1.1. Regularities within morphemes

In English roots of the form sCVC, the two Cs generally cannot be both labial, both velar,

both nasal, or both [l].1 The generalization is quite strong (see Berkley 1994 for statistical
findings on this and related phenomena in the English lexicon), and hypothetical

exceptions, though pronounceable, sound somewhat ill-formed (?[slIl], ?[skQN]).2

Generalizations like this one are often attributed to morpheme structure constraints

1
Although such sequences are common across word or morpheme boundaries: It’s Lily! or Ask Angry Joe.
2
A search of the online Oxford English Dictionary for sCVC words only (i.e., not the full set of
sC(C)VC(C) words, which follow similar restrictions) found, collapsing variant spellings and
pronunciations, just 3 words with two labials (Spam, spume, spoom), 9 words with two velars (skoke, skeck,
skowke, skeg, skig, scak, scoke, scag, scug), 3 words with two nasals (smon, snam, snum), and no words
with two ls. Most of these words were unfamiliar to me.

1
(introduced by Halle 1959 as “morpheme structure rules”)3—language-specific

conditions that rule out some set of possible morphemes as ill formed.

Morpheme structure constraints are static in the sense that they can be observed

only as a property of existing words; they do not drive alternations. Although slill sounds

strange, it is pronounceable and does not require any “repair”.

Morpheme structure constraints are rarely exceptionless. For example, English

words like [spΘm] ‘Spam (brand name of processed meat product)’ and [skEg] ‘skeg (oat

species; part of ship’s keel; fin of surfboard; plum species; nail; stump of a branch; tear in

cloth)’ violate the sCVC restriction described above. There needs to be some mechanism

that allows these words to escape the constraint.

1.1.1.1. Zimmer’s conundrum

What is the role of morpheme structure constraints in the grammar, since they do not

drive alternations? In Optimality Theory (OT; see §1.5), often include a proof that the

correct surface forms result no matter what the input (Richness of the Base: Prince &

Smolensky 1993, Smolensky 1996a). For example, if a language lacks morphemes of the

form CiVCi, the analysis includes a demonstration that the input /pop/ is repaired to (say)
[pot]. A problem with this type of demonstration, of course, is that the analyst generally

does not know what the correct surface form for the input /pop/ should be ([pok], [kop],

[po]...)—it might even be [pop].

In the case of morpheme structure constraints at least, it is doubtful that such

proofs are necessary, because the learner has no reason to posit underlying forms that are

significantly different from the surface forms. For example, by Lexicon Optimization

(Prince & Smolensky 1993; Itô, Mester, & Padgett 1995), the learner would construct the

3
although root structure constraint would be more apt in most cases.

2
underlying form /pok/ for a morpheme that is always pronounced [pok]; similarly, she

would construct /kop/ for [kop], and so on. If she never hears [pop], she will not

construct /pop/, and so there is no need for the grammar to repair /pop/, because no such

lexical entries exist. If the constraint against morphemes of the form CiVCi plays no role

except to repair inputs that may not exist anyway, then perhaps it does not belong in the

grammar.

Inkelas, Orgun, and Zoll 1997 make a similar argument for Labial Attraction, a

constraint on vowels in Turkish roots.4 Inkelas et al. propose a overspecification as a

mechanism for tagging words as exceptions to constraints. Nonexceptional segments in

morphemes are underspecified, and their feature values can be filled in by markedness

constraints at no faithfulness cost. In different morphological contexts, different values

will be filled in, resulting in alternation. Exceptional segments, on the other hand, are

fully specified, and high-ranked faithfulness constraints prevent tampering with those

underlying specifications. The tableau in (1) illustrates the analysis for Turkish final

devoicing: underspecified /kitaB/ (B stands for a bilabial stop unspecified for voicing)

undergoes final devoicing, but overspecified /etyd/ does not.

4
Labial Attraction is a systematic exception to Round Harmony: normally, a high vowel must agree in
[round] with a preceding vowel (e.g., *Atu), but if the preceding vowel is [A] and the intervening consonant
is labial, then a high, back vowel will be [+round] instead of [-round] as expected. Round Harmony drives
alternations, applying across a suffix boundary, but Labial Attraction holds only within morphemes (and
even within morphemes, there are exceptions).

3
(1) Under- and overspecification
/kitaB/+/a/ IDENT-IO[HIGH] C/__# = [-VOICE] C =[+VOICE]
‘book-dative’
a . kitaba *!
b kitapa
/kitaB/ IDENT-IO[HIGH] C/__# = [-VOICE] C =[+VOICE]
‘book-nominative’
c kitab *!
d . kitap *
/etyd/ IDENT-IO[HIGH] C/__# = [-VOICE] C =[+VOICE]
‘etude’
e . etyd *
f etyt *! *

Inkelas et al. conclude, however, that for a static pattern such as Labial Attraction,

special tagging is not necessary. Without alternations, nothing drives the learner to

construct underspecified lexical entries. Therefore, faithfulness constraints do all the

“work”, and there is no role in the grammar for constraints like Labial Attraction.

Zimmer (1969) attempted to find psychological evidence for Labial Attraction

and two other Turkish morpheme structure constraints, and found that many speakers had

internalized a different version of Labial Attraction than the one linguists had

formulated.5 Zimmer speculates on why this should be so:

The question of course arises as to how speakers of a language can get away with
such erroneous notions [the “wrong” version of Labial Attraction]. This, however,
is not really very mysterious. The mistaken generalizations we have attributed to
speakers of Turkish do not involve productive phonological rules. Both groups
presumably learn lexical items in their fully specified form and then simply repeat
them; the MSC’s [morpheme structure constraints] in question do not fill in
values for incompletely specified segments. […] Since these generalizations [that
speakers make about vowel cooccurrences], and those made in this area by other
speakers, have no observable consequences in the course of the normal use of the
language, they are not subject to correction in the same way in which a wrongly
learned productive rule would be.

5
The linguists’ constraint: [u] is required after [A] followed by a labial consonant. The constraint exhibited
by some of the speakers: [u] is required after [A] followed by any consonant.

4
 The conundrum is, if Labial Attraction does no “work” in the grammar of

Turkish, why had speakers internalized any version of it at all?

1.1.2. Regularities within morphologically complex words

Regularities are also to be found in morphologically complex words. For example,

English words suffixed with -ic generally have penultimate stress, regardless of the stress

pattern in the base.

(2) Stress in English words with -ic

artíst-ic cf. ártist
laparoscóp-ic cf. láparoscope
cholerá-ic cf. chólera

There are a few exceptions to this generalization, such as chóler-ic (cf. chóler) and Árab-

ic (cf. Árab).

Regularities in polymorphemic words are “productive” in the sense that if a

speaker knows only the related base, it is up to her to create a word that follows or does

not follow the generalization. (By contrast, if a speaker knows the word slill, she has no

choice but to pronounce it slill.) For example, should the -ic form of carob be carób-ic or

cárob-ic (or something else)? Compared to morpheme structure constraints, regularities

in polymorphemic words thus have more opportunity to make themselves felt in the

language, as new affixed forms are coined much more frequently than new morphemes.

Regularities in morphologically complex words might seem at first glance to

naturally belong in the grammar (and so Zimmer’s conundrum would not arise), but when

there is evidence that the words are listed as separate lexical entries (see §2.2.3), the

situation is the same as with morpheme structure constraints: speakers would not need to

5
learn the regularity in order to produce existing words correctly. But if speakers do apply

the regularity to novel affixed words, this fact must be accounted for somehow.

1.1.3. Regularities across words

Regularities also exist in the mappings among related words. For example, many English

verb roots ending in [… IN(C)] form their past tense by changing [I] to [Q], although

there are several competing patterns:

(3) English present-past mappings

present past
sing sang
ring rang
sink sank
drink drank

but

fling flung
bring brought
blink blinked

This is not a generalization about the shape of past-tense forms, but rather a

generalization about the mappings between present- and past-tense forms. Like

regularities within morphologically complex words, regularities in the mappings between

words have the property of productivity: when a speaker forms the past tense of novel

spling, for example, she must decide whether it should be splang, splung, splinged, or

perhaps something else.6 Thus, mapping regularities also have opportunity to make their

presence known. And like regularities in morphologically complex words, regularities in

mappings do not need to be learned in order to produce existing words correctly.

6
Bybee & Moder 1983 performed an experiment that required speakers to do just this task. See §5.3.1 for a
discussion.

6
 1.2. Exceptions to lexical patterns

It was mentioned above that lexical regularities tend to have exceptions (Spam, Árabic,

blinked), but the distribution of exceptions often is not random. In the two cases

discussed in this dissertation (Chapters 2 and 4), the exceptions themselves are highly

patterned: although it is not predictable whether any given word will be an exception,

words with certain phonological properties are more likely than others to be exceptions.

There are not enough exceptions to the sCVC morpheme structure constraint or to the

generalization that -ic carries penultimate stress to look for patterns within the
exceptions, but we can see many such patterns in English past tense. For example, a verb

is more likely to follow the [I]-[Q] mapping if it has a velar nasal in the coda than if it has

an alveolar or bilabial nasal (begin, began; swim, swam) (see Bybee & Slobin 1992 for a

discussion of regularities in the distribution of English past-tense mappings).

Frisch, Broe, and Pierrehumbert (1996), expanding on Pierrehumbert 1993,

examined the distribution of exceptions to an Arabic morpheme structure constraint that

forbids consonants of the same place of articulation within a root. They showed that far

from being random, exceptions to the constraint are distributed such that the more similar

two consonants are, the less likely they are to cooccur. For example, /t...d...k/ and

/t...z...k/ both violate the constraint against homorganic consonants within a root, but

because t and d are more similar than t and z (they share membership in more natural

classes), roots of the form /t...d...X/ are more common than roots of the form /t...z...X/.

Frisch et al.’s account of the Arabic facts is discussed in the following section. See Frisch

and Zawaydeh (to appear) for evidence on the psychological reality of this constraint.

7
1.2.1. Regularities in a separate system: the Stochastic Constraint Model

The Stochastic Constraint Model (Frisch, Broe, & Pierrehumbert 1996, Frisch 1996) is an

attempt to model lexical regularities. Frisch et al. propose constraints that are functions

from phonological characteristics to acceptability values, which should predict

experimental well-formedness judgments and lexical frequency.7 The functions are of the

form acceptability = 1/(1+eK+Sx), where x is the numerical value of the phonological

characteristic, and K and S are parameters that determine the location and sharpness of

the boundary between acceptable and unacceptable.

To account for the Arabic constraint, a function is proposed that takes as its x the

similarity between two consonants and returns an acceptability value between 0 and 1.8
The acceptability value was compared against lexical frequency, and the match was

found to be good. Frisch 1996 compared this model to several others and found that it

was a better fit to the Arabic lexicon.

The Stochastic Constraint Model models knowledge of well-formedness, and

explains patterns in the distribution of exceptions to morpheme structure constraints. But

constraints in this model play a very different role from that of constraints in OT. To

quote Frisch 1996, “[the stochastic constraint] does not influence what the output is for

7
The mechanism relating well-formedness and lexical frequency in unclear, but we can say there is a two-
way relationship. On the one hand, lexical frequency shapes acceptability values by determining what
values the learner assigns to the parameters of the stochastic constraint. On the other hand, acceptability
values could shape lexical frequency by influencing how rare words or loans are “repaired” (low-
acceptability words would tend to drift towards repairs that enhance their acceptability), and influencing the
shape of newly coined words.
8
This is somewhat of a simplification. First, the function is acceptability = A/(1+eK+Sx), where A need not
be 1. In directly modeling lexical frequency (observed number of occurrences/expected number of
occurrences) without the mediating step of acceptability, Frisch 1996 uses other values of A to get a better
fit. Second, Frisch 1996 actually multiplies together three different constraints to get a total acceptability
value: one constraint is a function on the similarity of the first two consonants in a triliteral, one is on the
similarity between the second and third, and one is on the similarity between the first and third.

8
any particular input, but rather it constrains the space of possible inputs and outputs in a

probabilistic manner.” (p. 92) The mental system represented by the Stochastic

Constraint Model would have to exist alongside the system for mapping inputs to outputs.

This dissertation proposes a model in which the same system that maps inputs to outputs

can encode lexical regularities and patterns in the distribution of exceptions to those

regularities.

1.3. Preview of the proposal

It is conceivable that knowledge of lexical regularities resides outside the grammar—or

even that no discrete knowledge of the regularities exists at all. Speaker behavior that

appears to reflect such knowledge could merely be the result of some on-line procedure

such as consultation of a sample of the lexicon or matching to associative memory. These

two strategies are discussed at greater length in Chapter 6 and shown to be ill suited to

the regularities discussed in this dissertation. As argued there, the speaker must possess

knowledge that is abstracted away from the lexicon itself. The only linguistic subsystem

commonly proposed that contains such knowledge is the grammar. Therefore, the

approach taken here will be to incorporate knowledge of lexical regularities directly into

the grammar.

To accomplish that goal, this dissertation proposes a model of grammar that

allows the primacy of listed information to coexist with knowledge of lexical regularities.

Existing words’ behavior is encoded in their lexical entries; that information is preserved

through high-ranking faithfulness constraints and constraints that force listed information

to be used if available. Lexical regularities are encoded through low- and variably ranked

constraints, which are irrelevant for existing words, but determine the pronunciation of

novel words.

9
 The ranking tendencies of these subterranean constraints are learned through

exposure to the lexicon, using Boersma’s (1998) Gradual Learning Algorithm, which is

shown to be capable of learning rates of lexical variation: constraints that are violated by

many words become low-ranked, and constraints that are violated by few words become

high-ranked, even if none of those constraints are relevant for existing words once the

grammar reaches its adult state (in this case, because high-ranking faithfulness constraints

determine the optimal candidate).

Chapter 2 presents Tagalog nasal substitution, a sporadic morphophonemic

phenomenon. A statistical examination of the lexicon reveals that the distribution of

exceptions to nasal substitution is patterned. Experimental evidence is presented for the

psychological reality of nasal substitution and its subregularities. The chapter implements

the model for the case of nasal substitution, showing how the subterranean constraints

governing nasal substitution and its patterns produce rates of substitution on novel words

and acceptability ratings for novel words that are similar to the experimental results. In

particular, the paradoxical result that speakers perform nasal substitution at a low rate on

novel words, but rate certain types of nasal-substituted novel words as highly acceptable

is explained in terms of the listener’s probabilistic reasoning about her interlocutor’s

underlying form (in rating a novel word, the listener must entertain the possibility that for

her interlocutor, the word is not novel).

Chapter 3 shows how probabilistic interactions between speakers and listeners

perpetuate lexical patterns as new words enter the language. Bayesian reasoning on the

part of the listener results in a bias in favor of nasal-substituted pronunciations: although

they are the minority pronunciation for a novel word, listeners disproportionately tend to

add them to their lexicons (whereas unsubstituted pronunciations tend to be ignored). The

chapter presents the results of introducing novel words into a computer-simulated speech

10
community, attempting to replicate the rates of substitution for various stem types that

can be observed in Spanish loans.

Chapter 4 applies the model to vowel height alternations in Tagalog. Although

vowel raising under suffixation is nearly universal in native words, many loanwords from

Spanish and English have resisted raising. The chapter argues that the main predictor of

whether a word will resist raising is how amenable it is to being construed as reduplicated

(raising is then prevented, because it would disrupt reduplicative identity). It is argued

that a purely phonological mechanism (Aggressive Reduplication) drives such

morphosyntactically unmotivated reduplicated construals. This second case is of interest

because the subregularity involved is quite abstract, and does not emerge

straightforwardly from associative memory.

1.4. Tagalog

Because nearly all the data discussed in the body of this dissertation are from Tagalog,

this section covers some essential facts about the language, and gives details on how

lexical data were obtained. Although this dissertation’s main goal is to present a model of

lexical regularities, I hope that it will also be useful as a source of detailed information on

several aspects of Tagalog phonology.

Tagalog (Austronesian, Malayo-Polynesian, Western Malayo-Polynesian, Meso

Philippine, Central Philippine, Tagalog) is the national language of the Philippines (in

this role, it is sometimes called Pilipino). It has over 15 million first-language speakers

worldwide (Ethnologue 1996), and is used to some degree by 39 million Pilipinos. First-

language speakers are mainly in Luzon and Mindoro.

The language has long had contact to varying degrees with Chinese, Malay, and

languages of Indonesia and India; a moderate number of loanwords from these languages

11
are still in use. During the time of the Spanish occupation of the Philippines (mid

sixteenth through nineteenth centuries), there was extensive contact with Spanish;

starting with the U.S. occupation (first half of the twentieth century) and continuing to

today there has been extensive contact with English. There are now large numbers of

loanwords from Spanish and English.

1.4.1. Phonology sketch

The phoneme inventory of Tagalog is given in (4).

(4) Tagalog phoneme inventory

p t k ? i u
b d g e o
s h a
m n N
l
R
w j

The phonemes /d/ and /R/ were probably once allophones of the same phoneme (and were

represented identically in the pre-Hispanic syllabary): within native roots, they are in

complementary distribution, with [R] intervocalically and [d] elsewhere. Root-final /d/

always alternates between [d] when word-final and [R] when intervocalic because of

suffixation. Root-initial /d/ is always [d] when word-initial, and may be either [d] or [R]

when intervocalic because of prefixation. Spanish loans, however, introduced many [d]s

and [R]s in other positions.

The situations of /i/, /e/ and /u/, /o/ are similar: the high/mid distinction was

probably once purely allophonic (only two heights are distinguished in the syllabary),

with mid vowels restricted to final syllables, and high vowels elsewhere. For extensive

discussion of the situation today, see Chapter 4.

12
 Other sounds are frequently used in loanwords, such as [S], [tS], [dZ], [«Õ] and

sometimes [f].

The basic syllable structure is CV(C), although onset clusters are commonly

found in loanwords, and coda clusters occasionally. Most roots are disyllabic. Either

stress or length is contrastive.9 I will not take a position on which (for two opposing

views, see e.g. Schachter & Otanes 1972 and French 1988), and both are marked in all

examples (long vowels with no marked stress are secondary-stressed).

Tagalog is rich in morphology. There are many derivational prefixes, which are

often stacked several deep. There are two inflectional (and sometimes derivational)

infixes, -in- and -um-, which are inserted between the first C and V of the stem (the result

may be a verb, noun, or adjective depending on the construction).10 There are two
suffixes, -in and -an, which also play a variety of roles. When a vowel-final word is

suffixed, the allomorphs -hin and -han are used. There is also reduplication: the first C

and V can be copied (usually inflectional; I refer to this as REDCV), or the first two

syllables (derivational). Some examples of Tagalog affixes are shown in (5).

9
There are two types of word: those with a long, stressed penult, and those with a short penult and a
stressed ultima. There are a few loans that some speakers pronounce with antepenultimate stress and length.
In native words, a long/stressed penult must be open, but in some loans, it is closed. In derived words, there
may be length and secondary stress on the antepenult or earlier syllables.
10
In loans with complex onsets, the position of the infix varies (between the two onset consonants or
between the onset and nucleus). See Ross 1996.

13
(5) Examples of Tagalog affixes
bare stem: laki@ ‘size, bulk’

prefixation: ma-laki@ ‘big’

ma-pag-ma-laki@ ‘smug’

infixation: l-um-aki@ ‘to grow big’

suffixation: laki-ha@n11 ‘to enlarge (object focus12)’

reduplication: laù-laki@ ‘will grow big’

ma-laki@-laki@ ‘fairly large’

1.4.2. Notes on the data

Tagalog data of three types are presented: experimental data, lexical statistics, and

examples. The experimental data are discussed in detail in §2.3. The lexical statistics are

based on English (1986), a two-volume Tagalog-English, English-Tagalog dictionary.

The dictionary was compiled by Leo English, a (non-native speaker of Tagalog) priest

who lived in the Philippines for 30 years, and Teresita Castillo, a native speaker of

Tagalog. The exact methods for determining which pronunciations to include are not

known, and probably involved consensus among Castillo and the several other Tagalog

speakers who assisted. Because of the large size of the corpus and the frequent

disagreement among speakers as to the correct pronunciation of individual words, the

dictionary was used as the sole source of lexical statistics, producing a large, consistent

11
also lak-ha@n. See §4.7.2 for a discussion of syncope.
12
Every Tagalog sentence (with a few exceptions) has what may loosely be called the focus: a noun phrase
that bears the enclitic si (for proper names of people) or ?aN (for all other noun phrases); the other noun
phrases in the sentence bear the enclitic kaj/sa (if indirect object, goal, etc.) or ni/naN (if direct object or
subject). There are also corresponding focus and nonfocus pronouns. The verbal morphology indicates the
thematic role of the focused noun phrase. For example, in a sentence with the verb laki-ha@n, the object
being enlarged would be marked with ?aN, the person enlarging it with ni, and the instrument being used to
enlarge with sa. See Schachter & Otanes 1972 for a thorough description of Tagalog syntax.

14
source of pronunciations. Thus, although an individual word discussed in Chapter 2

might be pronounced with nasal substitution (see §2.2.1) by some speakers and without

by others, the overall statistics should be representative of the speech community.

Examples given in the text are drawn from English’s dictionary, from reference

sources such as Schachter and Otanes 1972 and Ramos and Bautista 1986, and from my

own observations of spoken and written Tagalog. I am not a native (or even fluent)

speaker of Tagalog, but have studied the language both as a linguist and in the classroom.

Transcriptions are IPA (Handbook of the International Phonetic Association

1999), with the exception that an acute accent is used to indicate stress. In some tables

and charts, where phonetic fonts were not available, “N” is used for [N], “?” for [?], and

“r” for [R]. Tagalog orthography is also used in some tables and charts; it is identical to

IPA except that “ng” is used for [N], “r” for [R], and “y” for [j], and [?] is not written.

1.5. Appendix: OT basics

The analytical framework used here is Optimality Theory (OT: Prince & Smolensky

1993). The machinery of Correspondence Theory (McCarthy & Prince 1995) is also

employed extensively. It is not possible of course to give a complete explanation of

Optimality Theory here, but a brief overview is possible. See Archangeli and Langendoen

1997 or Kager 1999 for a full introduction to OT.

OT employs two functions, Gen and Eval. Gen takes an underlying representation

(“input”) and returns a (possibly infinite) set of possible surface forms (“output

candidates”). Some output candidates might be identical to the input, others slightly

modified (for example by deleting one segment), others unrecognizable. Eval chooses the

candidate that best satisfies a set of ranked constraints; this optimal candidate becomes

15
the surface representation. The ranked constraints are violable, in the sense that the

optimal candidate may still violate some constraints.

The constraints are of two types: Markedness constraints enforce well-formedness of

the output itself, for example by forbidding consonant clusters. Faithfulness constraints

enforce similarity between the input and the output, for example by requiring all input

segments to appear in the output.

In standard OT, the constraint set is strictly ranked: a candidate that violates a high-

ranking constraint more than other candidates do can never redeem itself by satisfying
lower-ranked constraints. Eval can be thought of as choosing the subset of candidates that

violates the top-ranked constraint the fewest times, then of this subset, selecting the sub-

subset that violates the second-ranked constraints the fewest times, and so on until only

one candidate remains.

The “tableau” (a standard expositional device in OT) in (6) illustrates this procedure

for the input /ilp/ (upper left corner) in a hypothetical mini-language. Each of the output

candidates a, b, and c is flawed in some way: c, the candidate that looks most like the

input, has a consonant cluster; this violates the constraint against consonant clusters,

*CC, as indicated by the asterisk in the cell at the intersection of *CC’s column and

candidate c’s row. *CC is a Markedness constraint. Candidate b has deleted a segment,

and candidate a has inserted a segment; these candidates violate the Faithfulness

constraints DON’TDELETE and DON’TINSERT, respectively.13

In this language, *CC is the highest-ranked constraint (ranking is indicated by left-to-

right ordering of the constraints’ columns—we can also write

*CC>>DON’TDELETE>>DON’TINSERT). Eval first eliminates candidate c from the

13
These two constraint names are shorthands. See §2.4.3 for some standard constraint names and
definitions.

16
competition because it alone violates *CC. The elimination is represented by the

exclamation mark; the shading in the cells to the right represents the fact that candidate

c’s violations of lower-ranked constraints are now irrelevant. Eval next eliminates

candidate b, because of its violation of DON’TDELETE; now just one candidate remains

(a), so it is optimal, as indicated by the pointing finger. All of DON’TINSERT’s cells are

shaded, because it is now irrelevant. In this language, then, an input string /ilp/ is

pronounced [ilip]; in another language, the constraint ranking might be different and

would choose a different candidate.

(6) Sample OT tableau

/ilp/ *CC DON’TDELETE DON’TINSERT
a . [ilip] *
b [il] *!
c [ilp] *!

OT was chosen as the analytical framework here because it allows straightforward

expression of the idea that when the lexicon cannot determine some aspect of a word’s

pronunciation, the likelihood that a particular option will be chosen depends on that

option’s well-formedness along a variety of conflicting dimensions (see §2.7).

17
 2. The model as applied to nasal substitution

2.1. Chapter overview

This chapter presents a model of lexical regularities through the example of nasal

substitution in Tagalog. Section 2.2 describes the phenomenon of nasal substitution and

its distribution in the lexicon. Section 2.3 presents the results of an experiment aimed at

assessing the psychological reality of nasal substitution in production and judgment of

well-formedness. Section 2.4 gives a grammar for nasal substitution, with constraints that

encode the regularities in its distribution. Section 2.5 considers several possibilities for

how potentially nasal-substituting words are represented in the lexicon. Section 2.6

shows how the grammar in §2.4 could be learned from exposure to the lexicon, using

Boersma’s (1998) Gradual Learning Algorithm. Section 2.7 describes the speaker’s

probabilistic use of the grammar for novel and existing words. Finally, §2.8 describes

how the listener uses the grammar to determine her interlocutor’s underlying form and to

arrive at acceptability judgments.

18
 2.2. Nasal Substitution

2.2.1. The phenomenon

Nasal substitution is a phenomenon that occurs somewhat sporadically in the Tagalog

lexicon. When certain prefixes are attached to a stem beginning in a sonorant, they appear

as paN-, maN-, or, less often, naN-, which is derived morphologically from maN-.14 (e.g.,

hukbo@ ‘army’, paN-hukbo@ ‘military’). But when these same prefixes attach to an

obstruent-initial stem, either they appear with place assimilation to the obstruent, as pam-

/pan-/paN-, mam-/man-/maN-, nam-/nan-/naN- (e.g., po?o@k ‘district’, pam-po?o@k

‘local’), or the final nasal of the prefix and the obstruent appear to combine into a nasal

that is homorganic to the original obstruent (e.g., mag-biga@j ‘give’, ma-miga@j

‘distribute’). It is the second case that is known as nasal substitution. In (7) are shown

examples, for every consonant in the Tagalog inventory, of substitution and

14
There are a variety of productive morphological constructions that participate in nasal substitution, but in
all of them, the prefix complex ends in paN-, maN-, or naN- (even though, morphosyntactically, it may be
preferable to think of the affixes as a whole, since the meaning of the prefix complex is often not
compositional). There are also some unproductive constructions that can trigger nasal substitution, whose
prefix complexes end in, taN-, tuN- siN-, hiN- (the only common one), kaN-, and kuN- (e.g., biù@laN ‘number’,
tam-bila@N ‘digit’; bali@k ‘upside-down’, tum-bali@k ‘return’; pu@ùno? ‘leader’, si-mu@ùno? ‘grammatical
subject’; ku@ùto ‘louse’, hi-Nutu@ù-hin ‘to pick out lice’; pata@j ‘corpse’, ka-maùta@-jan ‘death’; baba@?
‘descent’, mag-pa-kum-baba@? ‘humble’). The fairly productive construction mag-kaN-RCV, for verbs of
accidental result (dapa@? ‘face down’, mag-kan-da-Ra@ùpa? ‘to fall on one’s face’), never produces
substitution.

This set exhausts the prefixes that end in N, except for a group that I do not consider real prefixes,
because they seem more like members of a compound: wala@N-, (?i)sa@N-, (ka)siN-, pagi@N- and magi@N- (e.g.,
ba@ùjad ‘payment’ wala@N-ba@ùjad ‘free’; da@ùli? ‘finger-width’ san-da@ùli? ‘one finger width’; ?iti@m ‘black’,
kasi@N-?iti@m ‘as black as’; bu@ùNa ‘fruit’ pagigi@N-bu@ùNa ‘conversion into a fruit’; su@ùka? ‘vinegar’ magi@N-
su@ùka? ‘to become vinegar’). These are all two syllables long (except for optionally shortened (?i)sa@N- and
(ka)siN-), can bear their own stress, produce semantically transparent words, never induce nasal
substitution, and often fail to undergo nasal assimilation. In addition, wala@? ‘does not have/exist’ and ?isa@?
‘one’ also occur as free-standing words, which require the “linker” -N- under certain circumstances.

19
nonsubstitution, using a variety of common morphological constructions that can trigger

substitution.

(7) Nasal-substituting prefixes with various stems

p pighati@? ‘grief’ pa-mi-mighati@? ‘being in grief’
po?o@k ‘district’ pam-po?o@k ‘local’
t pag--tu@ùloj ‘staying as guest’ kaù-pa-nulu@ùj-an ‘fellow lodger’
tabo@j ‘driving forward’ pan-tabo@j ‘to goad’
k kamka@m ‘usurpation’ ma-pa-Namka@m ‘rapacious’
kaliski@s ‘scales’ paN-kaliski@s ‘tool for removing scales’
? ?isda? ‘fish’ maù-Ni-Nisda@? ‘fisher’
?ulo@l ‘silly’ maN-?ulo@l ‘to fool someone’
b mag-biga@j ‘to give’ ma-miga@j ‘to distribute’
bigka@s ‘pronouncing’ mam-bi-bigka@s ‘reciter’
d dala@ùNin ‘prayer’ ?i-pa-nala@N-in ‘to pray’
dini@g ‘audible’ pan-dini@g ‘sense of hearing’
g ginda@j15 ‘unsteadiness on feet’ pa-Ni-Ninda@j ‘unsteadiness on feet’
ga@ùwaj ‘witchcraft’ maN-ga-ga@waj ‘witch’
s su@ùlat ‘writing’ maù-nu-nula@t ‘writer’
pan-su@ùlat ‘writing instrument’
h hukbo@ ‘army’ paN-hukbo@ ‘military’
m marka@ ‘mark’ paN-marka@ ‘marker’
(no examples of n16)
N Na@lit ‘grinding of teeth’ paN-Na-Na@lit ‘grinding of teeth’
R Rasjo@n ‘ration’ paN-Rasjo@n, pan-Rasjo@n ‘for rationing’
l la@gom ‘assimilation’ ma-pan-la@gom ‘monopolistic’
w mag--wisi@k ‘to sprinkle’ paN-wisi@k ‘sprinkler’
j jamo@t ‘annoyance’ maN-jamo@t ‘to annoy’

A few remarks on the examples in (7): First, when nasal substitution occurs in

conjunction with reduplication, both base and reduplicant are substituted (pa-mi-

mighati@? rather than *pa-mi-bighati@? or *pa-mi-mbighati@?); when no nasal substitution

occurs, the assimilated nasal precedes only the reduplicant (mam-bi-bigka@s rather

than *mam-bi-mbigka@s). I adopt Wilbur’s (1973) and McCarthy and Prince’s (1995)

15
One of only 2 instances of substitution of g that I found.
16
Nasal-initial roots are few in Tagalog. The absence of any n-initial roots that have potentially nasal-
substituting derivatives is probably accidental.

20
proposal that “overapplication” of nasal substitution in pa-mi-mighati@? results from

reduplicative correspondence. Note that the overapplication shows that a nasal resulting

from substitution belongs to the stem (although it may also belong to the prefix in some

sense; see the discussion of coalescence in §2.4), whereas a prefix nasal that merely

assimilates is not part of the stem.

Second, it is not clear whether nasal substitution is possible on nasal-initial stems:

nasal-initial stems are rare to begin with, and among those that do exist, it is not always

possible to tell what the prefix is. For example, in ma-manhi@d ‘to become numb’, from
manhi@d ‘numb’, it is not clear whether the prefix is simply ma- (which can also form

verbs, with similar semantics), or maN- with nasal substitution.17 There do exist

unambiguous constructions (such as maN+REDUPLICATION—there is no potentially

confusable ma+REDUP), but no cases of nasal-initial stems in these constructions.

Third, glottal stop is problematic. Many researchers have assumed that initial

glottal stop in Tagalog is simply predictably inserted in vowel-initial words (since there

are no strictly vowel-initial words); the preservation of initial glottal stop in prefixed

words like mag-?a@ùwaj ‘to fight’ (or maN?ulo@l) would then be regarded as the effect of a

tendency to align morpheme boundaries with syllable boundaries (for a formal theory of

alignment, see McCarthy & Prince 1993, Cohn & McCarthy 1998). And a word like

17
Schachter and Otanes (1972) argue that these verbs are maN-prefixed, because their gerunds are formed
by changing m to p and reduplicating, as are the gerunds of uncontroversially maN-prefixed verbs (taùkot
‘fear’, ma-na@ùkot ‘to intimidate’, pa-na-na@ùkot ‘intimidating’). In contrast, ma- verbs’ gerunds are formed
by replacing ma- with pagka- (maù-bujo@ ‘to get involved’, pagka-bujo@ ‘getting involved’). But Carrier
(1979) points out that some m → p & RCV gerunds do come from ma- verbs (pa-li-li@ùgo? ‘bathing’ from ma-
li@ùgo? ‘to take a bath’).

Carrier (1979) argues against the maN-with-substitution analysis for nasal-initial stems, because
some of the nasal-initial stems that take ma-/maN- do not substitute when combined with paN-, and so
should not substitute with maN- (paN-no?o@d ‘for watching’). But, I have found many stems that substitute
with maN- but not with paN- (bunto@t ‘tail end’, ma-munto@t ‘to finish last’, pam-bunto@t ‘tailpiece’).

21
maùNiNisda@? would be failure of alignment rather than true nasal substitution, either with

the nasal of the prefix becoming associated to the stem, or with reduplicative

correspondence causing the second N to be inserted.18 Since glottal stop is phonemic

word-finally, I prefer to regard word-initial glottal stop as phonemic rather than

epenthetic (why pick glottal stop as the epenthetic segment rather than something else?),

and I view maùNiNisda@? as nasal-substituted, although, as will be seen below, the

distribution of “substituted” glottal stops in the lexicon is puzzling.

2.2.2. Distribution of exceptions

I collected all 1,736 words from English (1986) that had an obstruent-initial stem and a

potentially nasal-substituting prefix, and found two trends. First, substitution is most

likely with a front stem-initial consonant (p or b) and least likely with a back consonant

(k or g). Second, substitution is more likely if the stem-initial consonant is voiceless than

if voiced. Both trends can be seen in (8), which combines data from all constructions (t

and s are also combined, to better illustrate the two trends; t and s are separated in the

more detailed charts that follow). 19

18
A similar proposal, considered and rejected by Carrier (1979), is that there is a phonemic difference
between truly glottal-stop-initial and truly vowel-initial stems, which determines whether or not nasal
substitution will appear to occur. Thus ?isda@? would be underlyingly /isda@?/, and ?ulo@l underlyingly
/?ulo@l/. There are some glottal/vowel-initial stems whose derivatives vary in whether or not they substitute,
but this does not refute Carrier’s idea: such stems would be underlyingly vowel-initial, but in some
derivatives morpheme-specific alignment constraints would force an epenthetic glottal stop.
19
Previous accounts of the lexical distribution of nasal substitution have noted (not quite correctly) that g
never substitutes (Bloomfield 1917, Schachter & Otanes 1972); that d and g rarely substitute (Blake 1925);
that voiceless consonants substitute more than voiced ones (De Guzman 1978, but see fn. 20); and that
morphology matters (Schachter & Otanes 1972, De Guzman 1978).

22
(8) Rates of nasal substitution for entire lexicon

100%
percentage of words that substitute

10 26 17
90% 100 70 97
80%
70%
60%
unsubstituted
50%
40% substituted
30%
20%
10% 253 430 185 177 25
0% 1

p t/s k b d g
stem-initial obstruent

Different constructions have different overall substitution rates. The bar charts in

(9) show rates of substitution for each stem-initial obstruent in the most common affix

patterns. The breakdown by affix is suggested in part by De Guzman (1978), who

distinguished adversative from nonadversative verbs,20 and instrumental adjectives (ti@tik

‘writing’, pa-ni@tik ‘used for writing’) from reservative adjectives (baNke@te ‘banquet’,

pam-baNke@te ‘for a banquet (said of clothes, food, etc.)’).21

20
Adversative verbs are hostile or harmful to the patient (e.g., bato@ ‘stone’, ma-mato@ or mam-bato@ ‘to
throw stones at’). Nonadversative verbs include inchoatives (paja@t ‘thin’, ma-maja@t ‘to become thin’),
statives (butikti@k ‘teeming with’, ma-mutikti@k ‘to teem with’), professional verbs (gamo@t ‘medicine’, maN-
gamo@t ‘to practice medicine’), habitual verbs (sigari@ljo ‘cigarette’, ma-nigari@ljo ‘to be a smoker’),
distributives (k-um-u@ùha ‘get’, ma-Nu@ùha ‘to gather things’), and repetitives (binta@ùna ‘window’, ma-
minta@ùna ‘to keep looking out a window’).
21
De Guzman claimed that in non-adversative verbs, substitution is obligatory for all obstruents and that in
adversative verbs, substitution is obligatory for voiceless Cs but optional for voiced Cs and glottal stop. (9)
shows that there are some counterexamples to the first clause of the claim; although the classification of
some verbs could be argued over, there are some nonsubstituting verbs that are definitely nonadversative

23
 The constructions illustrated in (9) are adversative-verb-forming maN-;
nonadversative-verb-forming maN-; paN+RCV-, which forms mainly gerunds, but also
some less predictable nominalizations (tahi@? ‘stitch’, pa-na-nahi@? ‘sewing’); maN+RCV-,
which forms professional or habitual nouns (ba@ùtas ‘law’, mam-ba-ba@ùtas ‘legislator’);
noun-forming paN- (instrumentals, gerunds, and unpredictable nominalizations, e.g.,
gu@ùgol ‘expense’, paN-gu@ùgol ‘spending money’); and reservative-adjective-forming paN-
(no other constructions had enough examples with each segment to make a chart
meaningful).
Within each chart, each obstruent is scaled for comparison. For example, the first
column in the first graph says that there are a total of 39 p-initial stems listed in English
(1986) that took the paN-RCV- construction, and of those, all are substituted. The fifth
column shows that there are 35 b-initial stems, of which 29 substitute, 1 varies, and 5 do
not substitute.

(gi@ùgil ‘tremble, thrill’, maN-gi@ùgil ‘to tremble, thrill’). There are no counterexamples to the second clause
of the claim. De Guzman further claims that in instrumental adjectives, substitution is optional for voiceless
Cs and impossible for voiced Cs and glottal stop. Instrumental adjectives are not included in (9) because
there were too few tokens; there were indeed no substituted voiced Cs, but there were only 5 tokens of b,
none of d, and 2 of g.

24
(9) Rates of substitution for various prefixes

Substituted Varies Unsubstituted

paN+RCV-

100% 1
5 7 14 7
1
80%
60%
40%
20%
39 22 41 25 29 3 17
1
0%
p t s k b d g ?
stem-initial segment

maN+RCV-
100% 1
2
12 12 6
80% 20
60%
3
40%
20%
18 19 25 20 15 7
1 1
0%
p t s k b d g ?

25
 maN- (adversative)

100%
12 9 21
6

50%

11 12 39 11 6 4
0%
p t s k b d g ?

maN- (other)

100% 1 6 4
3 13

50%
7

65 39 32 74 66 34
0%
p t s k b d g ?

26
 paN- (noun)

100% 3
8 6
17 27 22
80% 7 7
26
60% 18 13
5
40% 8
20%
27 25 20 7 11 6
0% 1
p t s k b d g ?

paN- (reservative adjective)

100%
3 5 7 4 5
80%
5
5
60% 17
40% 3
3
20% 3
3 2 2 1
1
0%
p t s k b d g ?

27
 To determine the statistical significance of the voicing and place-of-articulation
effects, I used contingency table analysis, a way of determining whether two nominal
variables are independent of each other. Glottal stop is omitted from the statistical results,
because although it mostly patterns as the most posterior voiceless stop (substituting a bit
less often than k), in adversative maN- verbs, ? inexplicably substitutes less than 20% of
the time, whereas the other voiceless stops always substitute. As noted above, it is
unclear whether ? actually undergoes nasal substitution (rather than simple deletion) at
all.
To test whether the voicing effect was significant, we can construct a table with
the observed number of voiced and voiceless consonants22 that were unsubstituted or
substituted,23 as in (10) and a similar table with the “expected” values—the values that we
would see if voicing and substitution were independent of each other—as in (11).

(10) Voicing and nasal substitution: observed frequencies

unsubstituted substituted total
voiceless 46 578 624
voiced 217 142 359
total 263 720 983

(11) Voicing and nasal substitution: expected frequencies

unsubstituted substituted total
voiceless 166.950 457.050 624.000
voiced 96.050 262.950 359.000
total 263.000 720.000 983.000

22
Using just the 6 most common constructions. All other constructions account for only an additional 66
words.
23
Varying cases are omitted, because a smaller table yields more-conservative significance results.

28
 The table of expected frequencies uses the same totals as the table of observed
frequencies, and fills in the other (boldface) values proportionally: since in total, 624/983
= 63.48% of the words are voiceless-initial, 63.48%, or 166.950, of the 263 unsubstituted
words should be voiceless-initial. Conversely, since 263/983 = 26.75% of the words were
unsubstituted, 26.75%, or 166.950, of the 624 voiceless-initial words should be
unsubstituted.
Inspecting the two tables visually, it is clear that the observed and expected values
are quite different. It was expected that about 457 voiceless-initial stems would
substitute, but 578 did; it was expected that about 96 voiced-initial stems would fail to
substitute, but 217 did. In other words, substitution is more common than expected
among voiceless-initial stems, and less common than expected among voiced-initial
stems.
To test the significance of the differences between the observed and expected
values, χ2, which is the sum, for all table cells (excluding the totals), of
(observed-expected)2/expected.

In this case, χ2 = 327.572. If two nominal variables like substitution and voicing are
known, given the number of rows and columns in the table, the probability p that any
given value of χ2 or a higher value would be obtained by chance is known. In this case,
p < 0.0001.

It would be ideal to test for the voicing effect within each place of articulation and
within each morphological construction, since it might be that, for instance, a
disproportionately large number of voiceless-initial stems in a construction that has a
high independent rate of substitution is skewing the results. The numbers are too small to
do this kind of breakdown, but it should be apparent from inspection of the charts in (9)
that the voiceless-initial stems are not concentrated in the highly-substituting

29
constructions, and that within every construction, the voiceless-initial stems substitute
more frequently.
Similar contingency tables can be constructed for nasal substitution and place of
articulation. Here, we must break the data into voiceless and voiced cases, since we
already know that voicing has a strong effect, and the proportion of voiced- vs. voiceless-
initial stems is not steady across place of articulation. Observed and expected frequencies
are given in (12) and (13).

(12) Place of articulation and nasal substitution: observed frequencies

voiceless unsubstituted substituted total
labial 6 163 169
dental 25 276 301
velar 15 139 154
total 46 578 624

voiced unsubstituted substituted total

labial 80 128 208
dental 58 12 70
velar 79 2 81
total 217 142 359

(13) Place of articulation and nasal substitution: expected frequencies

voiceless unsubstituted substituted total
labial 12.458 156.542 169.000
dental 22.189 278.811 301.000
velar 11.353 142.647 154.000
total 46.000 578.000 624.000

voiced unsubstituted substituted total

labial 125.727 82.273 208.000
dental 42.312 27.688 70.000
velar 48.961 32.039 81.000
total 217.000 142.000 359.000

30
 In (12) and (13), there are more rows, so trends are harder to spot. To make them
more apparent, (14) lists (observed-expected)/expected for each cell. A large negative
value means that the observed value was much lower than expected, and a large positive
value means that it was much higher than expected.

(14) Place of articulation and nasal substitution: (observed-expected)/expected values

voiceless unsubstituted substituted
labial -0.518 0.041
dental 0.127 -0.010
velar 0.321 -0.026

voiced unsubstituted substituted

labial -0.364 0.556
dental 0.371 -0.567
velar 0.614 -0.938

Recall that the place effect predicts that labials should be substituted more often
than expected (positive value in the top-right cell of (14)) and unsubstituted less often
than expected (negative value in the top-left cell), velars should be the opposite, and
dentals should fall somewhere in between. The tables in (14) show that in both the
voiceless and voiced cases, labials are substituted more often than expected (although the
effect is weak for voiceless p) and are unsubstituted less often than expected; velars are
substituted at about the expected rate when voiceless and much less often when voiced,
and are unsubstituted more often than expected in both cases. Dentals and velars can be
compared by noting that the tendency to be unsubstituted more often than expected is
greater than expected in velars than in dentals in both the voiceless (0.321 vs. 0.127) and
voiced (0.614 vs. 0.371) cases. In the voiced case, the tendency to be unsubstituted less
often that expected is much stronger among velars than among dentals (-0.935 vs. -
0.551); in the voiceless case, the difference between velars and dentals is tiny (although
in the right direction: -0.026 vs. -0.010)

31
 We can perform a χ2 test for the place-of-articulation effect too, but the results are
less meaningful, because they tell us only that (12) and (13) are significantly different,
not whether the front-to-back trend is significant. The χ2 value for voiceless consonants is
5.264; for a table this size, the a probability of obtaining such a large χ2 by chance if
place of articulation and substitution were independent is p= 0.07. The χ2 value for
voiced consonants is 103.345, p < 0.0001. It is not surprising that the place differences
are small among the voiceless consonants, because in four of the six morphological
constructions included there is a ceiling effect—nearly all the voiceless consonants of any
place of articulation are substituted.
Finally, (15) summarizes the results of performing pairwise contingency-table
analyses between pairs of consonants. The test used was Fisher’s Exact Test,24 which
enumerates all tables having the same row and column totals as the table of observed
values. Each such table’s probability of occurring, assuming no association between the
variables (initial obstruent and nasal substitution), can be calculated. The probabilities for
the tables that are skewed in the same direction as the observed table, to the same degree
or more extremely, are added to find the probability p that such a skewed table could
have arisen by chance if the two variables were independent.

(15) Pairwise differences in rate of substitution

expected difference Fisher’s Exact Test
voicing effect p>b p < 0.0001
t,s>d p < 0.0001
k>g p < 0.0001
place effect p>t p = 0.0528
t,s>k p = 0.6038
b>d p < 0.0001
d>g p = 0.0034

24
All statistical results were calculated in Statview.

32
2.2.3. Productivity of nasal substitution

There are several ways in which nasal substitution appears unproductive. First, despite
the lexical trends described above, it is of course not completely predictable which words
will undergo substitution—substitution is not even predictable among derivatives of the
same stem, as illustrated in (16). Note the lack of a strict implicational hierarchy for
substitution among the constructions paN-, paN+REDCV-, maN+REDCV-, and maN-.

(16) Differing behavior among derivatives of the same stem

biga@j ‘gift’
pam-biga@j ‘gifts to be distributed’
pa-mi-miga@j ‘act of giving away’
ma@ù-mi-mi@gaj ‘distributor’
ma-miga@j ‘to distribute (actor focus)’

bugbo@g ‘wallop’
pa-mugbo@g ‘wooden club used to pound clothes during washing’
pam-bu-bugbo@g ‘act of clubbing or pounding; assault’
mam-bugbo@g ‘to wallop’

bu@los ‘harpoon’
pa-mu@los ‘harpoon’
mam-bu-bu@los ‘harpooner’

bu?o@? ‘whole’
pam-bu?o@? ‘something used to produce a whole’
pa-mu-mu?o@? ‘becoming whole; coagulation’
ma-mu?o@? ‘to solidify; to clot’

Second, although the semantic connection between stem and derivative is always
apparent, exact meanings are sometimes unpredictable, especially with certain prefixes,
such as verbal maN-. Note that semantic idiosyncrasy is found in both substituted and
unsubstituted words:

33
(17) Semantic unpredictability with nasal-substituting affixes
?aba@N ‘watcher’
maN-?aba@N ‘to wait near people who are eating, hoping to get
some food’
baba@ù?e ‘woman’
mam-baba@ù?e ‘to have a mistress’
si?i@l ‘oppressed by a ruler’
ma-ni?i@l ‘to strangle to death’
?iba@ùbaw ‘surface’
paùN-?iba@ùbaw ‘veneer’
ki@ùta ‘visible’
paù-Nita@ù?-in, paù-Nita?-i@n25 ‘apparition, omen’
tu@ùbig ‘water’
ma-nubi@g ‘to urinate’
bali@k ‘return’
pa-mali@k ‘hand rudder’
ga@ntSo ‘hook’
maN-ga-ga@@ùntSo ‘con man’

Third, certain affixes can cause unpredictable stress/length shifts. Note that this
idiosyncrasy too occurs in both substituted and unsubstituted words (but see (62)):

(18) Unpredictable stress/length shifts associated with nasal-substituting affixes

tahi@? ‘sewing’
maù-na-na@ùhi? ‘seamstress’
cf. puna@ ‘remark’
maù-mu-muna@ ‘critic’
?a@ùmak ‘town’
maN-?a-?ama@k ‘resident of town’
cf. ka@ùRit ‘sickle’
maù-Na-Na@ùRit ‘person whose job it is to cut grass with a sickle’

25
This stem is exceptional: it has a final glottal stop only when suffixed.

34
 tu@ùbig ‘water’
ma-nubi@g ‘to urinate’
cf. ki@ùkil ‘carpenter’s file’
ma-Ni@ùkil ‘to chisel; to ask for money’
si@ùpit ‘claws’
pan-sipi@t ‘(type of) rat-trap’
cf. ga@ùmas ‘weeding’
paN-ga@ùmas ‘tool for weeding’

The result is that for many words with nasal-substituting affixes, a speaker must
know a number of facts not predictable from other words containing the same stem—
whether or not the word undergoes substitution, the meaning of the word, and the stress
of the word—and thus must maintain a separate lexical entry for that word (for a
discussion of other ways to encode the unpredictable information, see §2.5).
If most or all words with nasal substitution are fully listed, there is no need to
represent nasal substitution in the grammar: each word is simply pronounced the way it is
listed (see §1.1.1.1). The sticking point here is whether or not nasal substitution is part of
speakers’ competence. If it is, it should be accounted for (somehow). The following
section addresses this question experimentally.

35
 2.3. An experiment

2.3.1. Introduction

I conducted an experiment aimed at answering two questions: (i) Is nasal substitution

productive? (ii) Are speakers aware of the lexical patterns within nasal substitution? If
the answer to either of these questions is yes, then perhaps nasal substitution belongs in
the grammar—certainly it must be accounted for somewhere in the system that governs
linguistic behavior, whether in the grammar or in some other subsystem. As discussed in
§1.3, this dissertation takes the approach that absent a clear understanding of how other
subsystems could account for a particular linguistic behavior, the behavior should be
accounted for by the grammar wherever plausible.
Nine native speakers of Tagalog living in Los Angeles participated. As shown in
(19), they ranged in age from 18 to 69, and had emigrated from the Philippines 3 to 20
years earlier (age at emigration did not correlate with productivity of nasal substitution).

(19) Personal characteristics of experiment participants

Participant # Age Age at emigration from Philippines
1 27 7
2 46 35
3 43 40
4 69 66
5 43 34
6 56 50
7 40 30
8 18 8
9 37 25

2.3.2. Task I: productivity

In the first task, participants were shown a series of cards, each of which had a
crude illustration of a person performing a farming or craft activity, with two sentences

36
(in regular Tagalog orthography, with accent marks26) printed at the top. A sample card is
shown in (20).

(20) Sample card for Task I

The sentences were designed as a “wug” test (Berko 1958) for the maN+RCV-
construction, which forms professional and habitual nouns (similarly to English -er):
participants had to produce the maN+RCV- form of a novel stem, which involved deciding
whether or not to perform nasal substitution. For example, in the sentence shown in (21),
the novel root is bugna@t, presented in a construction (pag+RCV-) that does not permit
substitution. To fill in the blank, the participant would probably choose one of maN-bu-

26
Accent marks—which are optional and not commonly used—indicate nonpenultimate stress and the
presence or absence of final glottal stop. I used accent marks in this standard way, but also placed accent
marks over penultimate stressed syllables, to ensure that the intended stress was always clear.

37
bugna@t (no substitution, no assimilation), mam-bu-bugna@t (assimilation only), or ma-

mu-mugna@t (substitution).

(21) Sample sentence pair for Task I

Pagbubugnát ang trabaho niya. Siya ay ________________.
to-bugnat (topic) job his/her He/she (inversion)
His/her job is to bugnat. He/she is a ____________.

The experiment was carried out in individual sessions. Starting with two real-
word examples (blanks filled in) and then two real-word training items, the participant
took each card and read the sentences aloud, filling in the blank. Participants in Group A
(4 participants) were given some real words mixed in with novel words, and were told
that many of the words were rare and that if they didn’t know a word or its maN+RCV-
form, they should just guess. Participants in Group B (5 participants), were given only
novel words after the training items, and were told that the words were invented and there
were no right or wrong answers. (See §0 for a complete list of stimuli).
The purpose of the illustrations was to encourage participants to think of the
words as real. Since none of the participants grew up in a rural environment, it was
plausible that they would not be familiar with farming and craft terms. There is a large
part of the Tagalog vocabulary known as “deep Tagalog”—affixed words which have
been largely replaced by Spanish and English loanwords—so the idea that an unfamiliar
word could still be real and native should not seem too implausible to Tagalog speakers.
When Group A participants were told at the end of the experiment that most of the words
were in fact novel, three of the four expressed mild surprise; one said that he had so
suspected.

38
2.3.2.1. Results of Task I

The main result from Task I, shown in (22),27 was that substitution rates were much lower
than the rates found in the lexicon for maN+RCV-, but were higher than zero. In other
words, nasal substitution is neither very productive nor completely unproductive. Note
that Group B included one participant (#3, a Tagalog instructor at a university), who had
a very high rate of substitution. If she is omitted, the rate of substitution for Group B is
much lower. The difference between Groups A and B (A has a slightly higher
substitution rate) is not significant. To give some idea of the amount of inter-speaker
variation, (23) gives overall substitution rates for each participant; the four columns on
the left are speakers from Group A, and the five columns on the right are speakers from
Group B.

27
Token counts shown in (22) are for all speakers combined. Because Group A has one few speaker than
Group B, token counts are not the same between the two groups. One token was omitted (from Speaker #3)
because it could not be clearly classified as substituted or unsubstituted (maùNaNatha@l for taha@l—perhaps
interference from katha@? ‘literary work’, maùNaNatha@? ‘author’?)

39
(22) Rates of substitution on novel words

Substituted Unsubstituted

Group A

100%
80%
7
8 9 8 9
60% 11 11
15
40%
20% 5
4 3 4 3
1 1 1
0%
p t s k b d g ?

Group B

100%

80%
11 11 10 11
60% 11 12
14 20
40%

20% 5
4 3 4 3 4
0% 1

p t s k b d g ?

40
(23) Overall rates of substitution on novel words, broken down by participant

substituted unsubstituted

100%

80%

60%

40%

20%

0%
1 4 5 9 2 3 6 7 8
Group A Group B
Speaker #

Group B’s low rate of substitution (compared to the proportion of existing words
that substitute) is not surprising. This group was told they were dealing with novel words,
and it makes sense not to perform nasal substitution in coining a novel derived word, in
order to promote recoverability of the stem for the listener (especially since nasal
substitution neutralizes voicing and continuancy distinctions in the stem). With an
established word that would be familiar to the listener, recoverability is less of a concern.
The low rate of substitution for Group A might seem puzzling, though, because
this group was told they were dealing with real words, and so should be making guesses
that would match rates of substitution in the lexicon. But Group A was told they were
dealing with rare real words, and so they may still have been matching lexical
frequencies—the lexical frequencies found in rare words. Bloomfield (1917) asserted that
nasal substitution was more frequent among common words, and although I have no

41
lexical-frequency data against which to test this assertion systematically, it seems
plausible.28

2.3.3. Task II: acceptability

The second experimental task was designed to determine whether or not participants’
grammars include the patterns of voicing and place of articulation seen in nasal
substitution. Substitution rates in the first task were too low to probe for effects of
voicing and place. Task II was administered immediately after Task I: starting with four
novel-word practice items (substituted and unsubstituted for each of two stems) each
participant (whether from Group A or Group B) was given cards with the same
illustrations and the same sentences as in Task I, but this time with the blanks filled in, as
shown in (24). Each root was presented twice (but not consecutively; order was
randomized), once substituted and once unsubstituted.

(24) Example stimuli for Task II

Kung pagbubugnát ang trabaho niya, siya ay mamumugnát.
Kung pagbubugnát ang trabaho niya, siya ay mambubugnát.
‘If her/his job is to bugnat, she/he is a bugnat-er’

The participant read the sentences aloud, then stated his or her rating of the sentence pair,
on a scale from 1 (bad) to 10 (good).

2.3.3.1. Results of Task II

Participants’ acceptability judgments generally reflected lexical frequencies. (25) shows

the combined average for each segment of the rating given to a substituted stimulus

28
Cf. English verbs: irregulars tend to have higher frequency than regulars, in part because low-frequency
irregulars are more likely to regularize over time (Bybee 1985).

42
minus the rating given to the corresponding unsubstituted stimulus. A positive number
means that over all, participants rated the substituted stimulus higher; a negative number
means that over all, participants rated the unsubstituted stimulus higher.

(25) Acceptability judgments: substituted - unsubstituted; error bars indicate 95%

confidence interval

4
3

0
-1 p t s k b d g ?
-2

-3
-4

-5

The positive numbers for voiceless-initial roots and negative numbers for voiced-
initial roots mean that over all, participants tended to prefer the substituted stimuli for
voiceless-initial roots and tended to prefer the unsubstituted stimuli for the voiced-initial
roots, reflecting the voicing effect. And, except for the unexpectedly low ratings for p,29
acceptability judgments also reflected the place effect. The voiceless/voiced difference is

29
The p-t and p-s differences are not significant. I investigated the possibility that the low ratings for
substituted p were the result of a neighborhood effect, but they do not appear to be: for each stimulus word,
I counted the number of substituting and nonsubstituting words in its phonological neighborhood. The
neighborhood was defined as the set of words sharing 5 segments, in the right positions (with empty codas
counting as segments), with the target word. The average number of substituting words in the
neighborhoods of the p stimuli was 2 (average number of unsubstituted = 0), and the average number of
substituting words in the neighborhoods of the s and t stimuli was also 2 (average number of unsubstituted
= 0.33).

43
highly significant—p < 0.0001 by Scheffé’s F.30 The place effect is not very significant:
because of the low values for p-initial stems, the overall difference between bilabials and
dentals is not even in the right direction. The difference between bilabials and velars is in
the right direction, but is not significant (p < 0.0736 by Scheffé’s F). The difference
between dentals and velars is in the right direction and significant (p = 0.0168 by
Scheffé’s F).
An ANOVA on voicing, place, and speaker shows that there was no significant
interaction between voicing and place, meaning that the magnitude of the voicing effect
does not vary significantly by place of articulation, and the magnitude of the place effect,
such as it is, does not vary significantly by voicing. There were, however, significant
interactions between voicing and speaker (F = 3.088, p = 0.0056) and between place and
speaker (F = 3.402, p = 0.0002), meaning that the voicing and place effects had different
strengths for different speakers. There was no significant difference in acceptability
ratings between Group A and Group B.

30
For the ANOVA and Scheffé’s results, some data had to be omitted into order to balance cells. Data for
s-initial stems were omitted (to avoid having twice as many data points for voiceless dentals as for other
categories); data for one of the da-initial stems was omitted (to avoid have 25% more data points for d than
for other segments); and data were excluded for participant #6, who made several errors in reading aloud
the stimuli (not applying substitution, although the stimulus was substituted; the errors were all on velar-
initial stems, which can be confusing to read because the digraph “ng” is used to represent N).

44
 2.4. The grammar

2.4.1. Desiderata for an analysis

The experimental results described above suggest that nasal substitution and its patterns
must be modeled in the grammar, in a way that accounts for the following facts: existing
words with nasal-substituting affixes are listed; speakers rarely perform nasal substitution
on novel words or rare words; and listeners prefer nasal substitution on voiceless
obstruents over voiced, and front over back.
The basic model that I will propose involves high-ranking input-output
correspondence constraints that cause established words to be pronounced as listed, with
lower-ranked markedness constraints that come into play when no listed form is available
(as with a novel word). This section presents the constraints involved in nasal
substitution, and shows how they interact to produce novel utterances and to produce
utterances based on listed words. Subsequent sections show that the grammar proposed is
learnable from the lexical data, that the grammar predicts appropriate behavior by both
speakers and listeners, and that the interaction of speakers and listeners maintains lexical
patterns.

2.4.2. Paradigm Uniformity

Paradigm Uniformity, also known as Output-Output Correspondence, enforces similarity

among related words (Crosswhite 1996 and 1998, Steriade 1996, Kenstowicz 1997,
Benua 1998). For any word, there are potentially many other words to which it could be
seen as related: ma-miga@j is clearly related to the bare-stem word biga@j, perhaps related
to other derivatives of the stem biga@j, and perhaps even related to other words with the
prefix maN-. It is clear that nasal substitution reduces similarity between the nasal-

45
substituted word and unsubstituted derivatives of the same stem, including the bare stem,
violating Output-Output Correspondence constraints. I will use PU as a shorthand for
those correspondence constraints that enforce similarity between an unsubstituted stem
like biga@j and the substituted form of that stem found in ma-miga@j and are violated by
nasal substitution (e.g., IDENT-OO[SONORANT], IDENT-OO[VOICE] for voiceless-initial
stems). Candidates with nasal substitution violate PU once, and candidates without nasal
substitution do not violate PU.

2.4.3. Input-Output Correspondence

PU is one of the forces that discourage substitution in novel words. Input-Output

Correspondence is part of the force that allows substitution in words that are listed as
substituted (USELISTED, discussed below, is the other crucial part).
Input-Output Correspondence enforces similarity between an input and an output,
and thus encourages substitution if the input is a substituted word, but discourages
substitution if the input is an unsubstituted word or a prefix+stem combination. Adopting
the view of Lapoliwa (1981), Newman (1984), and Pater (1996), nasal substitution is a
coalescence of two segments, as illustrated in (26).

(26) Nasal substitution as coalescence

/ m1 a2 N3/ + / b4 i5 g6 a7 j8 /

[ m1 a2 m3,4 i5 g6 a7 j8 ]

Matching subscripts indicate that a segment in the output is the correspondent of a

segment in the input, so /maN3/+/b4igaj/ → [mam3,4igaj] means that the surface segment
[m3,4] corresponds to both the input segment /N3/ and the input segment /b4/. The

46
coalescence analysis allows output [m3,4] to straightforwardly inherit some of the features
of /N3/ (manner features) and some of the features of /b4/ (place features). If one of the
input segments were actually deleted, the analysis would be more complicated, requiring
constraints that preserve the features of an input segment even if that segment is not
present in the output.
Coalescence can produce featural misidentity between the prefix nasal and the
coalesced nasal—/N3/ is [dorsal], but [m3,4] is [labial]—and between the underlying stem-
initial obstruent and the coalesced nasal—/b4/ is [-sonorant], but [m3,4] is [+sonorant];
thus, nasal substitution violates IDENT-IO constraints. Coalescence also alters the
precedence relations between segments in the underlying string: in the input, segment 3
strictly precedes segment 4, but in the output, it does not.
There is a difference, though, between substitution of a synthesized prefix+stem
combination and substitution of an unsubstituted listed word (if that listed word is a
phoneme string—see §2.5 for consideration of other possibilities). In /maN3/+/b4igaj/ →
[mam3,4igaj] , the precedence relation that is interrupted is between segments that do not

belong to the same lexical entry (/N3/ and /b4/); within the prefix and within the stem, all
precedence relations are preserved. If coalescence applies to a single listed word,
however, as in /mam3b4igaj/ → [mam3,4igaj] , however, the precedence relation that is
disturbed is between two members of the same lexical entry. Pater (1996) differentiates
between LINEARITY, which is violated by any coalescence, and ROOTLINEARITY, which is
violated only by coalescence within a root. I will instead make the distinction between
MORPHORDER, which is violated by disturbing the linear order of morphemes (such as by
coalescing members of two different morphemes) and ENTRYLINEARITY, which is
violated by disturbing the linear order of segments (as by coalescence) within a lexical
entry coalescence:

47
(27) Constraints against coalescence
MORPHORDER
If morpheme µ1 precedes µ2 in the input, then all the segments of µ1 must
precede all the segments of µ2 in the output.
ENTRYLINEARITY
If segment X precedes segment Y within a lexical entry, A is the output
correspondent of X, and B is the output correspondent of Y, then A must
precede B.

Pater justifies ROOTLINEARITY by the fact that roots often contain a richer
contrast set than affixes, but it could also be seen as justified by work such as Cho (to
appear), which suggests that timing relations between gestures belonging to different
morphemes are much more variable than timing relations between gestures belonging to
the same morpheme, implying that violating timing relations such as precedence within a
lexical entry is more strongly avoided than violating timing relations across lexical
entries.
The table in (28) summarizes the role of Input-Output Correspondence in nasal
substitution by showing the CORR-IO31 violations of a variety of input-output pairs.

31
“CORR-XY” stands for any constraint affecting correspondence between X and Y (IDENT[F]-XY, MAX-
XY, DEP-XY, etc.)

48
(28) Corr-IO constraints: sample violations
‘to biga@j’ IDENT IDENT DEP MAX MORPH ENTRY
[PLACE] [SON] ORDER LINEARITY
/maN3/+/b4igaj/ → [mam3,4igaj] * * *
/maN3/+/b4igaj/ → [mam3igaj] * *
/maN3/+/b4igaj/ → [mam4igaj] * *
/maN3/+/b4igaj/ → [mam3b4igaj] *
/mam3i4gaj/ → [mam3i4gaj]
/mam3i4gaj/ → [mam3bi4gaj] *
/mam3i4gaj/ → [mam3b3i4gaj] * *32
/mam3b4igaj/ → [mam3,4igaj] * *
/mam3b4igaj/ → [mam3igaj] *
/mam3b4igaj/ → [mam4igaj] * *
/mam3b4igaj/ → [mam3b4igaj]

2.4.4. Listedness

This section introduces a constraint USELISTED, which requires that a single lexical entry
be used as input (rather than a prefix+stem combination). If no such entry is available,
USELISTED is irrelevant, because it is violated by all candidates, but if such an entry is
available, USELISTED requires that it be used.
It is usually assumed that the input to a tableau is a particular lexical entry or
combination of lexical entries; CORR-IO constraints evaluate each output candidate’s
faithfulness to that one input. I will assume instead (as in (28)) that each candidate is an
input-output pair—different candidates can have different inputs—and CORR-IO
constraints evaluate correspondence within each pair. The real “input” to a tableau that is
shared by all candidates is the morphosyntactic and semantic features that the speaker
wishes to express, which I will call the intent; there may be more than one lexical item or
combination of lexical items that could express that intent. This means that Gen, the
component of the grammar that generates the candidate set, must generate a complete set

32
“Splitting” a segment can be thought of as a violation of ENTRYLINEARITY, because in the input, segment
3 does not precede itself, but in the output, it does.

49
of outputs for each input that is available in a given tableau. As in (28), two distinct
candidates may share the same33 output, but have different inputs.
USELISTED enforces a preference for candidates whose inputs consist of a single
lexical entry, rather than a string of morphemes:

(29) USELISTED
The input portion of a candidate must be a single lexical entry.
(1 violation if not true)34

The tableaux in (30) illustrate the operation of USELISTED. I assume that high-
ranked constraints enforce morphosyntactic and semantic identity between intent and
output, preventing some unrelated lexical entry or prefix+stem combination from being
used.35 In (30), these constraints are included in the shorthand constraint MEANING, which
I omit from subsequent tableaux. In the first tableau, candidate a, which uses a single
lexical entry, satisfies both MEANING and USELISTED. Candidate b satisfies MEANING,36
but violates USELISTED, because it uses a combination of two lexical entries. Candidate c

33
The outputs are not exactly the same, because their segments are in correspondence with the segments of
different inputs.
34
It might be desirable to make USELISTED sensitive to the number of lexical entries beyond a binary
one/many opposition (i.e., preferring a candidate that uses a lexicalized prefix-stem combination plus a
suffix over a candidate that concatenates prefix+root+suffix afresh), but the constraint as defined will
suffice for present purposes.
35
Or perhaps the restriction is in GEN (the function that generates the set of candidates) itself. Using high-
ranking constraints instead is attractive, though, because it allows speech errors in which the wrong input is
(e.g., deviant for devious) to be described as the result of very rare rankings (see §2.4.7).
36
A prefix+stem combination does not completely satisfy MEANING when the meaning of the existing
single lexical entry is idiosyncratic but it satisfies what would presumably be the highest-ranking MEANING
constraints. For example, if a speaker wants to talk about a rudder (for which there is a listed word,,
/pamalik/ ‘rudder’), her linguistic intent is not perfectly satisfied if she synthesizes /paN/+/balik/ (however
she decides to pronounce it), which should mean just ‘tool for returning’. But /paN/+/balik/ would satisfy
her intent better than an input that lacked the meaning ‘tool’, or was not a noun, or meant ‘tool for digging’.

50
uses a single lexical entry, but it violates MEANING, because it is not Actor-Focus (it is
Patient-Focus). Candidate d violates MEANING because it is [-distributive] (it would
simply mean ‘to give’). In the second tableau, bugnat is a novel stem, and so there is no
lexical entry /mamugnat/ available, and all possible candidates violate USELISTED.

(30) Violations of USELISTED

Intent: V, Actor-Focus ‘to distribute’ MEANING USELISTED
(a) /mamigaj/ → [mamigaj]
(b) /maN/+/bigaj/ → [mamigaj] *
(c) /?ipamigaj/ → [?ipamigaj] *
(d) /mag/+/bigaj/ → [magbigaj] * *

Intent: V, Actor-Focus ‘to bugnat’ MEANING USELISTED

(e) /maN/+/bugnat/ → [mamugnat] *
(f) /mag/+/bugnat/ → [magbugnat] * *

Are all lexical entries equally available? Surely the leap during word-learning
from unknown word to fully available lexical entry is not instantaneous. More-frequent
words seem to have stronger lexical entries—they are, for example, faster to recognize
(Rubenstein, Garfield, & Millikan 1970; Forster & Chambers 1973). Frisch (to appear)
reports experimental results in which subjects who were exposed to a novel word twice
rated it more “word-like” than subjects who were exposed to a novel word just once,
suggesting that a word is not immediately accepted the first time it is heard. The model
here assumes that rather than simply being listed in the mental lexicon or not, lexical
entries range in strength from 0 (not at all listed) to 1 (always available for use). Strength
of a lexical entry in this model is a function of the number of times a speaker has heard
the word, although in real life there are probably other factors, such as who the speaker
has heard the word from and in what context.

51
 There are two ways of implementing “gradient listedness” in the grammar. One is
to replace USELISTED with a family of inherently ranked constraints such as

USE100%LISTED >> USE90%LISTED >> ... >> USE10%LISTED >> USELISTED

where USEX%LISTED is satisfied by a candidate whose input lexical entry is X% listed or

more. Other constraints could be inserted into this hierarchy. For example, if

USE40%LISTED >> PU >> USE30%LISTED

a nasal-substituted derivative with 30% listedness (i.e., of whose listedness the speaker is
30% certain, or whose lexical entry’s strength is 30% of the maximum possible strength)
or lower will not be used because Paradigm Uniformity to the base forbids nasal
substitution. But a nasal-substituted derivative with 40% listedness (or higher) would
override PU and be used. This is illustrated schematically in (31): Candidate a, the
faithful parse of the single lexical entry, fails because it violates PU; candidate b satisfies
PU, but violates CORR-IO. Candidate d is the optimal candidate because, although it
violates USE30%LISTED, it satisfies PU, which is more highly ranked. But in the second
half of the tableau, candidates are available that satisfy USE40%LISTED, and so candidate
(e) is optimal despite its violation of PU.

(31) Interaction of a family of USEX%LISTED constraints and Paradigm Uniformity

ENTRY USE40% PU USE30%
LINEARITY LISTED LISTED
(a) /manala/ (30% listed) → [manala] * *!
(b) /manala/ (30% listed) → [mantala] *! *
(c) /maN/+/tala/ → [manala] * *! *
(d) . /maN/+/tala/ → [mantala] * *
(e) . /manili/ (40% listed) → [manili] *
(f) /manili/ (40% listed) → [mansili] *!
(g) /maN/+/sili/ → [manili] *! * *
(h) /maN/+/sili/ → [mansili] *! *

52
 The other way to approach gradient listedness is to have a single constraint,
USELISTED, with the availability of a given lexical entry in any utterance equal to the
listedness of the entry. For example, a word that is 30% listed has a 30% probability of
being available in any given tableau. The ranking CORR-IO, USELISTED >> PU produces
the result /manala/ → [manala] 30% of the time (upper tableau in (32)—listed /manala/
is available as an input), and the result /maN/+/tala/ → [mantala] 70% of the time
(lower tableau in (32)—synthesized candidates only).

(32) Interaction of a unitary USELISTED constraint and Paradigm Uniformity

ENTRY USE PU
LINEARITY LISTED
(a) . /manala/ → [manala] *
(b) /manala/ → [mantala] *!
(c) /maN/+/tala/ → [manala] *! *
(d) /maN/+/tala/ → [mantala] *!
(g) /maN/+/tala/ → [manala] * *!
(h) . /maN/+/tala/ → [mantala] *

In contrast, we would see /manili/ → [manili] 40% of the time, and /maN/+/tili/
→ [mantili] 60% of the time. This may seem like an obvious empirical difference
between the unitary-USELISTED approach and the USEX%LISTED approach, which
produced uniformly /manala/ → [manala] and uniformly /manili/ → [manili] in (31),
but under the stochastic constraint ranking scheme introduced below, the difference is not
so clear. For that reason, I will use unitary USELISTED.

2.4.5. Constraints specific to nasal substitution

Nasal substitution is some 5000 years old (see fn. 83). The original phonetic motivation
might have been consonant-cluster avoidance, as suggested in Archangeli, Moll, and
Ohno 1998; post-nasal lenition; or an attempt to avoid a non-crisp edge (prefix nasal and
stem-initial consonant sharing place of articulation, as required by nasal assimilation), as

53
suggested in Pater 1999b. I suspect that modern Tagalog nasal substitution is divorced
from any phonetic or prosodic motivations, and simply exists as an arbitrary alternation.37
Accordingly, I will propose a constraint, NASSUB (short for “nasal-substitute”) that
simply requires nasal substitution:

(33) NASSUB38

* W
µ µ
| |
X X
| |
[+nasal] [-sonorant]

A morpheme-final nasal must not be immediately followed by an obstruent within

the same word. 39

I will assume that NASSUB penalizes failure to substitute in both synthesized

prefix+stem candidates, and in candidates whose input is a single listed word.40 This is
because even although a morphologically complex lexical entry like /mamigaj/ contains
no morpheme boundaries, its segments are coindexed to related lexical entries: the first

37
Note that the prefixes mag- and pag- also produce consonant clusters—and, with velar-initial stems, non-
crisp edges (e.g., mag-kilati@ùs-an ‘to appraise each other’ from kila@ùtes, kila@ùtis ‘carat’), unless some
mechanism requires the g and k to have separate-but-identical features. But these prefixes do not induce
coalescence, even though the identity violations would be no worse those incurred in nasal substitution.
38
Representations in constraint definitions should be interpreted as nonexhaustive at the edge of each tier.
For example, in (33), other morphemes may come before or after the two shown, but not between. When
tiers are missing, the information on those tiers should be considered irrelevant. For example, in (34), the
two segments may belong to different morphemes or to the same morpheme.
39
Where “word” must be defined so as to exclude the compounding-like prefixes discussed in fn. 14, which
never trigger nasal substitution.
40
Although this assumption is not crucial to the model proposed here—once a word is listed as
unsubstituted, ENTRYLINEARITY almost always prevent NASSUB from having any effect.

54
three segments (mam) are coindexed with the segments of the lexical entry for the prefix
/maN-/, and the last five segments (migaj) are coindexed to the segments of the lexical

entry for the word /bigaj/. The candidate /mamigaj/ → [mamigaj] satisfies NASSUB,
because there is no sequence of a distinct nasal and obstruent coindexed to two different
morphemes.
Turning to the constraints that produce the patterns in the lexical distribution of
nasal substitution,41 I attribute the higher rate of substitution on voiceless-initial stems to
a constraint *NC8, a constraint forbidding a sequence of a nasal and a voiceless obstruent:

(34) *NC8

* W
X X
| |
[+nasal] [-voice, -sonorant]

A [+nasal] segment must not be immediately followed by a [-voice, -sonorant]

segment within the same word.

Hayes and Stivers (1996) give a phonetic motivation for *NC8: the raising of the
velum during the nasal-to-oral transition expands the oral cavity, slowing the buildup of
the supraglottal air pressure that would otherwise “turn off” voicing. An NC8 sequence
thus requires extra effort (such as glottal abduction) to keep the obstruent voiceless.

41
This is a form of Emergence of the Unmarked (McCarthy & Prince 1994): although nasal substitution
itself is not motivated by pure markedness, the patterns in its distribution seem to reflect considerations of
markedness.

Newman (1984) finds an implicational hierarchy reflecting similar effects in related languages in
which nasal substitution is predictable if the stem-initial obstruent is known: If the language substitutes g, it
also substitutes d, and if a language substitutes d, it substitutes b; similarly, substitution on k implies

55
Hayes and Stivers propose that the articulatory difficulty of NC8 clusters drives postnasal
voicing. Pater (1996) discusses *NC8 as the motivation for postnasal voicing, Indonesian
nasal substitution (which applies only to voiceless obstruents), nasal deletion,
and denasalization.42
*NC8 favors substitution in voiceless-initial stems. A word like mantukad, without
substitution, violates *NC8, but manukad, with substitution, does not. *NC8 is irrelevant for
voiced-initial stems, since it is violated by neither substitution (mandukad) nor
nonsubstitution (manukad).
If *NC8 is ranked high enough to produce an effect in nasal substitution, why is
*NC8 violated so freely word-internally? One answer is the distinction made above
between MORPHORDER and ENTRYLINEARITY:

(35) Coalescence within vs. across listed items

/maN1/+/p2ili/ ENTRY *NC8 MORPH
LINEARITY ORDER
. mam1,2ili *
mam1p2ili *!
/ban1t2a@ùj/
ban1,2a@ùj *!
. ban1t2a@ùj *

substitution on t,s and p. In addition, substitution on b implies substitution on p, d on t,s, and g on k. Thanks
to Joe Pater for pointing out this interesting finding.
42
Pater 1999b proposes instead that Alignment is the driving force behind Indonesian nasal substitution,
and that IDENT-IO for pharyngeal expansion (see Steriade 1995) is what restricts nasal substitution to
voiceless obstruents: voiced obstruents require pharyngeal expansion to maintain transglottal airflow
despite a vocal tract obstruction, and so are [+pharyngeal expansion], but voiceless obstruents—which lack
transglottal airflow—and nasals—which lack a vocal-tract obstruction—are [-pharyngeal expansion]. So,
fusing a voiced obstruent and a nasal violates IDENT[PHARYNGEAL EXPANSION], but fusing a voiceless
obstruent and a nasal does not. This approach might work for Tagalog as well (with stochastic constraint
ranking).

56
 If ENTRYLINEARITY is very highly ranked, *NC8 will not be able to shape the
lexicon root-internally the way it seems to have done for nasal substitution (by a
mechanism proposed below).43
To introduce the constraints that produce the place-of-articulation effect, consider
the chart in (36), showing the distribution of various consonants in various positions
within the root in Tagalog. The numbers are from a database of about 4,600 disyllabic
Tagalog roots,44 with reduplicated roots excluded.

43
But see fn. 42: adopting Pater’s (1999b) approach to Indonesian, the voicing effect would be driven by a
difference in faithfulness (rather than a difference in markedness), in which case there is no drive to
coalesce nasal-obstruent clusters word-internally. Under the learning mechanism discussed below, though,
there is no way to prevent *NC8 from being learned with a fairly high ranking, so we would still have to rely
on ENTRYLINEARITY to prevent root-internal coalescence.
44
All the native, disyllabic roots in English 1986 were recorded. The count shown is by type—each root is
counted just once, no matter how many affixed forms it has. The restriction to disyllabic roots is necessary
because monosyllables are clitics (pronominal and discourse), which may not obey the same morpheme
structure constraints as lexical roots, and roots of more than two syllables are—at least historically—
polymorphemic. Because of evidence that speakers may treat words that appear polymorphemic as
polymorphemic, even without morphosyntactic motivation (see Baroni 1998, Hammond 1999, and Chapter
4 of this dissertation), words with more than two syllables might therefore also escape root structure
constraints.

57
(36) Distribution of consonants in roots of the form C1V(C2)C3V(C4)

C1 C2 C3 C4

100%
59
98 61
258
90%
201 184
421 353 355 265
80% 178
234 169
575
335 383

70% 172
307

614
60% 227
245
368 301 253
27 247
50%
31 174
50
40% 39 159 66
147 192
41 236 197
30% 159 275
452
354 264
463
20% 503 522
496
165
321 218

10% 167 9
73
52 32
32 46
0%
p t s k b d/r g m n N l j w ? h

58
 Note that in general, fronter consonants are better represented root-initially.45 For
example, about 45% of ps are root-initial (C1), but only about 28% of ks are root-initial.
Note further that obstruents are better represented initially than are sonorants. There are
very few root-initial nasals, both over all and as a proportion of nasals in all positions;
among the nasals, m is better represented root-initially than n or N. This consonantal
distribution suggests that (and would provide evidence to the learner that) root-initial
nasals are disfavored, but that among the root-initial nasals, the fronter ones are less
disfavored.
I propose the following family of constraints against root-initial nasals: *[rootN,
*[rootn, *[rootm (abbreviated *[N, *[n, *[m).

(37) *[N, *[n, *[m

* [root X [root X [root X

| | |
[+nasal, +dorsal] [+nasal, +coronal] [+nasal, +labial]
A root must not begin with [N] ([n], [m]).

This family of constraints disfavors substitution.46 For example, ma-nukad, with

substitution, violates *[n, because the n that results from substitution is root-initial (as
well as prefix-final). But man-tukad, without substitution, does not violate *[n], because
the n belongs to the prefix only. The ranking *[N >> *[n >> *[m (which could be inherent

45
Ingram (1974) proposes “fronting” as an acquisition strategy: a front-to-back order is preferred for both
consonants and vowels within a word (i.e., ...p...t..., ...p...k..., and ...t...k... are preferred to ...t...p..., ...k...p...,
or ...k...t...; ...i...u... is preferred to ...u...i...).
46
in synthetic candidates as well as in candidates with single-lexical-entry inputs, because the n in a lexical
entry /manukad/ would be coindexed to the t of the related word /tukad/. This assumption does not
materially affect the model presented here, however (see fn. 40).

59
or learned) would disfavor substitution most on posterior places of articulation.47 For
example, if *[N >> *[n >> NASSUB >> *[m, then all else being equal, substitution would
occur on a labial-initial root, but not on a coronal- or velar- initial root:

(38) *[N >> *[n >> *[m

/maN/+/bala/ *[N *[n NASSUB *[m
(a) . mamala *
(b) mambala *!
/maN/+/dala/
(c) . mandala *!
(d) manala *
/maN/+/gala/
(e) . maNgala *!
(f) maNala *

Is there any functional motivation for dispreferring root-initial nasals, or for

especially dispreferring root-initial back nasals? Among voiceless obstruents, the place
effect could be seen as a fine-tuned version of *NC8. Recall that the phonetic motivation
proposed by Hayes and Stivers (1996) for *NC8 is that the expansion of the oral cavity
during velum-raising encourages voicing. Their model also found that frontness of the
obstruent encourages voicing, because there is a greater expanse of flexible cheek wall
that can expand outward and reduce supralaryngeal pressure. This would explain why p
substitutes more often than k. But it does not explain why b substitutes more often than d,
since turning off voicing is not necessary in mb, nd, and Ng clusters—indeed, the
frontness of b would make voicing easier to maintain, 48 and thus the cluster mb would be
less marked (and so less subject to repair by coalescence) than nd or Ng.

47
Cf. English, in which root-initial N is not permitted at all.
48
See Ohala & Riordan (1980), who found that passive cavity expansion maintained voicing longer for b
than for d or g.

60
 Another possibility, expanding on Pater 1999b (see fn. 42), is that IDENT-IO
violations are greater when substituting a fronter consonant. Pater proposes that the
reason voiced obstruents do not substitute in Indonesian is that if they did, it would
violate IDENT-IO[PHARYNGEAL EXPANSION]: voiced obstruents are [+pharyngeal
expansion]—they require active expansion of the pharynx, or some other exertion, to
maintain voicing—but nasals are [-pharyngeal expansion], because voicing is maintained
by venting air out the nose. Fronter consonants should require less pharyngeal expansion,
because more cheek area is available for passive expansion, and so coalescing a b with a
nasal is less of a violation of (some gradient version of) IDENT-IO[PHARYNGEAL
EXPANSION]. The place effect among voiceless consonants is then a puzzle, though,
because no voiceless consonants require any pharyngeal expansion.
Whatever the reason, the Tagalog lexicon manifests a dispreference for root-
initial nasals, so I will simply assume the *[NASAL constraint family. Although there may
be a reason for the family to be inherently ranked *[N >> *[n >> *[m, this ranking is
learnable from the lexicon (see §2.6), so it need not be assumed.

2.4.6. Summary of constraints

The table in (39) summarizes the constraints relevant to determining whether a word is
pronounced with nasal substitution.

61
(39) Constraints affecting nasal substitution
Constraint Effect
PARADIGM UNIFORMITY discourages N.S. (nasal substitution)
NASSUB encourages N.S.
*NC8 encourages N.S. for voiceless-initial stems
*[N discourages N.S. for velar-initial stems
*[n discourages N.S. for coronal-initial stems
*[m discourages N.S. for bilabial-initial stems
MORPHORDER discourages N.S. in prefix+stem concatenations
ENTRYLINEARITY encourages N.S. if word is listed with substitution
discourages N.S. if word is listed without substitution.

As noted above, if a word has a listed form, and it is available, and

ENTRYLINEARITY is ranked high, the word will be pronounced as listed:49

(40) Input-Output Correspondence requires use of listed form

ENTRY USE *[N *NC8 *[n NAS PU MORPH *[m
LIN LISTED SUB ORDER
(a) . /mambu@ùla?/ → *
mambu@ùla?
(b) /mambu@ùla?/ → *! * *
mamu@ùla?
(c) /maN/+/bu@ùla?/ → *! *
mambu@ùla?
(d) /maN/+/bu@ùla?/ → *! * * *
mamu@ùla?

ENTRY USE *[N *NC8 *[n NAS PU MORPH *[m

LIN LISTED SUB ORDER
(e) . /mamalsa@/ → * *
mamalsa@
(f) /mamalsa@/ → *! *
mambalsa@
(g) /maN/+/balsa@/ → *! * * *
mamalsa
(h) /maN/+/balsa@/ → *! *
mambalsa@

49
Candidate f in (40) results from splitting underlying m into m and b. Epenthesizing the b instead would
produce a homophonous candidate (not shown) that satisfies ENTRYLINEARITY but violates high-ranking
DEP.

62
 But when no listed form is available, as in a novel word, ENTRYLINEARITY is
satisfied by all candidates, and USELISTED cannot be satisfied by any candidate, so both
are irrelevant; the lower-ranked constraints decide. The tableau in (41) illustrates how the
constraint ranking in (40) would treat a novel root beginning with each obstruent.

(41) Coining of novel words, using the ranking in (40)

maN- form of ENTRY USE *[N *NC8 *[n NAS PU MORPH *[m
/pala/ LIN LISTED SUB ORDER
(a) . /maN/+/pala/ → * * * *
mamala
(b) /maN/+/pala/ → * *! *
mampala
(c) . /maN/+/tala/ → * * * *
manala
(d) /maN/+/tala/ → * *! *
mantala
(e) . /maN/+/sala/ → * * * *
manala
(f) /maN/+/sala/ → * *! *
mansala
(g) . /maN/+/kala/ → * * *
maNkala
(h) /maN/+/kala/ → * *! * *
maNala
(i) . /maN/+/bala/ → * * * *
mamala
(j) /maN/+/bala/ → * *!
mambala
(k) . /maN/+/dala/ → * *
mandala
(l) /maN/+/dala/ → * *! * *
manala
(m) . /maN/+/gala/ → * *
maNgala
(n) /maN/+/gala/ → * *! * *
maNala

Under this ranking, in which PU is fairly low, the ranking of *NC8 with respect to
the three anti-root-initial-nasal constraints (*[N, *[n, *[m) creates a place-of-articulation
cutoff among the voiceless obstruents; in this case, labials and coronals substitute, and

63
dorsals do not. The ranking of NASSUB with respect to the three nasal constraints places a
cutoff among the voiced obstruents; in this case, only labials substitute.

2.4.7. Stochastic constraint ranking

Of course, this cannot be the constraint ranking for the language, because not all novel b-
initial stems (for example) were substituted in the experiment. There is no one ranking
that would be compatible with the experimental results above on novel stems, because for
every consonant tested, there were some tokens in which speakers substituted it, and
some in which they did not.
For this reason, I will adopt stochastic constraint ranking, as proposed in Hayes
and MacEachern 1998, Boersma 1998, Boersma and Hayes 1999, and Hayes (to appear).
Stochastic constraint ranking is similar to variable constraint ranking (as in Anttila 1997).
In Anttila’s system, certain ranking pairs within a hierarchy are fixed, and all ranking
permutations of the constraints that respect those fixed pairs are equally possible. For
example, with constraints C1, C2, C3, and C4 and the ranking C1 >> {C2, C3} >> C4, there
is a 50% probability of speaker’s using the ranking C1 >> C2 >> C3 >> C4 in any given
utterance, and a 50% probability of using C1 >> C3 >> C2 >> C4.
Stochastic constraint ranking differs from variable constraint ranking in that rather
than having only two types of ranking between any two constraints (completely fixed and
completely free), any ranking is possible, but some are more probable than others. This is
implemented by assigning each constraint a probability distribution centered on a
particular ranking value. In any given utterance, an actual value is generated for each
constraint, at random but in accordance with the constraint’s probability distribution.50

50
And, in Boersma’s system, using a quantity called “ranking spread”. Full details are given below, in “The
Speaker”.

64
The dominance relations in the constraint hierarchy are determined by these actual
values. For example, consider the hypothetical constraint system in (42). C1 has a fairly
high ranking value, C2 and C3 are somewhat lower, and C4 is quite a bit lower.

(42) Hypothetical constraint system

C1 C2 C3 C4
probability density

high ranking low ranking

In nearly all of the linear rankings that would be produced by this system on
various occasions, C1 outranks the other three constraints, because its distribution is
centered on a much higher ranking value. This means that it would be possible, but
vanishingly unlikely,51 for C1 to be ranked low enough, and/or any other constraint to be
ranked high enough, for C1 to be dominated. Similarly, it is very improbable that C4 will
outrank any other constraint. But C2 and C3 overlap considerably, which means that their
ranking with respect to each other varies quite a bit. This system is different, however,
from an Anttila-style C1 >> {C2, C3} >> C4 system in that it encodes a weak tendency for
C2 to outrank C3 rather than completely free ranking between the two.
Stochastic constraint ranking allows us to model a situation in which nasal
substitution rarely occurs in any novel word, but it is more likely to occur on a voiceless-

51
See §2.7 for calculations of probability.

65
or front-initial stem: PU and MORPHORDER will tend to prevent substitution, but
substitution will occur on a voiced-initial segment whenever NASSUB outranks PU and
the relevant *[NASAL constraint, and on a voiceless-initial segment whenever either
NASSUB or *NC8 outranks PU and the relevant *[NASAL constraint. This means that there
are more rankings under which, say p would substitute than b, making it more likely that
p will substitute. As for the place effect, if *[N tends to outrank *[n, which in turn tends to

outrank *[m, it is more likely that NASSUB (or *NC8, if relevant) will outrank *[m,
allowing substitution, than that it will outrank *[n or *[N. The following sections show
how such a constraint system would be learned and used.

66
 2.5. Representations: encoding exceptionality

It was argued in §2.2.3 that potentially nasal-substituting words must have their own
lexical entries, both to ensure that the word is reliably substituted or unsubstituted, as the
case may be, and to list additional unpredictable information, such as stress shifts and
opaque meanings.52 An equivalent53 approach would be for every stem to list the
unpredictable information about its derivatives, as in (43).

(43) Sample lexical entry for stem-listing approach (cf. (16))

[bugbo@g], Noun, ‘wallop’

derivative phonological notes semantic notes
paN- (tool for doing X) [+nasal subst.] when washing clothes
paN+REDCV- (act of doing X) [-nasal subst.]
maN- (to perform an X) [-nasal subst.]

This section considers some other alternatives to full listing: substitution

diacritics, underspecification, and allomorph listing. All three will be discussed in terms
of separate lexical entries for each derivative of a stem, but could also be combined with
the stem-listing approach (for example, (43) lists substitution diacritics in the stem’s
subentries).

52
The only exception would be variably pronounced words with no other unpredictable semantic or
phonological characteristics. Section 0 takes up the question of whether a three-way distinction can be
captured without listing all existing words.
53
equivalent for present purposes, that is. This stem-listing approach and full listing might make different
predictions about behavior in lexical access tasks.

67
2.5.1. Substitution diacritics

Rather than a full string of phonemes, a derived word’s lexical entry could consist of a
string of morphemes, plus diacritics indicating additional unpredictable information, such
as nasal substitution (see the discussion of diacritic-based exceptionality in §1.1.1.1).54
This approach shares properties of full listing (each word has its own lexical entry) and
stem-listing (only unpredictable information is listed). We could assign the special
diacritic to nonsubstituting words, to substituting words, or to both. If the diacritic is
applied only to substituting words, we need some mechanism to distinguish between
listed, nonsubstituting words and novel words—that is, we must ensure that a listed,
diacritic-less word (almost) never undergoes substitution, whereas a novel word (also
diacritic-less) may well undergo it. Similarly, if only nonsubstituting words bear the
diacritic, we need a mechanism to distinguish the behavior of a diacritic-less listed word
(which must undergo substitution) and a novel word (also diacritic-less, which may or
may not substitute).
Absent such a mechanism, every word that is consistently substituted or
consistently unsubstituted must bear the diacritic [+NasSub] or [-NasSub]. To make the
grammar sensitive to the difference, the constraint NASSUB could be split into two
constraints (high-ranked NASSUB[+] and low-ranked NASSUB[-]), or its definition could be
modified so that it does not apply to [-NasSub] words.55

54
The presence of the diacritic would make a word subject to special constraints or to a special constraint
ranking.
55
Restricting NASSUB to only [+NasSub] words would not work, because NASSUB must be able to apply to
newly coined words, which would not have any diacritic. Variable words might be words that lacked a
diacritic.

68
 The diacritics approach is equivalent, for present purposes, to full listing: novel
words’ behavior is variable and depends solely on the grammar; the lexicon determines
the behavior of established words.

2.5.2. Underspecification

The underspecification approach of Inkelas, Orgun, and Zoll 1997 (see §1.1.1.1) assigns
a fully specified feature matrix to a segment that resists an alternation (Faithfulness
constraints preserve the underlying feature values no matter what), and an underspecified
feature matrix to a segment that does alternate (Markedness constraints fill in context-
appropriate feature values).
Underspecification might work well if all the derivatives of a single stem behaved
uniformly: representations for a hypothetical nonsubstituting stem palid (with full
specification) and a hypothetical substituting stem pilad (with underspecification) are
shown in (44). Faithfulness constraints would prevent [-nasal] segments from merging
with prefix-final N, but [0nasal] segments would be free to merge.

(44) Partial lexical entries for underspecification approach

palid Pilad
| |
[-nasal] [0nasal]

Because multiple features are involved, the underspecification approach would

also need to ensure that when the P in /Pilad/ becomes [+nasal], it also becomes [+voice],
[+sonorant], and so on, and that a [-voice] specification does not prevent coalescence into
a nasal.56

56
Nasal-initial stems (which would be [+nasal]) are also a problem. As discussed in §2.2.1, it is unclear
whether or not they can undergo substitution, but it is clear that sometimes they do not (e.g., paN-marka@
‘marker’). Because IDENT-IO[NASAL] could not prevent substitution on a [+nasal] segment, MORPHORDER

69
 But in any case, as discussed in §2.2.3, a stem’s derivatives do not behave
uniformly. The underspecified/fully specified contrast, then, would be implemented in
the derived words themselves, which buys little, since an underspecified segment like the
P in /maNPilad/ would always be in the same context (nasal-substituting).

Another use of underspecification would be for novel words: the initial obstruents
of stems themselves could be underspecified ([0nasal]), so that when stems were
combined for the first time with a substitution-inducing prefix, it would be up to the
grammar to determine whether or not nasal substitution would apply: MORPHORDER and
the *[NASAL constraints would discourage substitution; NASSUB and *NC8 would
encourage it. The stem-initial segments of existing derived words, on the other hand,
would be fully specified as [-nasal] if unsubstituted and [+nasal] if substituted, and high-
ranking IDENT-IO[NASAL] would preserve the underlying feature values. Again, this
version of underspecification would be largely equivalent to full listing.

would have to somehow be formulated or parametrized so as to prevent substitution on [+nasal] segments

but not on [0nasal] segments.

70
2.5.3. Allomorph listing

The final approach to be considered is allomorph listing. If the derivatives of a stem

behaved uniformly, we might say that a nonsubstituting stem had just one allomorph—
continuing the example from (44), /palid/—whereas a substituting stem had two—/pilad/
and /milad/.57 For stems with two allomorphs, the best one would be selected according to
context (N-final prefix or not—the prefix would also have to have two allomorphs).
Adapting the allomorphs approach to the unpredictable behavior of a stem’s
derivatives, we could let each derivative’s lexical entry specify which allomorph it
selects. In this case, the only empirical difference between a stem with no nasal-
substituted allomorph and a stem with a substituted allomorph that no derivatives happen
to select would be that novel derivatives of the first kind of stem would most likely be
unsubstituted at first—a substituted allomorph might later develop—because a
substituted pronunciation could arise only from the grammar. Novel derivatives of the
second kind of stem would be more likely to substitute, since a substituted pronunciation
could arise either from the grammar or from selecting the existing, substituted
allomorph.58 Aside from this difference between classes of stems, the allomorphs
approach is equivalent in effect to diacritics for derivatives.

57
Actually, several allomorphs would be necessary in order to deal with other phonology that a derived
word (including potentially nasal-substituted words) might undergo, such as vowel raising with suffixation
(see Chapter 4), syncope (see §4.7.2), and stress shifts.
58
See Steriade 1999 for evidence that the pronunciation of a new derived word depends on the available
allomorphs for the word’s stem

71
 2.6. The Learner

Section 2.4.7 proposed that constraints are stochastically ranked. But “stochastic” does
not mean “freely variable”: the learner must determine ranking values for each constraint,
which will then determine the probability of any particular total ranking of constraints.
This section gives a brief explanation of Boersma’s (1998) Gradual Learning Algorithm,
and then shows what kind of grammar is learned using the constraints introduced and a
mini-lexicon. In particular, I will show how the Gradual Learning Algorithm can rank
constraints even when their presence is unnecessary in tableaux for existing words;
subsequent sections exploit this result.
Boersma’s Gradual Learning Algorithm was designed to learn a stochastic
grammar (see §2.4.7) from variable data. The algorithm is error-driven: it generates
hypothetical outputs, in proportion to the frequencies generated by the constraint ranking
achieved so far. Schematically, a grammar consisting only of the constraints PU and
NASSUB would begin with the two constraints equally ranked. For the input /mamigaj/,
outputs [mamigaj] (correct) and [mambigaj] (incorrect) would each be produced 50% of
the time:

(45) Learning, starting with two equally-ranked constraints

NASSUB >> PU (probability .5)
/mamigaj/ NASSUB PU
. mamigaj *
mambigaj *!
PU >> NASSUB (probability .5)
/mamigaj/ PU NASSUB
, mambigaj *
- mamigaj *!

72
 Learning occurs when an output is incorrect, as in the second tableau (incorrectly
selected candidate indicated by ,; “real” winner, not selected under this ranking,
indicated by -). The constraint violations of the incorrect winner (mambili) are
compared to those of the correct output (mamili) and constraint rankings are adjusted
accordingly: all constraints on which the incorrect output does better than the correct
output are demoted, and all constraints on which the correct output does better than the
incorrect output are promoted. Note that only two candidates are relevant to adjusting the
constraint ranking: the incorrect winner and the correct output. The adjustment does not
take into account the constraint violations of the other candidates, since they were
correctly ruled out by the ranking used. Note also that each candidate is an input-output
pair: the learner does not have to, for example, consider all possible inputs that could
have generated the correct or incorrect output.
In this case, if mambili is incorrectly chosen as the winner, PU is demoted—since
mambili has fewer violations of it than mamili does—and NASSUB is promoted—since

mambili has more violations of it than mamili does. Adjustments are initially large, and

become smaller and smaller as learning progresses, so that as the learner approaches its
“adult” state, the grammar is not very susceptible to change.
I applied the Gradual Learning Algorithm (using Hayes 1999) to a set of
substituted and unsubstituted words, composed of hypothetical stems each with a nasal-
substituting prefix, assuming that each was fully listed as a whole word. The corpus
reflected the numbers of substituted and unsubstituted words in the lexicon59 for all
constructions combined. The table in (46) summarizes the composition of the mini-
lexicon used for learning.

59
Only type frequencies were used, because token frequencies were not available.

73
(46) Mini-lexicon for learning
initial number of words
segment
substituted unsubstituted
p 21 1
t&s 36 3
k 15 1
b 15 8
d 2 6
g 0 8

Along with the correct candidate (the faithful rendering of the lexical entry), each
tableau had three incorrect candidates: the unfaithful rendering of the lexical entry (e.g.,
/pamuntol/ → [pampuntol] , or /paNkundol/ → [paNundol] ), the unsubstituted

prefix+stem (/paN/+/puntol/→ [pampuntol] ), and the substituted prefix+stem

(/paN/+/puntol/→ [pamuntol] ). The constraints used were those given in §2.4.
Since all the words were fully listed, ENTRYLINEARITY and USELISTED together
suffice to select the correct output. On every learning trial in which an incorrect output is
produced, ENTRYLINEARITY or USELISTED is promoted, but adjustment of other
constraints also occurs. For example, if the ranking in (47) is generated, the incorrect
candidate ,/pamuntol/ → [pampuntol] is selected instead of the correct candidate
-/pamuntol/ → [pamuntol] . So, NASSUB, *NC8, and ENTRYLIN must be promoted; PU

and *[m must be demoted.

(47) Sample learning trial

PU NAS *[N *NC8 *[m *[n USE MORPH ENTRY
SUB LISTED ORDER LIN
, /pamuntol/
←* ←* ←*
→ pampuntol
- /pamuntol/
*!→ *→
→ pamuntol
/paN/+/puntol/
* * *!
→ pampuntol
/paN/+/puntol/
*! * * *
→ pamuntol

74
 If the lexical entry in question had instead been /panuntol/, the *[NASAL constraint
to be demoted would have been *[n, and if the lexical entry had been /paNuntol/, *[N
would have been demoted. The proportion of words that are substituted in the mini-
lexicon is higher for labials (36 out of 45 are substituted) and coronals (38 out of 47) than
for velars (15 out of 24). Since *[NASAL constraints are demoted only when the correct
output is substituted (and the grammar instead selects an unsubstituted output), *[m and
*[n are demoted more often than *[N. In other words, even though in the target grammar
the *[NASAL constraints play no role in determining the optimal output, their relative
ranking is learned because the Gradual Learning Algorithm adjusts the rankings of all
constraints on which the correct and incorrect candidates differ.60
When ENTRYLINEARITY and USELISTED climb high enough in grammar that no
more incorrect outputs are generated, learning stops. Therefore, the initial constraint
adjustment increment must be small enough that there is opportunity to learn about the
lower-ranked, seemingly irrelevant constraints before ENTRYLINEARITY and USELISTED
take over.61

60
Although it is clear among voiced obstruents that there is a much higher rate of substitution for [b] than
for [d], the large number of substituted voiceless coronals (the [t]s and the [s]s) obliterates the
labial/coronal distinction. If the mini-lexicon is devised so that each obstruent type is equally represented
(e.g., 10 [p]s, 10 [t]s and [s]s, etc.) and the rate of substitution within each type is reflected, rather than
absolute numbers of substituted words, a sharp ranking difference emerges between *[m and *[n as well as
between those two and *[N.

Evidence for the ranking of the *[NASAL constraints could also come from the distribution of roots
in the lexicon (see (36)), although these were not included in the learning procedure. For example, there are
few roots beginning in /N/, and so there would be few instances in which the learner had to demote *[N
because a candidate that obeyed it (e.g., /Nata/ → [kata]) had mistakenly won; there are more roots
beginning in /m/, and so more instances in which *[m would be similarly be demoted.
61
This seeming inefficiency is not troubling if we consider that in the early stages of learning, the child
may be ill-equipped to guess which words as really listed for adults and which are synthesized, and may not
have enough evidence about the underlying form to know whether ENTRYLINEARITY is ever violated. So
learning that involves USELISTED, ENTRYLINEARITY, and other non-phonotactic constraints should proceed
cautiously.

75
 Using an initial learning increment of 0.1 and a final increment of 0.0001 over
2000 trials produced satisfactory results (in each trial, one output is generated for each
word in the mini-lexicon). The average constraint rankings over twelve such runs are
shown in (48); error bars indicate standard deviations.62

(48) Ranking values arrived at by Gradual Learning Algorithm

115

110
ranking value

105

111.05 111.05
100
104.30 103.53
99.71 100.29 101.07 99.63
95 99.02

90

m
ed

n
N
PU
n

T
er

Su

*[
Li

*N

*[

*[
st

rd
try

as
Li

N
se

ph
En
U

or
M

The following section shows what kind of production behavior occurs with this
grammar.
Note that in the case of nasal substitution, the high ranking of ENTRYLINEARITY is
essential to assuring that listed words are pronounced faithfully. This high ranking is
assured because although different words give conflicting evidence to the learner about
the ranking of most constraints (PU, NASSUB, *[m, etc.), every word gives evidence in
the same direction for ENTRYLINEARITY—the correct candidate always obeys

62
The standard deviation, that is, of the ranking values arrived at over the twelve runs, which could be
imagined as twelve different learners’ exposures to the same data.

76
ENTRYLINEARITY. This result generalizes to other cases of exceptionality: if existing
words’ stable behavior is encoded in some property of their lexical entries, then the
constraint(s) requiring faithfulness to that property will always become high ranked,
because correct candidates always obey them.

77
 2.7. The Speaker

Section 2.6 presented the typical ranking values that a learner arrives at after exposure to
the lexicon. The ranking values determine the probability that a given candidate will be
optimal in a particular tableau, but there is a certain amount of calculation involved. This
section goes through the steps that yield the frequencies at which the grammar predicts
various outcomes for both listed and novel words.

2.7.1. Probability of a candidate’s being optimal

As described in §2.4.7, in the Boersmian model, a constraint ranking is chosen

probabilistically for each utterance, in accordance with the ranking values in the
grammar. Once the ranking is chosen, the optimal output for a given input is fully
determinate. But, in my model, the availability of inputs in a given utterance is also
decided probabilistically (on the basis of Listedness values). Therefore, the probability of
occurrence for any output, given the speaker’s linguistic intentions, depends
probabilistically on both the grammar and the lexicon.
Before giving actual numbers for nasal substitution, some explanation of the
method for calculating these probabilities: The probability of a lexical entry’s being
available is straightforward. As discussed in §2.4.4, it is a function of how many times
the word has been heard, (as well as, ideally, from whom and in what context). §3.4
discusses the function further.

(49) Availability as a function of listedness

P(Available(Entry)) = Listedness(Entry)

If the set of available inputs is known, the probability that a particular input-
output pair will be chosen as optimal is just the probability that a constraint ranking under

78
which that pair is optimal will be generated. The set of such rankings can be determined
by inspecting a tableau. For example, in the schematic tableau in (50), in order for
candidate a to be optimal, it must be superior to both b and c. For a to be superior to b,
a’s violations of C2 and C4 must be outweighed by b’s violations of C1 and/or C3. In other

words, either C1 or C3 must outrank C2, and either C1 or C3 must outrank C4. Similarly,
for a to be superior to c, C1 must outrank C4.

(50) Hypothetical tableau

C1 C2 C3 C4
(a) * *
(b) * *
(c) * *

Any ranking that meets the condition in (51) will produce a as the optimal candidate.

(51) Ranking requirements for candidate a in (50) to be optimal

(C1 >> C2 OR C3 >> C2) AND C1 >> C4

Before showing how to calculate the probability of obtaining a ranking consistent

with complex requirements like those in (51), let us first consider the simplest case, with
only two candidates and two constraints:

(52) Simple hypothetical tableau

C1 C2
(a) *
(b) *

Computing the probability of C1>>C2 is fairly simple and is described in §2.11. In brief,
in a given utterance, each constraint is assigned a “selection point”, or actual value, based
on the constraint’s ranking value in the grammar and a certain degree of random noise.
Therefore, P(Ci>>Cj) depends only on the difference in ranking value between Ci and Cj.

79
Probabilities of Ci>>Cj for integer differences in ranking value from -10 to 10 are given
in (53).

(53) Probability of Ci's outranking Cj in a given utterance

rV(Ci) - rV(Cj) -10 -9 -8 -7 -6 -5 -4 -3 -2 -1
P(Ci>>Cj) .0002 .0007 .002 .007 .02 .04 .08 .14 .24 .36

0 1 2 3 4 5 6 7 8 9 10
.50 .64 .76 .86 .92 .96 .98 .993 .998 .9993 .9998

The situation is more complicated if we want to calculate P(C1>>C2 AND

C1>>C3). We cannot simply multiply (P(C1>>C2) * P(C1>>C3)), because P(C1>>C2)

and P(C1>>C3) are not independent. A method for calculating probabilities of complex
ranking requirements is given in §2.11, with a sample calculation in Mathematica given
in §2.12.
We can now begin to calculate actual probabilities of outcomes from the grammar
learned in §2.6.

2.7.2. Generating a listed form

When a listed word exists, the probability that it will be faithfully used is very high, but
never quite 1. The probability at which unfaithful outcomes occur—or at which the listed
form is ignored in favor of forming the word afresh—is quite low given the grammar
learned in §2.6, low enough to be in the realm of speech errors.
For a listed, substituted form of a p-initial stem with the maN+RCV- prefix
complex (/mamumuntol/), the four outcomes I will consider here are faithful
/mamumuntol/ → [mamumuntol] ; unfaithful /mamumuntol/ → [mampupuntol] ;

unsubstituted, newly formed /maN/+/RCV/+/puntol/ → [mampupuntol] ; and substituted,

newly formed /maN/+/RCV /+/puntol/ → [mamumuntol] :

80
(54) Four candidates for a listed, substituted word
USE ENTRY MORPH *NC8 NASSUB PU-maN *[N *[n *[m
LISTED LIN ORDER +RCV-
/mamumuntol/
* *
→ [mamumuntol]
/mamumuntol/
* * *
→ [mampupuntol]
/maN/+/RCV/+/puntol/
* * *
→ [mampupuntol]
/maN/+/RCV /+/puntol/
* * * *
→ [mamumuntol]

For the faithful output /mamumuntol/ -> [mamumuntol] to occur, (i) the input
/mamumuntol/ must be available; (ii) PU must be outranked by *NC8, or NASSUB, or

USELISTED and ENTRYLINEARITY; and (iii) *[m must be outranked by *NC8, or NASSUB,
or USELISTED and ENTRYLINEARITY. If /mamumuntol/’s listedness is 0.953, for example,
the probability of (i) is 0.953. The joint probability of (ii) and (iii) is 0.9999,63 so the
probability of /mamumuntol/ → [mamumuntol] ’s being the optimal output given that
/mamumuntol/ is 95.3% listed is 0.953 * 0.9999 = 0.953.

We can similarly calculate the probability that /mamumuntol/ → [mampupuntol]

will be the optimal candidate:

P(/mamumuntol/ is available) = 0.953

P((PU or *[m >> ENTRYLIN) and (PU or *[m >> *NC8) and (PU or *[m >>
NASSUB) and (USELISTED >> ENTRYLIN)) = 0.00003
P(/mamumuntol/ → [mampupuntol] ) = 0.00003

Thus, /mamumuntol/ → [mampupuntol] is possible, but extremely unlikely.

We can also calculate P(/maN/+/RCV/+/puntol/ → [mampupuntol] ) and
P(/maN/+/RCV/+/puntol/ → [mamumuntol] ), both small but not minuscule:

63
Using the method in §2.11, this is the result of integrating pdf(z*NC8, zNasSub, zUseListed, zEntryLin, z*[m) over the
region where the requirements in (ii) and (iii) are met.

81
 P(/maN/+/RCV/+/puntol/ → [mampupuntol] )
= P(/maN/+/RCV/+/puntol/ → [mampupuntol] | /mamumuntol/ is not available) *
P(/mamumuntol/ is not available)
+ P(/maN/+/RCV/+/puntol/ → [mampupuntol] | /mamumuntol/ is available) *
P(/mamumuntol/ is available)
= 0.600 * 0.047 + 0.00003 * 0.953
= 0.029

P(/maN/+/RCV/+/puntol/ → [mamumuntol] ) =
= P(/maN/+/RCV/+/puntol/ → [mamumuntol] | /mamumuntol/ is not available) *
P(/mamumuntol/ is not available)
+ P(/maN/+/RCV/+/puntol/ → [mamumuntol] | /mamumuntol/ is available) *
P(/mamumuntol/ is available)
= 0.399 * 0.047 + 0 * 0.953
= 0.019
(P(/maN/+/RCV/+/puntol/ → [mamumuntol] | /mamumuntol/ is available) = 0
because candidate /maN/+/RCV/+/puntol/ → [mamumuntol] 's constraint
violations are a superset of candidate /mamumuntol/ → [mamumuntol] 's.)

We can perform the same calculations to determine the likelihood of each

outcome if the 95.3% listed input /mampupuntol/ exists (assuming there is no listed input
/mamumuntol/64):

64
If there are two listed entries for the word, the calculations are still straightforward, but there are six
candidates in the tableau (two for the first entry, two for the second entry, and two for the prefix+stem
combination). But the model given in 3 of how the listener updates her lexicon prevents two competing
entries from becoming fully listed, so this case is not considered here.

82
(55) Candidate probabilities if /mampupuntol/ exists
USE ENTRY MORPH *NC8 NASSUB PU-maN *[N *[n *[m
LISTED LIN ORDER +RCV-
/mampupuntol/
* *
→ [mampupuntol]
/mampupuntol/
* * *
→ [mamumuntol]
/maN/+/RCV/+/puntol/
* * *
→ [mampupuntol]
/maN/+/RCV/+/puntol/
* * * *
→ [mamumuntol]

P(/mampupuntol/ → [mampupuntol]) = 0.953

P(/mampupuntol/ → [mamumuntol]) = 0.00003
P(/maN/+/RCV/+/puntol/ → [mampupuntol]) = 0.029
P(/maN/+/RCV/+/puntol/ → [mamumuntol]) = 0.019

The following table summarizes the same results for all six types of initial obstruent, to
five decimal places:

(56) P(input|output) for various stem-initial obstruents

p t/s k b d g
/substituted/
P(/substituted/ .95251 .95249 .95219 .95250 .95247 .95213
→ [substituted])
P(/substituted/ .00003 .00004 .00019 .00004 .00005 .00022
→ [unsubstituted])
P(/maN/+/RCV/+/X/ .02852 .02868 .02969 .04429 .04441 .04500
→ [unsubstituted])
P(/maN/+/RCV/+/X/ .01894 .01879 .01793 .00317 .00307 .00265
→ [substituted])
/unsubstituted/
P(/unsubstituted/ .94566 .94566 .94568 .95246 .95246 .95246
→ [unsubstituted])
P(/unsubstituted/ .00363 .00363 .00361 .00007 .00007 .00006
→ [substituted])
P(/maN/+/RCV/+/X/ .02849 .02864 .02950 .04426 .04436 .04478
→ [unsubstituted])
P(/maN/+/RCV/+/X/ .02223 .02208 .02121 .00322 .00312 .00270
→ [substituted])

83
 The high ranking values of USELISTED and ENTRYLINEARITY tend to swamp
differences among stem-initial segments and between the two constructions, but as we
will now see, the differences become greater when there is no listed form.

2.7.3. Generating a novel form

When there is no listed form, the only possible candidates are /maN/+/RCV/+/X/ →
[unsubstituted] and /maN/+/RCV/+/X/ → [substituted] . The probabilities of the two

outcomes for each stem-initial obstruent are given in (57), which shows that the overall
rate of substitution on novel words will be fairly low. There are slight differences in
probability of substitution among the three places of articulation, and there is a sharp
difference between voiced and voiceless segments.

(57) Probabilities of outcomes when no listed form exists

p t/s k b d g
P(/maN/+/RCV/+/X/ .60066 .60385 .62198 .93314 .93527 .94413
→ [unsubstituted]
P(/maN/+/RCV/+/X/ .39934 .39615 .37802 .06686 .06473 .05587
→ [substituted]

We can see, then, that the grammar produces the desired result for speakers: very
high faithfulness to listed words, and low but nonzero substitution on novel words.
Chapter 3 shows how the probabilistic interaction of speakers and listeners shapes the
establishment of new words in the lexicon.

84
 2.8. The Listener

2.8.1. Introduction

In addition to the behavior of the learner and the speaker, the model must also account for
the behavior of the listener. Most work on perception/comprehension in OT has focussed
on how the listener retrieves the underlying form given the utterance she hears
(Smolensky 1996b, Tesar 1998, Boersma 1998, Pater 1999a). The meat of that problem
here is not calculating the segmental content of the input, but rather deciding whether the
input was a single listed word or a concatenation of morphemes. This section discusses
how the listener makes this decision, which is crucial to determining the probability that a
new polymorphemic word will eventually be assimilated into the lexicon as substituted or
as unsubstituted. This section also discusses how the listener arrives at a judgment of how
acceptable an utterance is; in particular, I will show how the model produces
acceptability judgments similar to those seen in the experiment.

2.8.2. Reconstructing the underlying form

The idea of lexicon optimization was introduced Prince and Smolensky (1993) and
elaborated by Itô, Mester, and Padgett (1995) and Smolensky (1996b): given an output
produced by another speaker, the listener chooses the input such that the input-output pair
is maximally harmonic. A schematic example is shown in (58).

(58) Choosing the optimal input

[bak] NOCODA DEP-C
. /bak/ → [bak] *
/ba/ → [bak] * *!

85
Because the output is held constant, violations of pure markedness constraints (in this
case NOCODA) and of correspondence constraints not involving the input (e.g., CORR-
BR) are the same for every input. Therefore, CORR-IO constraints alone (here, DEP-C)
determines the optimal input, and the optimal input is the one that is most similar to the
actual output. Differences between input and output then exist only when driven by
alternations.65 For example, in Hale and Reiss’s (1998) model of grammar- and lexicon-
learning, when different outputs are recognized as containing the (semantically and
morphosyntactically) same morpheme, in order to avoid synonymy they are learned as
having the same input, which must then violate Input-Output Correspondence at least
sometimes.
Without adopting the details of any particular version of input recognition in
Optimality Theory, I will assume that the adult listener is capable of recognizing that
hypothetical [mamumuntol] —uttered in a context that supplies morphosyntactic and
semantic information—may be composed of the familiar morphemes maN, RCV, and
puntol.66

65
Or, as in Prince and Smolensky 1993 (p. 196), by violations of *SPEC, which prohibits underlying
material. The tension between *SPEC and Input-Output Correspondence is the tension between storing as
little information as possible in the lexicon and changing the input as little as possible when uttering it.
66
An interesting question is what the listener does if the stem puntol is not familiar. The listener must then
decide whether the stem is puntol, buntol, or muntol (tuntol, etc. are easily ruled out by faithfulness
constraints on obstruent place of articulation).

The model predicts that the probability that the listener would select a particular stem—
P(/puntol/|[mamumuntol])—is proportional to two other probabilities: first, the prior probability of that
stem’s existence—P(/puntol/)—which can be calculated from lexical statistics on the frequency of word-
initial p, the frequency of cooccurrence of p and l within a word, etc.; and second, the probability that
[mamumuntol] would be produced given the stem under consideration—P([mamumuntol]|/puntol/)—
which is straightforwardly calculable from the constraint ranking.

In the experiments described above in §2.3.3, though, the listener knows the segmental content of
the stem, because it is presented in the prompt, so stem selection is not part of the task.

86
 But the listener also must consider the possibility that [mamumuntol] was
generated from a single listed form, such as /mamumuntol/ or /mampupuntol/. Assuming
that decisions about underlying forms are made stochastically, the listener must compare
the three probabilities in (59).

(59) Three possibilities on hearing [mamumuntol]

P(/maN/+/RCV/+/puntol/|[mamumuntol]) “the probability that the speaker’s input

was /maN/+/RCV/+/puntol/, given that
the output heard was [mamumuntol] ”
P(/mamumuntol/|[mamumuntol]) “the probability that the speaker’s input
was /mamumuntol/, given that the output
heard was [mamumuntol] ”
and

P(/mampupuntol/|[mamumuntol]) “the probability that the speaker’s input

was /mampupuntol/, given that the
output heard was [mamumuntol] ”

As shown in (60), we can rewrite these using Bayes’ Theorem. The theorem states:
P(A|B) = P(B|A)*P(A)/P(B)
“The probability of A given B is equal to the probability of B given A, times the
prior probability of A (i.e. the probability of A when nothing is known about B),
divided by the prior probability of B.”

(60) Bayesian inversion of probabilities compared by listener

P(/maN/+/RCV/+/puntol/ | [mamumuntol])
= P([mamumuntol] | /maN+RCV+puntol/) * P(/maN/+/RCV/+/puntol/) /
P([mamumuntol])
P(/mamumuntol/ | [mamumuntol])
= P([mamumuntol] | /mamumuntol/) * P(/mamumuntol/) / P([mamumuntol])
P(/mampupuntol/ | [mamumuntol])
= *P([mamumuntol] | /mampupuntol/) * P(/mampupuntol/) /
P([mamumuntol])

87
 Since the denominators are the same in all three expressions, the numerators
determine the differences in probability. The probabilities P([mamumuntol]|
/maN/+/RCV/+/puntol/), P([mamumuntol]| /mamumuntol/), and P([mamumuntol]|

/mampupuntol/) are calculated by the grammar. Given the grammar learned in §2.6, they

are equal to 0.39934, 0.99936, and 0.00003, respectively. But we still need to know the
prior probabilities P(/maN/+/RCV/+/puntol/), P(/mamumuntol/), and P(/mampupuntol/). In
other words, the listener must decide how likely it is that the speaker’s lexicon contains
this word as a single, pre-packaged entity (and that this lexical entry was used) versus
how likely it was that the speaker formed the word on the fly by concatenating a prefix
and a stem.
How does the listener make this decision? One possibility is that she relies solely
on the listedness of /mamumuntol/ or /mampupuntol/ in her own lexicon, taking each
word’s listedness as the probability that it was used by the speaker.
But a more cautious listener, capable of learning new words from interlocutors,
would also take into account the overall productivity of the maN+RCV- construction.
P(/maN/+/RCV/+/puntol/) should decrease as the listedness of a whole word

(/mamumuntol/ or /mampupuntol/) increases, and should increase as the productivity of

maN+RCV- increases. In other words, the more listed a whole word is for the listener, the

less likely that the speaker would have composed /maN/+/RCV/+/puntol/ on the fly—since
the speaker and listener belong to the same speech community, the listener can assume
that their lexicons will tend to be similar—and, the more productive the construction is,
the more easily the speaker could have employed it to generate a new word. Additionally,
P(/maN/+/RCV/+/puntol/) should be close to 0 if a whole word is 100% listed, regardless

of the productivity of maN+RCV- (no matter how productive the construction is, if the
word is already listed it will probably not be formed anew), and it should also be close to

88
0 if the productivity of maN+RCV- is zero, regardless of the listedness of any whole word
(even if the word is isn’t listed for the listener, if the construction is not productive, it
must have been listed for the speaker). The function shown in (61) has the desired
properties; the constants 3 and 6 were chosen (somewhat arbitrarily) because they
produce endpoints that are close to zero and one, and a gentle slope (rather than a strict
cutoff) centered on 0.5 on each axis.67

(61) P(/maN/+/RCV/+/puntol/) = 1/((1+e-3+6*Listedness(whole word))*(1+e3-6*Productivity(maN+Rcv)))

67
In a function of the form y = 1/(1+ea-bx) (a logistic function), b determines how steep the function is
(large absolute value for b means steep slope; positive b means y increases as x increases; negative b means
y decreases as x increases)and b/a is the location of the “half-way point”—the value of x for which y = 0.5.
Similarly, in multi-dimensional functions with multiple (1+e^(ai - bixi)) multiplied together in the
denominator, each bi determines the steepness of the function along the dimension xi, and ai/bi determines
where on the xi axis the function is centered.

89
where Listedness(whole word) is the listedness of whichever appropriate word is more
listed (max(Listedness(/mamumuntol/), Listedness(/mampupuntol))).
How does the listener assess the productivity of the construction maN-Rcv? There
are several cues available. One cue is the proportion of stems of the appropriate
morphosyntactic and semantic category that the listener has experienced as occurring in
the construction. For example, if maN-Rcv- is highly productive, the listener will have
heard the maN-Rcv- form of many stems; gaps would be accidental (and should tend to be
for rare stems). But if it is not very productive, only (or mostly68) those stems that have a
listed maN-Rcv- form can ever occur with maN-Rcv-, and so there will be many stems that
the listener has never heard with maN-Rcv-. If we can use dictionary entries as a rough
guide,69 sampling just the first stem on every tenth page70 with any nonstative verbal
derivative (as a rough diagnostic of suitability for the maN-Rcv- construction), 12 out of
152 have a maN-Rcv- derivative, yielding a productivity index of 0.079. Ideally, this index
would be weighted for frequency—the absence of a maN-Rcv- form for a low-frequency
stem should not count against productivity as much it would for a high-frequency stem.
A second cue is the correlation between the token frequency of each maN-Rcv-
word and the token frequency of its stem. If the construction is very unproductive, there
will be many separately listed maN-Rcv- words, whose frequencies are not affected by the

68
Speakers might occasionally use an unproductive construction to create a nonce form.
69
There is an obvious flaw in relying on the dictionary, of course, rather than a text or speech corpus,
because, depending on the lexicographer’s methods, a very productive construction may be less likely to
have its products listed in the dictionary (for example, in English 1986, only the infinitive of each verb is
listed, not the various aspects). In addition, for any construction, there are probably some missing derived
forms, causing all productivity indices to be artificially low.
70
Excluding nasal-initial stems.

90
frequencies of their stem, weakening the correlation. Since frequency data are not
available, though, we cannot calculate a productivity index based on this cue.
A third cue is the proportion of maN-Rcv- words that are phonologically or
semantically idiosyncratic. These words must have their own lexical entries to contain the
idiosyncratic information. Phonological idiosyncrasy in this case could include nasal
substitution and stress shifts. The behavior of maN-Rcv- words with respect to stress and
nasal substitution is summarized in (62). The cells in boldface are those that could be
considered idiosyncratic (either a stress change or nasal substitution), and they make up
119/195 = 61% of the total. Put another way, 39% of the maN-Rcv- words listed in the
dictionary lack idiosyncratic phonological characteristics, and thus a maximum of 39%
could lack their own lexical entries and be formed on the fly.

(62) Idiosyncrasies in maN-Rcv- words

stress change
none varies penultimate final → total
→ final penultimate
does not substitute 50 1 1 1 53
nasal
substitution

varies 3 0 0 0 3
substitutes 80 14 6 5 105
sonorant-initial 26 2 6 0 34
(cannot substitute)
total 159 17 13 6 195

If we took into account semantic idiosyncrasy, the figure might fall further. I will
not develop a formal metric of semantic idiosyncrasy here, but it is clear from casual
inspection of the various nasal-substituting constructions that some produce more
semantic idiosyncrasy than others do. For example, the meaning of a paN- (instrumental

91
adjective)71 word is almost completely predictable: paN-X means “used as a tool for X”. In
contrast, the meaning of a maN+Rcv- word can be considerably less predictable.
Manlulustaj ‘embezzler’ from lustaj ‘embezzle’ is straightforward enough, but

manliliNkis ‘boa constrictor’ from liNkis ‘tightly bound’ surely must have its own lexical

entry.
The productivity index for maN+Rcv- is, then, roughly somewhere between 0.08
and 0.39. For the sake of argument, let us assume it is 0.2, which the listener combines
with her listedness for this particular word, using the function in (61), to arrive at the
prior probability P(/maN/+/RCV/+/puntol/). If no whole word is listed at all for the
listener, P(/maN/+/RCV/+/puntol/) = 1/((1+e-3+6*0)(1+e3-6*0.2)) = 0.135.
Because the only alternatives to synthesized P(/maN/+/RCV/+/puntol/) that are
remotely probable are listed /mamumuntol/ and listed /mampupuntol/, the prior
probabilities P(/mamumuntol/) and P(/mampupuntol/) must add up to about 1 - 0.135 =
0.865 (still in the case that the listener has nothing listed). We want a function such that
P(/mamumuntol/)’s share of the 0.865 (i) is greater the more listed /mamumuntol/ is for

the listener, (ii) is smaller the more listed competing /mampupuntol/ is for the listener,
and (iii) is greater the larger the proportion of existing potentially-substituting words with
p-initial stems that undergo nasal substitution. Condition (iii) is necessary because in the

71
This raises the question of whether it makes sense to treat the various adjectival paN-s as separate
constructions (likewise nominal paN-, verbal maN-). It may be that adjectival paN- is really just one
construction, part of whose semantic function depends on the nature of the stem, so that the primary
meaning for a stem that denotes an action is instrumental, the primary meaning for a stem that denotes a
situation or class of people is reservative, and any other meaning can be considered idiosyncratic.

92
event that neither /mamumuntol/ nor /mampupuntol/ is listed at all for the listener, she
must rely on substitution rates in her lexicon72 to decide which would be more likely.
Consider the following function (again, constants are somewhat arbitrary—see fn.
67):
F(/mamumuntol/) =
1/((1+e2-6*Listedness(/mamumuntol/))(1+e-4+6*Listedness(/mampupuntol/))(1+e3-6*SubstProp(p)))

F increases with Listedness(/mamumuntol/), decreases with Listedness(/mampupuntol),

and increases with SubstProp(p), the proportion of potentially nasal-substituting words

based on p-initial stems that substitute (SubstProp(p) is 1, but the proportion for other
segments is lower). Similarly,

F(/mampupuntol/) =
1/((1+e2-6*Listedness(/mampupuntol/))(1+e-4+6*Listedness(/mamumuntol/))(1+e3-6*UnsubstProp(p)))

The units of F are arbitrary, since the purpose of F is to compute /mamumuntol/’s and
/mampupuntol/’s respective shares of 1- P(/maN/+/Rcv/+/puntol/). We can now use F to

calculate P(/mamumuntol/) and P(/mampupuntol/) by dividing up 1-

P(/maN/+/Rcv/+/puntol/) proportionally:

(63) Prior probabilities of /mamumuntol/ and /mampupuntol/

P(/mamumuntol/)
= (1-P(/maN/+/Rcv/+/puntol/)) * F(/mamumuntol/) _
F(/mamumuntol/)+F(/mampupuntol/)
P(/mampupuntol/)
= (1-P(/maN/+/Rcv/+/puntol/)) * F(/mampupuntol/) _
F(/mamumuntol/)+F(/mampupuntol/)

72
In a richer model, the listener could rely not just on substitution rates for p-initial stems, but also on
substitution rates for classes of stems related in other ways (other segments in the stem, number of
syllables).

93
For example, if Listedness(/mamumuntol/) = Listedness(/mampupuntol/) = 0,

F(/mamumuntol/)
= 1/((1+e2-6*Listdnss(/mamumuntol/))(1+e-4+6*Listdnss(/mampupuntol/))(1+e3-6*SubstProp(p)))
= 1/((1+e2-6*0)(1+e-4+6*0))(1+e3-6*1))
= 0.112
F(/mampupuntol/)
= 1/((1+e2-6*Listdnss(/mampupuntol/))(1+e-4+6*Listdnss(/mamumuntol/))(1+e3-6*UnsubstProp(p)))
= 1/((1+e2-6*0)(1+e-4+6*0)(1+e3-6*0))
= 0.006
so P(/mamumuntol/) = 0.865 * 0.112 / (0.112 + 0.006) = 0.824
and P(/mampupuntol/) = 0.865 * 0.006 / (0.112 + 0.006) = 0.041

It is now possible to begin calculating the probabilities in (60), which was the use
of Bayes’ Law by the listener to calculate the probability that the speaker was using a
particular input. In (64), the numerators are calculated using the figures arrived at above.

(64) Calculating (60) when listener has no listed form

P(/maN/+/puntol/|[mamumuntol]) = 0.399 * 0.135 / P([mamumuntol])
P(/mamumuntol /|[mamumuntol]) = 0.999 * 0.824 / P([mamumuntol])
P(/mampupuntol/|[mamumuntol]) = 0.00003 * 0.041 / P([mamumuntol])

The denominator can now be calculated also, by adding together the probability of
deriving [mamumuntol] from each possible source:

(65) Prior probability of the output

P([mamumuntol])
= P([mamumuntol]|/maN/+/Rcv/+/puntol/) * P(/maN/+/Rcv/+/puntol/) +
P([mamumuntol]|/mamumuntol/) * P(/mamumuntol/) +
P([mamumuntol]|/mampupuntol/) * P(/mampupuntol)
≈ 0.399 * 0.135 + 0.999 * 0.824 + 0.00003 * 0.041 = 0.878

Plugging this denominator into the equations in (65), we get:

94
(66) Final result for (64)
P(/maN/+/Rcv/+/puntol/|[mamumuntol]) ≈ 0.399 * 0.135 / 0.878 = 0.062
P(/mamumuntol/|[mamumuntol]) ≈ 0.999 * 0.824 / 0.878 = 0.939
P(/mampupuntol/|[mamumuntol]) ≈ 0.00003 * 0.041 / 0.878 = 0.000002

So, given an output [mamumuntol] , a listener with neither /mamumuntol/ nor

/mampupuntol/ will still be most likely to identify /mamumuntol/ as the input, because the

construction is not very productive, and because P([mamumuntol]|/mamumuntol/) is

much larger than P([mamumuntol]| /maN/+/Rcv/+/puntol/). Still, it is not outlandish to
guess that [mamumuntol] was synthesized (i.e., came from /maN/+/Rcv/+/puntol/)—the
listener will choose that possibility 6% of the time. She will almost never (1 time out of
every 500,000) guess that the input was /mampupuntol/.73
We can perform the same calculations for cases in which the listener hears
[mampupuntol] :

(67) Determining the input given output [mampupuntol]

P(/maN/+/Rcv/+/puntol/|[mampupuntol]) =
P([mampupuntol]|/maN/+/Rcv/+/puntol/) * P(/maN/+/Rcv/+/puntol/)
/ P([mampupuntol])
= 0.601 * 0.135 / 0.125 = 0.649
P(/mamumuntol/|[mampupuntol])
= P(/mamumuntol/) * P([mampupuntol]|/mamumuntol/) / P([mampupuntol])
= 0.004 * 0.824 / 0.125 = 0.025
P(/mampupuntol/|[mampupuntol])
= P(/mampupuntol/) * P([mampupuntol]|/mampupuntol /) / P([mampupuntol])
= 0.993 * 0.041 / 0.125 = 0.326

73
The reason P(/mampupuntol/|[mamumuntol]) is not quite zero is that (i) the stochastic grammar has a
slight chance of producing [mamumuntol] from /mampupuntol/, if NASSUB or *NC8 should outrank
ENTRYLINEARITY, and (ii) the prior probability of /mampupuntol/ is slightly greater than zero: although no
existing p-stem words fail to nasal-substitute in the maN+REDCV- construction, F makes room for the
possibility that a new one could come along.

95
 The difference between (66) and (67) is striking: even if the listener has no
relevant listed word, she is quite likely (94% probability) to conclude, after hearing
mamumuntol, that the speaker was using a listed, substituted word and update her lexicon

accordingly. After hearing mampupuntol, however, she is somewhat more likely to

conclude that the speaker formed the word on the fly than from a listed, unsubstituted
word (64% vs. 33% probability). This difference occurs partly because the difference
between P([mamumuntol]|/mamumuntol/) and P([mamumuntol] |/maN/+/Rcv/+/puntol/)
(0.999 vs. 0.399) is greater than the difference between
P([mampupuntol]|/mampupuntol/) and P([mampupuntol] |/maN/+/Rcv/+/puntol/) (0.993

vs. 0.601), and partly because the prior probability P(/mamumuntol/) is large and the
prior probability P(/mampupuntol/) is small (0.824 vs. 0.041). That is, (i) a nasal-
substituted pronunciation is 60 percentage points more likely to occur with a listed input
than if synthesized, whereas an unsubstituted pronunciation is only 40 percentage points
more likely to occur with a listed input than if synthesized; and (ii) for stems beginning
with p, the likelihood of a substituted listed form’s existing is greater than the likelihood
of an unsubstituted form’s existing.
The graph in (68) shows the difference between P(/substituted/|[substituted]) and
P(/unsubstituted/|[substituted]) for each stem-initial obstruent and for 4 different

listedness situations. Values greater than 0 indicate that for that obstruent and listedness
situation, a listener is more likely to update her lexicon when she hears a substituted word
than when she hears an unsubstituted word. For example, if the listener has neither a
substituted nor an unsubstituted word in her lexicon ( ), her likelihood of
recording a substituted p-stem word is about 60 percentage points higher than her
likelihood of recording an unsubstituted p-stem word.

96
 Nearly all the values are greater than zero; unless the listener has a listed
unsubstituted word and no listed substituted word in her lexicon ( ), or unless the
stem-initial segment is one that rarely undergoes nasal substitution (d or g74), the listener
is always more likely when she hears a substituted word than when she hears an
unsubstituted word to assume the speaker was using a listed word (and update her own
lexicon accordingly). This fact will be crucial in Chapter 3: despite the low rate of
substitution on novel words, a new word still has a good chance of eventually being
adopted by the speech community as substituted, since listeners will ignore most
unsubstituted instances of the word, assuming them to have been formed on the fly.75

74
For these obstruents, P([substituted] | /synthesized/) is low (and so P(/synthesized/ | [substituted]) is low),
but P(/substituted/) is low (and so P(/substituted/ | [substituted]) is also low).
75
I assume that only listeners update their lexicons. It is also possible that speakers update their own
lexicons in response to utterances they themselves have produced.

97
(68) Probability of listener’s guessing that speaker used a listed word: substituted -

unsubstituted

P(/sub/|[sub]) - P(/unsub/|[unsub])

1.0
0.8
0.6
0.4
0.2
0.0
-0.2 p t/s k b d g
-0.4
-0.6
-0.8
-1.0
stem-initial obstruent

L(/sub/) = L(/unsub/) = 0 L(/sub/) = L(/unsub/) = 0.5

L(/sub/) = 0, L(/unsub/) = 1 L(/sub/) = 1, L(/unsub/) = 0

2.8.3. Acceptability judgments

The other aspect of the listener’s behavior to be discussed here is the generation of
acceptability judgments. Following Hayes and MacEachern (1998) and Boersma and
Hayes (1999), I will assume that the listener’s acceptability judgment is a function76 of
the probability that her grammar could generate the utterance she has heard. This
probability is directly calculable from the ranking values of the constraints (as discussed
in §2.11), although it can also be approximated by running many trials of the constraint

76
Using the function for acceptability ratings from Boersma & Hayes (1999), Acceptability([substituted]) -
Acceptability([unsubstituted]) = log(1/P([substituted]) - 1) / log(0.2). The constant 0.2 was arrived at by
trial and error for a 7-point rating scale (rather than my 10-point scale), but it seems to work well here also.

98
system and seeing how often the form in question was generated. Calculating the
underlying form (a single word or a concatenation of morphemes) is also essential,
because the underlying form must be known in order to determine how well the utterance
satisfies CORR-IO constraints, violations of which reduce the utterance’s probability of
being generated.
Although the listener’s own probability of producing novel [mamumuntol]
(Listedness(/mamumuntol/) = Listedness(/mampupuntol/) = 0 for the listener), would be
low—P([mamumuntol]| /maN/+/Rcv/+/puntol/) = .399—when she hears someone else
say [mamumuntol] , she cannot be sure what her interlocutor’s input form was, and so her
estimate of P([mamumuntol]) for purposes of calculating an acceptability rating must
reflect all the possibilities, as shown in (69):

(69) P([mamumuntol])
= P([mamumuntol]|/mamumuntol/) * P(/mamumuntol/) +
P([mamumuntol]|/mampupuntol/) * P(/mampupuntol/) +
P([mamumuntol]| /maN/+/Rcv/+/puntol/) * P(/maN/+/Rcv/+/puntol/)
= 0.999 * 0.824 + 0.00003 * 0.041 + 0.399 * 0.135 = 0.878
(same numbers as in (64))

similarly,
P([mampupuntol])
= P(/mampupuntol/) * P([mampupuntol]|/mampupuntol/) +
P(/mamumuntol/) * P([mampupuntol]|/mamumuntol/) +
P([mampupuntol]| /maN/+/Rcv/+/puntol/) * P(/maN/+/Rcv/+/puntol/)
= 0.993 * 0.041 + 0.004 * 0.824 + 0.601 * 0.135 = 0.125

The result is that the probability of producing [mamumuntol] when the input can
only be guessed at is much higher than the probability of producing [mamumuntol] when
the input must be /maN/+/Rcv/+/puntol/ (0.399 vs. 0.878). The probability for

99
[mampupuntol] actually decreases (0.601 vs. 0.125), because the prior probability that

the speaker would one of the inputs that would be likely to produce [mampupuntol] as
output (/mampupuntol/ or /maN/+/Rcv/+/puntol/) is small. This may explain the
experimental results in §2.3.3.1: listeners judged novel substituted words to be fairly
acceptable (for voiceless-initial stems, they were judged more acceptable on average than
the unsubstituted forms of the same words), even though they produced them rarely. The
acceptability judgments were high because judges had to allow for the possibility that the
interlocutor (in this case, the hypothetical speaker whose utterances were written on the
cards shown to the judges) was using a word familiar to herself although unknown to the
judge.
How well does the model reproduce the experimental results? The graph in (70)
shows the model’s predictions.

(70) Predicted acceptability of substituted vs. unsubstituted for novel words

Acceptability(substituted) - Acceptability(unsubstituted)

1.5
1.0
0.5
0.0
-0.5 p t/s k b d g
-1.0
-1.5
-2.0

The model correctly predicts the distinction between voiced and voiceless, and
predicts a weak place-of-articulation effect. The graph in (71) shows the model’s
predicted values from (70) against the experimental values from (25).

100
(71) Predicted and experimental acceptability values (substituted - unsubstituted)

Acceptability(substituted) - Acceptability(unsubstituted)

2.5
2.0
1.5
1.0
0.5 output of model
0.0
experimental
-0.5
-1.0
-1.5
-2.0
-2.5
p t/s k b d g

How good a match is this? The experimental results and the output of the model
both reflect a voicing difference: for the voiceless stops, substitution is more acceptable
(even though it is the minority pronunciation). For the voiced stops, nonsubstitution is
more acceptable (for the model, substitution and nonsubstitution are equally acceptable
for /b/). Neither the experimental results nor the output of the model reflects the place
effect strongly: for the model, there is little difference among the voiceless obstruents and
a strong difference between /b/ versus /d/ and /g/, but no difference between /d/ and /g/.

2.9. Chapter Summary

This chapter has presented a model of lexical regularities using the example of nasal
substitution. It was argued that although the characteristics of existing words (substituting
or not) are determined by the lexicon, nasal substitution and its regularities are
nonetheless represented in the linguistic system. The model presented here attempted to
encode knowledge of nasal substitution directly into the grammar, by means of low-

101
ranked constraints that are relevant only for novel words (for existing words, high-ranked
USELISTED and CORR-IO require that the lexical entry be faithfully used).
The probabilistic rankings of the subterranean constraints are learnable through
exposure to the existing lexicon and result in variable speaker behavior for novel words
that reflects the patterns in the lexicon. The same probabilistic grammar used in speaking
can be used in listening, to make a probabilistic guess as to a speaker’s underlying form
Bayesian reasoning on the part of the listener results in a bias in favor of guessing that a
nasal-substituted utterance was generated from a single lexical entry (rather than from
morpheme concatenation). The grammar can also be used to generate acceptability
judgments for novel words (which are similar to the acceptability judgments seen
experimentally). Here, the listener’s uncertainty as to whether a novel-to-her word was
also novel for the speaker results in higher acceptability ratings for nasal-substituted
words than might be expected from the low rate of substitution on novel words when the
grammar is used for speaking.

102
 2.10. Appendix: experimental stimuli

For each obstruent (including ?), three novel-word stimulus stems were created. Each
stimulus was two syllables long, did not violate any morpheme structure constraints of
which I am aware,77 and would not be homophonous with an existing stem if substituted
(for example, since dapat already exists, sapat would not be considered as a novel stem).
There were no pseudoreduplicated novel stems.
For a given obstruent, each of the three stems had a different first-syllable vowel
(i, a, or u), and a prosodic pattern: penultimate stress/length; final stress and closed
penult; or final stress with open (short) penult. There were, however, four d-initial stems,
two flapped and two unflapped. As it turned out, flapping made no difference in
participant behavior. (72) gives the complete list of novel stems and the approximate
meanings conveyed by each stem’s accompanying illustration.

(72) Novel stimulus stems

pali@m ‘push a wheelbarrow’
pi@ùhig ‘get fruit down from tree by hitting with a stick’
punto@l ‘prune a tree’
taha@l ‘tie saplings together for support’
tikla@s ‘throw feed to chickens’
tu@ùkas ‘drive pigs into corral’
sawi@k ‘split cane’
siglo@t ‘carry water’
su@ùkad ‘weave a basket’
ka@ùpat ‘build a fence’
kiri@t ‘hoe earth’
kunta@l ‘call cattle’
ba@ùkad ‘stamp down earth over newly planted seeds’
bili@d ‘decorate ceramic jugs’
bugna@t ‘chisel strips of plank of wood’
dagsi@l ‘remove caught fish from hooks’ (flapped: pagdaragsi@l)
dampo@s ‘remove flowers from plant’ (not flapped: pagdadampo@s)

77
including a dispreference for identical consonants within the same root, unless it is pseudoreduplicated
(see §4.4 for examples of pseudoreduplicated roots).

103
 di@ùkib ‘sew fishing nets’ (not flapped: pagdidi@kib)
dugo@l ‘dig up plants’ (flapped: pagdurugo@l)
gana@t ‘train vines on supports’
gi@ùtap ‘smoothen edges of pot’
gutla@w ‘fish using a trap’
?ambo@j ‘fish using a net’
?i@ùlab ‘collect eggs from nests’
?u@ùbon ‘cool hot metal in water’
stimuli used as practice for Task II
ga@ùmat ‘pound grain’
pito@s ‘rake’

The criteria for choosing the real-word stimuli listed in (73) (used as practice
stimuli for both groups, and interspersed with novel stimuli for Group A) were just that
they have both an existing pag+RCV- form and a maN+RCV- form. An effort was made to
include some common and some rare real words. Some of the real-word stems are
sonorant-initial, and thus cannot undergo nasal substitution; in Task II (acceptability
judgments), only the unsubstituted forms of sonorant-initial stems were used.

(73) Real-word stimulus stems

Example stimuli for Task I (maN+RCV- form given)
bitha@j ‘sift’78
hi@ùlot ‘massage’
Practice stimuli for Task I (maN+RCV- form filled in by participant)
la@ùpi? ‘butcher’
bu@ùbo? ‘smelt’
Stimuli interspersed with novel stems in Group A
ha@ùbi ‘weave’
buno@? ‘wrestle’
saja@w ‘dance’
su@ùri? ‘analyze’
sulsi@ ‘mend’
li@ùlok ‘sculpt’
taNgo@l ‘defend (in court)’
hasi@k ‘sow seeds’

78
These are not the actual glosses of the stems when used bare (bitha@j means ‘sieve’), but rather the glosses
for the action to which both the pag+RCV- form and the maN+RCV- form refer.

104
 2.11. Appendix: Calculating probabilities of rankings

2.11.1. Pairwise ranking requirements

To calculate a pairwise ranking probability, P(Ci>>Cj), we can use the fact that given
two normally distributed populations I and J with means µi and µj and standard
deviations σi and σj, if we take samples of size ni from I and samples of size nj from J, the
difference between the means of the two samples, Mi - Mj, will be normally distributed,
with mean µi - µj and standard deviation sqrt((σi2 / ni) + (σj2 / nj)).79 This is illustrated in
(74).

(74) Mi - Mj
Population I has mean µi = 40 and standard deviation σi = 5.
Population J has mean µj = 35 and standard deviation σj = 2.
Sample ni = 10 points from I and nj = 20 points from J:

I
J

20 25 30 35 40 45 50 55 60

Find the mean of each sample: Mi = 42.2 and Mj = 35.5 (Mi - Mj = 6.7)
If we take enough samples, the mean value of Mi - Mj approaches µi - µj = 5, with
standard deviation sqrt((σi2 / ni) + (σj2 / nj)) ≈ 1.64

79
When ni = nj = 1, as in our case (see below), this means that the variance (σ2, the square of the standard
deviation) of Mi - Mj is equal to the sum of the variances of I and J.

105
 To see how this applies to our case, first a bit more detail on Boersma’s system is
necessary. An actual value, or selection point, for a constraint (“disharmony”, in
Boersma’s terms) is generated by adding to the ranking value a random variable with the
standard normal distribution, multiplied by a value called “ranking spreading” (following
Boersma and Hayes 1999, I use a rankingSpreading of 2):

(75) Arriving at a selection point for a constraint in a given utterance

selectionPoint = rankingValue + rankingSpreading * z
where z is a random variable, normally distributed with mean 0 and standard
deviation 1.

This means that the quantity (selectionPoint-rankingValue)/rankingSpreading (=

z) is normally distributed, with mean 0 and standard deviation 1. We can then employ the

method above to any two constraints Ci and Cj, taking samples of size ni = nj = 1 from the
distributions (selectionPointi-rankingValuei)/rankingSpreading and (selectionPointj-
rankingValuej)/rankingSpreading, which both have mean µi = µj = 0 and standard

deviation σi = σj = 1.
Since the sample sizes are 1, Mi and Mj are just the values of (selectionPointi-
rankingValuei)/rankingSpreading and (selectionPointj-rankingValuej)/rankingSpreading

on a given occasion. Then we have:

(76) Calculating P(Ci>>Cj)

Mi - Mj
= (selectionPointi-rankingValuei)/rankingSpreading - (selectionPointj-
rankingValuej)/rankingSpreading
so

rankingSpreading * (Mi - Mj) + rankingValuei - rankingValuej

= selectionPointi- selectionPointj

106
 P(Ci>>Cj)
= P(selectionPointi> selectionPointj)
= P(selectionPointi- selectionPointj> 0)
= P(rankingSpreading * (Mi - Mj) + rankingValuei - rankingValuej > 0)
= P(Mi - Mj > (rankingValuej - rankingValuei)/rankingSpreading)

Since we know the mean value of Mi - Mj (µi - µj = 0), and its standard deviation
(sqrt(1/1 + 1/1) = sqrt(2)), we can calculate the probability that Mi - Mj is greater than any
given quantity by integrating under the curve of Mi - Mj's probability density function
from that quantity to infinity. A probability density function (pdf) is a function of a
random variable defined such that the probability that the random variable lies between
two values a and b approaches pdf(a)*b as b approaches zero. For normally distributed
random variables like z or Mi - Mj, the probability density function is the familiar “bell
curve”. Intuitively, integrating under this curve over some region is equivalent to slicing
the region into a series of discrete subregions with boundaries ai to ai+b, and adding up,
for each subregion, the probability pdf(ai)*b that the random variable is in that
subregion. We make b approach zero so that the slices are infinitesimally small, and we
get the probability that the random variable lies somewhere in the whole region.
For example, if Ci has the ranking value 101 and Cj has the ranking value 100,
then (rankingValuej - rankingValuei)/rankingSpreading = -1/ 2 = -0.5. To find P(Ci >>
Cj) = P(Mi - Mj > -0.5), we integrate under the probability density function of Mi - Mj

(illustrated in (77)) from -0.5 to +infinity, and find that P(Ci >> Cj) = 0.64.

107
(77) P(Ci >> Cj) = P(Mi - Mj > -0.5) = .64

pdf of Mi - Mj area under curve

from -.5 to ∞ = .64

-7 -6 -5 -4 -2 -1 1 2 3 4 5 6 7

2.11.2. Complex ranking requirements

First, to see why pairwise ranking probabilities involving the same constraints are not
independent—and therefore why complex ranking probabilities such as P(C1>>C2 AND
C3>>C2) can’t be calculated by simply multiplying P(C1>>C2) and P(C3>>C2))—

consider the three-constraint system illustrated in (78). If C1>>C2 in a particular instance,

then it is likely that C1’s selection point was chosen from the upper end of its distribution,
and thus C1>>C3 is more likely. Similarly, we must be careful in calculating P(C1>>C2
AND C3>>C2), since P(C1>>C2) and P(C3>>C2) are not independent. For example, in the

three-constraint system illustrated in (78), if C1>>C2 in a particular instance, then it is

likely that C1’s actual value was chosen from the upper end of its distribution, and thus
C1>>C3 is more likely.

108
(78) Pairwise rankings are not independent
C2
0.2

0.2 C3
C1
0.1

0.1

0.0

-2 -1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
-

To help see why this is so, consider the case in which C1, C2, and C3 all have the
same ranking value. For any two of these constraints, Ci>>Cj and Cj>>Ci are equally
likely. Therefore, any of the six possible total rankings (shown in )) is equally likely:

(79) Possible total rankings of three constraints

a C1 >> C2 >> C3
b C1 >> C3 >> C2
c C2 >> C1 >> C3
d C2 >> C3 >> C1
e C3 >> C1 >> C2
f C3 >> C2 >> C1

Consider then the probability of P(C1>>C2 AND C1>>C3): If P(C1>>C2) = 0.5

and P(C1>>C3) = 0.5 were independent, we could multiply 0.5 * 0.5 to get P(C1>>C2
AND C1>>C3) = 0.25. But C1>>C2 and C1>>C3 in 2 out of the 6 equally possible total

rankings (a and b), so P(C1>>C2 AND C1>>C3) is actually 2/6 = 0.3ι. When C1 is highly
ranked (as in a and b), both P(C1>>C2) and P(C1>>C3) are increased.
Another way of thinking about this example is that the requirement C1>>C2 AND
C1>>C3 is equivalent to the requirement that C1 be the highest-ranked of the three

109
constraints. Since each of the three constraints has an equal chance of being ranked
highest, C1's probability of being ranked highest is 1/3.
How then can complex probabilities be calculated? One straightforward method is
to integrate the joint probability density function (like a probability density function of a
single variable, except that its domain is ordered n-tuples consisting of one value for each
of the random variables) of all the random variables involved over the region of interest.
For example, to find P(C1>>C2 AND C1>>C3), integrate pdf(z1, z2, z3) over the region
where C1>>C2 and C1>>C3, which is the region where
(rankingValue1+rankingSpread*z1-rankingValue2)/rankingSpread > z2 and

(rankingValue1+rankingSpread*z1-rankingValue3)/rankingSpread > z3. This operation

takes all the points (z1, z2, z3) such that C1>>C2 and C1>>C3, and sums the probabilities
that each of those points could occur. It is also possible to estimate complex probabilities
by simulation (run many trials of the grammar). This section will describe the direct
method, which yields exact probabilities.
Because z1, z2, and z3 are standard normal random variables, their joint probability
density function pdf(z1, z2, z3) is just

pdf(z1, z2, z3) = pdf(z1)* pdf(z2)* pdf( z3) =

(e − z1 / 2π )(e − z2 / 2π )(e − z3 / 2π )
2 2 2
/2 /2 /2

This function cannot be integrated symbolically, so all the probabilities used here were
obtained from numerical integration in Mathematica.80

80
See §2.12 for an example.

110
 2.12. Appendix: Sample calculation in Mathematica

The first calculation performed above in §2.7.2 is

P((*NC8>>PU OR NASSUB >> PU OR (USELISTED>>PU & ENTRYLIN>>PU))

& (*NC8>>*[m OR NASSUB>>*[m OR (USELISTED>>*[m & ENTRYLIN>>*[m)))

In order to come up with limits of integration for the joint probability density
function for Mathematica, the pairwise rankings must all be joined by AND, not by OR.
We can achieve this by partitioning the complex ranking requirement into a series of
mutually exclusive ranking requirements that together cover all possibilities:

P((*NC8>>PU OR NASSUB >> PU OR (USELISTED>>PU & ENTRYLIN>>PU))

& (*NC8>>*[m OR NASSUB>>*[m OR (USELISTED>>*[m & ENTRYLIN>>*[m)))
= P¬(PU>>*NC8 & PU>>NASSUB & (PU>>USELISTED OR PU>>ENTRYLIN)
& *[m>>*NC8 & *[m>>NASSUB & (*[m>>USELISTED OR *[m>>ENTRYLIN))
= P¬((*[m>>PU & PU>>*NC8 & PU>>NASSUB & USELISTED>>ENTRYLIN &
PU>>ENTRYLIN)
OR (*[m>>PU & PU>>*NC8 & PU>>NASSUB & ENTRYLIN>>USELISTED &
PU>>USELISTED)
OR (PU>>*[m & *[m >>*NC8 & *[m >>NASSUB & USELISTED>>ENTRYLIN & *[m
>>ENTRYLIN)
OR (PU>>*[m & *[m >>*NC8 & *[m >>NASSUB & ENTRYLIN>>USELISTED & *[m
>>USELISTED))
= 1-(P(*[m>>PU & PU>>*NC8 & PU>>NASSUB & USELISTED>>ENTRYLIN &
PU>>ENTRYLIN)
+ P(*[m>>PU & PU>>*NC8 & PU>>NASSUB & ENTRYLIN>>USELISTED &
PU>>USELISTED)
+ P(PU>>*[m & *[m >>*NC8 & *[m >>NASSUB & USELISTED>>ENTRYLIN & *[m
>>ENTRYLIN)
+ P(PU>>*[m & *[m >>*NC8 & *[m >>NASSUB & ENTRYLIN>>USELISTED & *[m
>>USELISTED))

To calculate the first item in the sum, P(*[m>>PU & PU>>*NC8 & PU>>NASSUB
& USELISTED>>ENTRYLIN & PU>>ENTRYLIN), we want to integrate pdf(zPU, z*[m, z*NC8,
zNasSub, zUseListed, zEntryLin) over the region where *[m>>PU & PU>>*NC8 & PU>>NASSUB

111
& USELISTED>>ENTRYLIN & PU>>ENTRYLIN. These ranking requirements can be put in
terms of the zi. For example:

*[m >> PU
rankingValue*[m + 2z*[m > rankingValuePU + 2zPU
z*[m > (rankingValuePU - rankingValue*[m + 2zPU)/2

Using the following variable names and with the following ranking-value
differences,

m zm
P zPU
T z*NC8
S zNasSub
U zUseListed
E zEntryLin
rankingValuePU - rankingValue*[m = 0.691
rankingValuePU - rankingValue*NC8 = -3.817
rankingValuePU - rankingValueNasSub = -0.570
rankingValuePU - rankingValueEntryLin = -11.335
rankingValueEntryLin - rankingValueUseListed = -0.004

we can express P(*[m>>PU & PU>>*NC8 & PU>>NASSUB & USELISTED>>ENTRYLIN

& PU>>ENTRYLIN) as

−3.82 + 2 P −.57 + 2 P −11.34 + 2 P

+∞ +∞ 2 2 2 +∞

ò( ò ò ò ò ò (e /(2π ) ( 6 / 2) )∂U )∂E )∂S )∂T )∂m)∂P

( − P 2 − m 2 −T 2 − S 2 − E 2 −U 2 ) / 2
( ( ( (
− ∞ .69 + 2 P −∞ −∞ −∞ −.004 + 2 E
2 2

which in Mathematica notation is

N[Integrate[(e^((-P^2-m^2-T^2-S^2-E^2-U^2)/2)/(2π)^(3/2), {P, -Infinity,

+Infinity}, {m, (0.691250+2P)/2, +Infinity}, {T, -Infinity, (-3.816917+2P)/2}, {S,
-Infinity, (-0.570333+2P)/2}, {E, (-11.334917+2P)/2}, {U, (-003750+2E)/2,
+Infinity}]]

where the N[] function instructs Mathematica to calculate a numerical result.

112
113
 3. Simulating the adoption of a new word

3.1. Chapter overview

This chapter shows how the model proposed in Chapter 2 perpetuates lexical patterns as
new words come in to the lexicon, still using the example of nasal substitution. Section
3.2 gives evidence from loanwords that nasal substitution and the pattern of its
distribution have indeed been replicated in new words. Section 3.3 outlines a model of
speaker-listener interaction that draws on the probabilistic behavior of speakers and
hearers described in Chapter 2. Section 3.4 describes a simulation of the speech
community designed to test whether the model in §3.3 can really produce the desired
results on new words. Section 3.5 gives the results of the simulation.

3.2. Assimilated loanwords

It is clear from examining the loanword vocabulary in Tagalog that new words
sometimes become listed with nasal substitution. English’s (1986) dictionary contains
only four potentially substituting derivatives of obstruent-initial English loanword stems,
and none of these are substituted. But Spanish stems have been in the lexicon longer and
have had more opportunity to accumulate derived forms. There are 152 potentially
substituting derivatives of obstruent-initial Spanish loanword stems.81 Of these, 97
substitute. This suggests that nasal substitution has been productive relatively recently—
productive not necessarily in the sense that it applies frequently to novel words, but in the
sense that as a novel word becomes assimilated into the lexicon it may become nasal-
substituted.

81
Including some indigenous Mexican words presumably imported through Spanish.

114
 There are too few examples to compare the rates of substitution for various
affixes in the Spanish loanword vocabulary to rates in the native vocabulary, but
combining all the affixal patterns together, we can get a rough idea of how well the
Spanish words are following the native patterns: the voicing effect seems to have been
present, and there is a higher rate of substitution for b than for d and g.82

(80) Substitution rates for Spanish stems, all affixal patterns combined

100% 4
5 8
80% 4
4 3
12
60% No
6
2 6 Vary
40% 36 35 1
17 Yes
20% 8
1
0%
p t/s k b d g ?
stem-initial obstruent

Assuming that the grammar of current Tagalog is fairly similar to the grammar of
Tagalog at the time these derived forms of Spanish stems were established (anywhere
from the mid sixteenth century to the present day83), we can use substitution rates in

82
Note the small number of words derived from Spanish stems beginning in d and g, despite the fact that
initial d, at least, does not seem to be underrepresented in Spanish (52 pages for p, 26 for t, 66 for [k], 21
for b, 39 for d, and 12 for [g] in The American Heritage Larousse Spanish Dictionary 1986—these page
counts are only rough approximations to root or stem counts, since many prefixed words are included). In
contrast, root-initial d and g, though not ill-formed, are quite underrepresented in the native Tagalog
vocabulary (see (36) in §2.4.5). Could Tagalog speakers somehow be selecting loans in such a way as to
perpetuate the lexical statistics of the native vocabulary?
83
The coining and establishment of a derived form of a Spanish stem could have occurred long after the
adoption of the stem itself. The relative scarcity of derived forms of English stems suggests that the
establishment of derived forms of loanword stems tends to occur long after the borrowing of the stem.

115
Spanish loanwords as an indication of how newly coined derived forms should eventually
develop: despite low initial rates of substitution, many words must eventually come to be
listed as substituted. In addition, given the Spanish data, a stem’s chance of eventually
being listed as substituted is probably influenced by the voicing effect, and possibly
influenced by the place effect.
This chapter proposes a model of the speech community—and a simulation of the
speech community under that model—that produces the following result: novel derived
forms have a low initial rate of substitution, but as they come to be listed, the proportion
that are substituted reflects proportions in the lexicon.
The crucial assumptions of the model are that (i) speakers generate outputs according to
the stochastic grammar they have learned from the lexicon (§2.6), (ii) listeners make a
probabilistic guess as to what input the speaker was using (§2.8.2) and update their
lexicons accordingly, adding new listed forms and changing listedness values of existing
forms. The results of §2.8.2 will be crucial in ensuring that the large number of
unsubstituted forms produced early in a word’s life does not guarantee that the word will
end up being listed as unsubstituted.
Many of the parameters of the simulation (such as values of constants) were
arrived at by trial and error. Some parameters could be changed greatly and the
simulation would still work; others’ exact values are crucial, even though there may be
no a priori justification for those values. Therefore, the simulation should be considered
an existence proof, rather than an assertion about the details of lexical evolution: it is
possible to create a successful simulation of lexical evolution in the speech community

Note that the past few centuries represent a very small portion of the history of nasal substitution.
Dempwolff (1969) traces nasal substitution to Proto-Austronesian itself, which would make it at least 5000
years old (Bellwood 1979).

116
that is consistent with the model proposed here and in Chapter 2, but the example here
may not be the only possibility, and may not be the possibility that is closest to reality.

117
 3.3. Model of the speech community

The structure of the model will be made more explicit in §3.4, which gives details of the
simulation, but essentially it is a synthesis of §2.7 and §2.8. When a derived form of a
stem does not yet exist, speakers who wish to utter it have no choice but to concatenate
morphemes on the fly. For example, if a speaker wishes to express the idea ‘one whose
job it is to puntol’, she must combine the morphemes /maN/, /REDCV/, and /puntol/. Given
the grammar in §2.6, there is an approximately 40% chance that the result of this
concatenation will be [mamumuntol], and a 60% chance that the result will be
[mampupuntol].
The speaker’s interlocutors hear either [mamumuntol] or [mampupuntol]. In order
to decide what adjustments, if any, to make to their mental lexicons, they must guess
what underlying form the speaker was using. Employing the Bayesian reasoning
discussed in §2.8, a person who hears [mamumuntol], and who has no listed word for
‘one whose job it is to puntol’ will guess (incorrectly) that the input was /mamumuntol/
94% of the time; she will guess (correctly) that the input was /maN/+/RCV/+/puntol/ 6%
of the time. When the guess is /mamumuntol/, the listener creates a lexical entry for
/mamumuntol/, and gives it a weak initial strength84—which means that this new lexical
entry is not yet very likely to be available to the listener for future utterances (as it builds
in strength, however, it will begin to influence the listener’s own utterances, and thereby
the lexicons of her listeners). When the guess is /maN/+/RCV/+/puntol/, the listener does
not update her lexicon at all—she already knows these morphemes. Similarly, if the same
person hears [mampupuntol], she will guess (incorrectly) that the input was

84
I assume the same initial strength for every once-heard word. See §3.7 for listedness-updating functions.

118
/mampupuntol/ 33% of the time and create a new lexical entry (which will eventually
influence her own speech); she will guess (correctly) that the input was
/maN/+/RCV/+/puntol/ 65% of the time and do nothing.
After there have been a few occasions to say ‘one whose job it is to puntol’, many
members of the speech community will have formed (weak) lexical entries for
/mamumuntol/ and/or competing /mampupuntol/, and their behavior as speakers and
listeners will be slightly changed: along with /maN/+/RCV/+/puntol/, /mamumuntol/ and
/mampupuntol/ will now be available occasionally as inputs to speakers, changing
slightly the frequencies at which speakers produce [mamumuntol] and [mampupuntol].
And listeners will be slightly more likely to guess /mamumuntol/ and /mampupuntol/ as
inputs, further strengthening them.
I assume in addition that lexical entries that differ only phonologically are in
competition:85 if a listener has lexical entries for both /mamumuntol/ and /mampupuntol/,
when she hears an utterance that she takes to be derived from /mamumuntol/, she will
both increase the strength of /mamumuntol/ and decrease the strength of /mampupuntol/
(and vice versa when she hears an utterance that she takes to be derived from
/mampupuntol/). Disparities in strength between competing lexical entries tend to grow
over time, because the stronger /mamumuntol/ becomes, the less likely the listener is to
“hear” /mampupuntol/, since P(/mampupuntol/) decreases as Listedness(/mamumuntol/)
increases (see §2.8.2).86

85
Cf. the blocking effect (Aronoff 1976): if one member of a stem’s paradigm has a certain meaning (e.g.,
fury), synonymous derivatives are blocked (*furiousness, *furiosity)
86
A prediction of this assumption (that the relationship between different pronunciations is antagonistic) is
that words with variable pronunciations even within speakers should tend to be low-frequency. For high-
frequency words, there should be enough tokens that any small difference in strength between the
competing lexical entries will eventually produce one clear winner.

119
 The usual result of competition between /mamumuntol/ and /mampupuntol/ is
that eventually one will emerge as strongly listed, and the other as very weakly listed.87
For example, with a p-initial stem like puntol, because of the rates at which listeners
guess that the speaker was using each input, the lexical entry /mamumuntol/ initially
tends to get strengthened more than /mampupuntol/. Speakers are then more likely to use
/mamumuntol/ as an input (with [mamumuntol] nearly always the output, because of
high-ranking ENTRYLINEARITY), with the result that listeners guess /mamumuntol/ even
more often, widening the gap between /mamumuntol/ and /mampupuntol/. A disparity in
strength between /mamumuntol/ and /mampupuntol/ in the early stages, then, if consistent
across the speech community, is self-reinforcing and leads to the eventual adoption of the
stronger option. A member of the next generation may not even form a lexical entry for
the weaker option at all (unless she hears a speech error such as /mamumuntol/ →
[mampupuntol] and guesses that the input was /mampupuntol/).
Which lexical entry—the substituted or the unsubstituted—eventually wins out
depends on an accumulation of many chance decisions by speakers and listeners. Which
lexical entry will tend to win out depends on the rate at which the on-the-fly input is
pronounced as substituted, and the rate at which listeners guess that a substituted
utterance derives from a single input versus the rate at which listeners guess that an
unsubstituted utterance derives from a single input.

87
In none of my simulations have different pronunciations remained in competition indefinitely, although
the situation is possible if P(/unsubstituted/ | [unsubstituted]) and P(/substituted/ | [substituted]) are close
enough to equal.

120
 3.4. How the simulation works

I constructed a simulation of a speech community following the model outlined above, in

order to verify that new words would eventually be assimilated into the lexicon as
substituted or unsubstituted at rates similar to those seen for Spanish-stem words in (80).
The simulated speech community has ten slots for members. The simulation
begins with eight slots filled: there are two people aged 20, two aged 40, two aged 60,
and two aged 80. Each person has a grammar consisting of ranking values learned by the
Gradual Learning Algorithm from exposure to a mini-lexicon, as in §2.6. Each grammar
represents one run of the Gradual Learning Algorithm, so each is slightly different.
Each run of the simulation involves the community’s deciding how to pronounce
one new word. On every trial within a simulation, one person is selected randomly as the
speaker, and two others as listeners. The speaker generates a constraint ranking (based on
the ranking values in her grammar), and produces the optimal candidate for the word
under consideration, given that ranking and the available inputs. Which inputs are
available is also determined probabilistically: the on-the-fly input (e.g.,
/maN/+/RCV/+/puntol/) is always available; the availability of inputs like /mampupuntol/
and /mamumuntol/ depends on the strength of those lexical entries (Listedness).
Each of the two listeners first decides whether or not to pay attention to the
speaker, based on a function of the speaker’s age (described in §3.7) such that younger
speakers are likely to be ignored (by age 14, the speaker is almost certain not to be
ignored). This prevents adults’ lexicons from being overly disrupted by children’s errors.
The listener then makes a probabilistic guess as to what input the speaker was using. If

121
/mampupuntol/ or /mamumuntol/ was the optimal input, its listedness is increased (and
the listedness of the other is decreased88) according to the function described in §3.7.
The details of the listener’s decision procedure require some elaboration. In
§2.8.2, the prior probabilities of inputs, and the probabilities of outputs given inputs, were
combined according to Bayes’ Law to derive the probabilities of inputs given outputs.
The prior probabilities of inputs were relatively simple to calculate: they were a fairly
simple function of the productivity of a construction and the strengths of relevant lexical
entries. But calculating the probabilities of outputs given inputs required either
integrating over many-dimensional areas with complicated boundaries. It might be
implausible to require listeners to perform multivariable calculus89 on every utterance, so
the simulation employs a simpler method, which produces values for P(output|input) that
are, on average, nearly accurate. As long as the simulation works, nearly accurate values
are in no way undesirable, since we have no direct evidence as to the accuracy of the
values for P(output|input) that listeners might use. The method used in the simulation is
given in (81).90

(81) Listener’s procedure for estimating P(output|inputi)

For each input being considered,
a. Generate a constraint ranking from the grammar
b. Run the input through the constraint ranking
c. If the result is the output under consideration, EstimatedP(output|inputi) = 1.
Otherwise, EstimatedP(output|inputi) = 0.

88
If the listedness of the input not heard is not decreased, eventually every word ends up with both forms
listed, and thus displays variable behavior. This seems empirically implausible, because it predicts that the
greatest variation will be found among high-frequency words.
89
Each P(output|input) calculation takes up to one minute in Mathematica.
90
Of course, the simulation also works if exact probabilities are used, so it is not a problem if humans are
in fact able to perform the exact calculations

122
 EstimatedP(output|inputi) is plugged into Bayes’ Law, as shown in (82):

(82) Estimating P(inputi|output)

EstimatedP(inputi|output) = P(inputi) * EstimatedP(output|inputi)
P(output)
where P(output) is the sum of P(inputi) * EstimatedP(output|inputi) for all inputi.

On any given occasion, EstimatedP(output|inputi) is far from the actual

P(output|inputi), of course (it is always either 0 or 1). But over many trials, the average
EstimatedP(output|inputi) is equal to the actual P(output|inputi). The source of the slight
inaccuracy in EstimatedP(inputi|output) over time is cases in which
EstimatedP(output|inputi) = 0 for all inputi: in these cases, P(output) (which is in the
denominator in (82)) would equal zero. We can either throw such cases out, or assign an
equal value to each EstimatedP(output|inputi); either approach skews the average values
of EstimatedP(inputi|output) somewhat. Fortunately, all-zero cases are rare enough in this
simulation (less than 1% of all trials) that the inaccuracy is minimal (less than a
percentage point).
Every 50 utterances of the word in question (1 “year”), each person has a chance
of “dying” and leaving her slot open; this chance increases with age. If there are empty
slots (as there are at the beginning of every simulation), there is a chance that a new
person may be “born” to fill it. Younger speakers are (unrealistically, of course) assumed
to have adult grammars and adult morphological parsing ability, but what they lack is an
adult lexicon: a newborn person in the simulation has no listed form for the word being
simulated. If the adults have already agreed on a listed form, the new person will quickly
acquire it, since she will be exposed to quite consistent data.

123
 3.5. Simulation results

The simulation was run for 120 novel words, 20 each for p, t, k, b, d, and g. Note that
since the mini-grammar used here is sensitive only to the initial segment of the stem,
there can be no intrinsic difference between one novel stem beginning with p and another
novel stem beginning with p. The reason for running the simulation multiple times is that
chance factors can lead to different results of different trials. Each word was used by the
speech community for 150 “years”. By that point, in every trial, every member of the
speech community over 20 years old was producing one pronunciation consistently, and
all were in agreement.91 (83) shows the results, with the distribution of substitution
among Spanish loans and in the whole lexicon repeated from (8) and (80) for
comparison.

(83) Simulation results for novel words after 150 “years”

100%
90% 3

80%
70%
60%
unsubstituted
50% 20 20 20 20 20
17 substituted
40%
30%
20%
10%
0%
p t k b d g

91
At earlier points in the simulation, though, there was always considerable within- and across-speaker
variation. The implication for real words with variable pronunciations is that they have not been used

124
(84) Nasal substitution in real Spanish loans

100% 4
5 8
4
80%
4 3
12
60%
2 6
40% 36 1
35
17
20% 8

0%
p t/s k b d g

(85) Nasal substitution in entire Tagalog lexicon

100% 10 26 17
100 70 97
80%

60%

40%

20%
253 430 185 177 25
0% 1

p t/s k b d g

enough to acquire a stable pronunciation. The model predicts, then, that words with variable pronunciations
should be low-frequency.

125
 Clearly, b-initial stems are not behaving as expected. The desired result was that
they be substituted about half the time. It turns out that rate of substitution on b produced
by the grammars learned in §2.6 is too low for a b-initial stem to end up substituted, even
with the listener bias described in §2.8.2. It is possible to construct grammars that, when
used in the simulation, produce the desired results for b and the other segments. For
example, the handcrafted grammar in (86) produces the results in (87).

(86) Hand-crafted grammar to produce the desired results for /b/-initial stems
Constraint Ranking value
USELISTED 122
ENTRYLINEARITY 128
MORPHORDER 105
PU 100
*NC8 104
NASSUB 105
*[N 106
*[n 105
*[m 102

(87) Simulation results using the handcrafted grammar in (86)

100%
1
90%
80%
5
70%
60%
unsubstituted
50% 10 10 10 10
9 substituted
40%
30%
5
20%
10%
0%
p t k b d g

126
Therefore, I do not regard the failure of b to substitute as a major problem for the model;
it may be that with small changes to the learner, or somewhat different learning data (for
example, token rather than type frequencies), grammars would be learned that would
produce the desired results. Note also that the model did not always produce all-or-
nothing results—as shown in (83), k-initial stems were substituted 85% of the time. So it
is not the case that mixed results such as those desired for b are difficult to obtain, just
that the rate of substitution for b generated by the learner-generated grammar was too low
to get a mixed result.

3.6. Chapter summary

This chapter has presented a model of the speech community that perpetuates lexical
patterns on new words, using the case of nasal substitution. The crucial element is the
listener bias in favor of nasal-substituted lexical entries discussed in §2.8: this bias allows
new words to eventually become listed as nasal substituted even if substitution is not the
majority pronunciation when the word is new.

127
 3.7. Appendix: Functions used in the simulation

(88) Deciding whether to pay attention to a speaker

P(paying attention) = 1 / (1 + e8 - speaker’s age)

The younger the speaker, the less likely that the listener will pay attention to
her utterance (a prerequisite for the listener’s updating her lexicon in response
to the speaker’s utterance)

(89) Prior probabilities of inputs (see §2.8.2)

P(/on-the-fly/) = 1/((1+e-3+6*Listedness(whole word))*(1+e3-6*Productivity(construction)))
where Listedness(wholeword) is the greater of Listedness(substituted) and
Listedness(unsubstituted)
P(/substituted/) = (1-P(/on-the-fly /))*F(/subst./)/(F(/subst./)+F(/unsubst./))
P(/unsubstituted/) = (1-P(/on-the-fly/))*F(/unsubst./)/(F(/subst./)+F(/unsubst./))
where
F(word) =
1/((1+e2-6*Lstdnss(word))(1+e-4+6*Lstdnss(competingword))(1+e3-6*ProportionThatSubst))

(90) Updating listedness

Listedness(word) = 1 / (1 + e4 - 0.15 * TimesHeard(word))
TimesHeard(word) is not a literal record of the number of times a particular
(pronunciation of a) word has been heard, and is not even stored long-term.
Instead, whenever the listener decides to increase a word’s listedness, she
calculates TimesHeard(word) from Listedness(word):
TimesHeard(word) = (4 - ln (1/Listedness(word) - 1)) / 0.15
She then increases TimesHeard(word) by 1, and recalculates Listedness(word).
The value for TimesHeard(word) can then be thrown away. When the listener
wants to decrease a word’s listedness, she performs the same procedure, but
instead of increasing TimesHeard(word) by 1, she decreases it by 0.5.
This means that TimesHeard reflects not the actual number of times an input has
been heard, but rather the cumulative effects of hearing the input and hearing
competing inputs.

128
 4. The model as applied to vowel height alternations

4.1. Chapter overview

This chapter applies the model developed in Chapters 2 to a different lexical regularity,
the distribution of exceptions to vowel raising in Tagalog loanwords. The regularity here
is of some intrinsic interest, and the analysis proposes a new phonological mechanism.
But the vowel-height case is also important to the main arguments of this dissertation
because it differs from nasal substitution in three respects. First, the pattern is found only
within loanwords, so the argument that it comes from the grammar (rather than from
statistical generalizations over the lexicon) is stronger. Second, the pattern itself is quite
abstract: in nasal substitution, words with the same stem-initial consonant behaved
similarly, but in vowel raising it is words whose internal similarity is of the same degree
that behave differently. Again, this argues for representation in the grammar rather than
emergence from the lexicon. Third, in nasal substitution different derivatives of the same
stem could behave differently, but in vowel raising the behavior of one relevant
derivative predicts the behavior of the rest; this has consequences for the structure of
lexical entries. The first and second points are taken up again in Chapter 5.
Section 4.2 presents the data on vowel height in Tagalog and the types of
exceptions that are found. Section 4.3 gives an analysis of those basic facts. Section 4.4
introduces Aggressive Reduplication, the mechanism that will be used to explain the
distribution of exceptions in loanwords, which is presented in §4.5. Section 4.6 argues
that the Aggressive Reduplication analysis is superior to other possibilities by
demonstrating that Aggressive Reduplication makes a prediction that other analyses do
not, and that that prediction is correct. Section 4.7 considers how vowel height should be

129
represented in lexical entries. Sections 4.8 and 4.9 discuss what a grammar for vowel
raising would look like, and how it could be learned.

4.2. Vowel height in Tagalog

In most of the Tagalog vocabulary, mid and high vowels are in near-complementary
distribution. Mid vowels are found only in final syllables, and [u] is found only in
nonfinal syllables. [i] can occur anywhere, and many words have [i] and [e] in free
variation in the final syllable. Examples in (91) illustrate some typical monomorphemic
native words.

(91) Distribution of mid and high vowels

ka@ùlos ‘grain leveler’
bagjo@ ‘typhoon’
baba@ù?e ‘woman’
bu@ùkas ‘tomorrow’
bunso@? ~ bunso@ ‘youngest child’
hi@ùbi ‘small dried shrimps’
ditdi@t ‘torn into strips’
pati@d ‘cut off’
ga@ùbi ~ ga@ùbe ‘taro’
pante@ ~ panti@ ‘dragnet’

Suffixation induces alternation, by making syllables that were once final

nonfinal:92

92
Tagalog has just two native suffixes, -in and -an, whose most common and productive function is to form
verbs (-in usually forms direct-object-focus verbs; -an usually forms indirect-object-focus verbs). These
suffixes are also used alone and in combination with prefixes in various other morphological constructions.
There are also loan suffixes such as -ero and -ista that sometimes combine with native stems.

In most suffixal constructions, stress and length (if any) are shifted one to the right: if the bare
stem has final stress, stress falls on the suffix; if the bare stem has penultimate stress and length, the penult
of the suffixed form has stress and length. This alternation could be thought of as preserving the (right-
aligned) original prosody of the stem. Some suffixal constructions induce different shifts or none at all, and
loanstems with long, stressed closed penults (very rare in native words) often behave differently.

130
(92) Suffixation-induced alternations
ka@ùlos ‘grain leveler’ kalu@ùs-in ‘to use a grain leveler on’
?abo@ ‘ash’ ?abu-hi@n ‘to clean with ashes’
baba@ù?e ‘woman’ ka-baba?i@ù-han ‘womanhood’
siste@ ‘joke’ sistiù-hi@n ‘to joke’

There are exceptions to all the generalizations just made, although they are
relatively few in the native vocabulary. There are many more exceptions in the loanword
vocabulary, which is discussed below.93
There are two classes of systematic exceptions to the generalization that mid
vowels are found only in final syllables. For completeness, they are described here and
accounted for to some extent, but they are not the main area of interest. First, in nonfinal
syllables containing an [aw] or [aj] diphthong,94 coalescence can occur, producing a
long/stressed mid or high vowel of the same backness and rounding as the glide. This
sometimes produces a nonfinal mid vowel:

(93) Vowel coalescence

?ajwa@n ~ ?e@ùwan ‘I don’t know’
hinta@j ka ~ ta@jka ~ te@ùka ‘Wait!’ (ka = ‘you’)
bajawa@N ~ bajwa@N ~ be@ùwaN ‘waist’
ka?unti@? ~ kawnti@?~ ko@ùnti? ‘a little’
ba?ina@t ~ bajna@t ~ bi@ùnat ‘relapse’
sajna@t ~ si@ùnat ‘slight fever’

The h that appears when a vowel-final stem is suffixed can be thought of as (i) epenthetic, (ii) part
of a postvocalic allomorph of the suffix, or (iii) part of the suffixal allomorph of the stem.
93
I make no claim that there is (or is not) a synchronic difference between the native and loanword
vocabularies; the native vocabulary is discussed first in order to make clear the basic pattern.
94
For coalescence to be possible, the glide must not be obligatorily the onset of the following syllable. The
diphthong may, however, be in free variation with a vowel-glide-vowel sequence (as in bajawa@N ~
bajwa@N). Or, it may be in free variation with a vowel-glottal-vowel sequence (as in ba?ina@t ~ bajna@t).

131
 The second systematic source of nonfinal mid vowels is V?V sequences in which
both vowels are nonlow. In these sequences, the vowels must match in backness, as
illustrated in (94).95 If the vowels are back, often the first is high and the second is mid,
but often both are mid. If the vowels are front, both vowels are usually high, but
occasionally both are mid.96

(94) Transglottal vowels

po?o@k ‘place’
pu?o@n ~ po?o@n ‘master’
su?o@t ‘clothing’
me?e@? ‘bleat’
le?e@g ~ li?i@g ‘neck’
bi?i@k ‘piglet’

Finally, there are also seemingly unsystematic exceptions in the native

vocabulary: words with mid vowels in nonfinal syllables,97 words whose final vowels
remain mid under suffixation, and words with final-syllable [u]. The list in (95) is close
to exhaustive: it includes all of the relevant items that were found in a database of the

95
Sequences not matching in backness might be absent because historical i?o, u?i, and u?e sequences have
become ijo, uwi, and uwe.
96
Why should a medial glottal stop license a nonfinal mid vowel? Steriade (1987) identifies translaryngeal
harmony (analyzed as spreading of a supralaryngeal feature node) as a cross-linguistically widespread
phenomenon in which total identity (except in laryngeal features) between vowels is encouraged across [?]
and [h]. Tagalog may not be a case of such harmony, in which [?] and [h] are supposed to behave the same:
I found only one case of a nonfinal mid vowel before [h] among the disyllabic roots (bohol ‘(shrub
species)’) compared to 13 cases of a nonfinal mid vowel before [?].

Whatever the historical origin, [o?o] and [e?e] sequences might synchronically be analyzed as
long, glottalized vowels (Steriade arrives at a similar conclusion for Yurok)—in that case, they are final,
and so should rightly be mid. These roots would have to escape the two-syllable minimal root requirement,
however.
97
For brevity, I will sometimes refer to a vowel in a word-final syllable as “final” even if it is followed by a
consonant. I will use “nonfinal” for a vowel in a nonfinal syllable (not for a vowel that is in a final syllable
but followed by a consonant).

132
4619 disyllabic98 roots in English’s 1986 dictionary, as well as all the relevant longer
words that I have encountered. Note that many of the words with nonfinal mid vowels
appear to have CV- or CVC- pseudoreduplication,99 and that the words that fail to raise
under suffixation have a nonfinal mid vowel of the same backness as the final mid vowel
(these facts will be relevant in the explanation of the distribution of exceptions).100

(95) Exceptional native words

Nonfinal [o]
pseudoreduplicated
?o@ù?o ‘yes’
toto?o@ ‘true’
ko@ùkok ‘crow of rooster; chickie’
to@ùto? ~ to@ùtoy ‘(affectionate term of address for little boy)’
goNgo@N ‘gruntfish sp.’
kato@ùto ‘comrade’
bako@ùko ‘fish sp.’
other
boho@l ‘shrub sp.’
?o@ù@ la ‘eagerness’
ko@ùkak ‘croak of frog’
Nonfinal [e]
pseudoreduplicated
de@ùde ‘baby bottle’
me@ùme ‘beddie-bye’
keNke@N ‘sound made by beating bottom of frying pan’
ne@ùne? ~ ne@ùneN ~ ni@ùni? ‘(affectionate term of address for little girl)’
heùùlehe@ùle ‘pretense of not liking’
he@ùle ‘lullaby’
other
ke@ùrwe ‘cricket’

98
The database was limited to disyllabic roots because longer roots are generally polymorphemic (at least
historically), and shorter roots are generally clitics.
99
Pseudoreduplication is discussed further in §4.4. What I mean by the term is that the last two syllables
are identical (except that the penult may lack the ultima’s coda), but no productive morphological process
of reduplication is at work.
100
Several of the words in (95) are baby-talk words, interjections, or onomatopoeic/mimetic words. As in
other languages, some well-formedness requirements seem to be relaxed in the “peripheral” vocabulary of
Tagalog (see Itô & Mester 1995).

133
 le@ùteN ‘cord’
ke@ùtoN ‘leprosy’
(raises when suffixed: ketu@ùN-in ‘to have leprosy’)
te@ùpok ‘victimized by hooligans’
he@ùto ~ ?e@ùto ‘Here it is!’
be@ùlat ‘Serves you right!’
le@ùkat ‘How could you?!’
pe@ùklat ‘scar’
te@ùkas ‘swindler’
se@ùlaN ~ se@ùlan ‘delicacy’
kula@ùlat ~ kule@ùlat ‘last’

Mid vowel that stays mid under suffixation101

de@ùde ‘baby bottle’ padede@ù-hin ‘give a baby a bottle’
toto?o@ ‘true’ toto?o@ù-hin ‘to be sincere’
po?o@t ‘hatred’ ka-po?ot-a@n ‘to hate’
(and all other o?o words; found no e?e words with suffixed derivatives)
Final [u]
sampu@? ‘ten’
?i@ùmus ‘headland’
kaso@j~ kasu@j ‘cashew’
bagko@s ~ bagku@s ‘on the contrary’
bambo@ ~ bambu@ ~ banbu@ ‘bludgeon’
da@ùto? ~ da@ùtu? ‘chieftain’
labi@w ~ labju@ ‘weeds that grow in a burned field’

101
These are the only exceptions to raising under suffixation that I have found. A vowel can also be made
nonfinal through disyllabic reduplication, and here raising is often optional, even in native words (e.g.,
ha@ùlo ‘mix’, haùlu-ha@ùlo or haùlo-ha@ùlo ‘(frozen desert/drink)’). The reason for this optionality may be the
presence of a prosodic break between the reduplicant and the base (see discussion following (101)).

134
 4.3. Analysis of vowel lowering/raising

Before moving on to the main subject of this chapter—exceptions to vowel height

alternations—I will briefly offer an analysis of the distribution of vowel height itself,
although no functional motivation.102 I propose the following phonotactic constraints:

(96) *NONFINALMID
* Word
σ σ
|
V
|
[-high, -low]
[-high, -low] vowels in nonfinal syllables are forbidden.

(97) *FINAL[u]
* σ ]Word
|
V
|
[+high, +back]
[+high, +back] vowels in word-final syllables are forbidden.

102
The vowel height alternations caused by suffixation are not nearly as ancient as nasal substitution (see
below) but phonetic motivation is still hard to find. Crosswhite (1999) proposes that lower vowels’ greater
sonority (greater jaw opening), makes them better suited to be long. Final lengthening might result in final-
syllable lowering (cf. Yokuts, whose long, high vowels lower—Newman 1944). But although Tagalog may
have some final lengthening, it also has many long vowels in nonfinal syllables, and these long vowels do
not lower (e.g., bu@ùkas ‘tomorrow’, hi@ùbi ‘small dried shrimps’). Compare Yidiø (Dixon 1977), whose high
vowels lower somewhat in short final syllables, and lower all the way to mid in long final syllables,
although nonfinal long vowels do not lower at all.

Could length-driven vowel lowering have arisen at a stage when there were no nonfinal long
vowels? Zorc (1972, 1983) argues for contrastive “accent” (length and/or stress) in Proto-Philippine, with
some words having a stressed, long penult and others a short penult and a stressed final syllable. Tagalog
vowel lowering is a fairly recent innovation, not shared by all Central Philippine languages; so if Zorc is
right, long penults would already have existed when vowel lowering began.

135
 The operation of the constraints is illustrated using Inkelas, Orgun, and Zoll’s
(1997) underspecification approach to exceptionful alternations (the analysis will be
modified in §0). Exceptionally high or mid vowels are fully specified as [+high] or [-
high]; vowels whose height is predictable are underspecified (indicated in the tableaux
below by capital O or E), with markedness constraints filling in the appropriate height, as
illustrated in (98).
In the first tableau, IDENT-IO[HIGH]103 is satisfied by all four fully-specified
candidates (a, b, d, e), since no height value is specified in the input. Thus it is the
markedness constraints that decide the matter, selecting [-high] for the vowel when it is
final (a), and [+high] when it is nonfinal because of suffixation (d) (the dashed line
between the two markedness constraints’ columns indicates that there is no evidence for
ranking one above the other).
The second and third tableaux show that a vowel must be mid if it is so specified
underlyingly, even when it is nonfinal. Raising an underlyingly [-high] vowel would
violate both IDENT-IO[HIGH]. Similarly, the fourth tableau shows that a final vowel must
be [u] if it is so specified underlyingly, because to make it mid would violate MAX[HIGH]
and DEP[HIGH].

103
Perhaps filling in feature values incurs some faithfulness violation; if so, assume the constraint violated
is low-ranked. Assume also that a high-ranking constraint prevents underspecified segments on the surface:
some value must be filled in.

136
(98) Tableaux illustrating underspecification analysis
Predictable alternation
/kalOs/ IDENT-IO[HI] *FINAL [u] *NONFINAL MID
a . [kalos]
b [kalus] *!
/kalOs+in/ IDENT-IO[HI] *FINAL [u] *NONFINAL MID
d . [kalusin]
e [kalosin] *!

Mid vowel in nonfinal syllable

/tekas / IDENT-IO[HI] *FINAL [u] *NONFINAL MID
g . [tekas] *
h [tikas] *!

Nonalternating mid vowel

/dede/ IDENT-IO[HI] *FINAL [u] *NONFINAL MID
i . [dede] *
j [dedi] *! *
/dede+hin/ IDENT-IO[HI] *FINAL [u] *NONFINAL MID
k . [dedehin] **
l [dedihin] *! *

[u] in final syllable (nonalternating)104

/sampu?/ IDENT-IO[HI] *FINAL [u] *NONFINAL MID
m . [sampu?] *
n [sampo?] *!
/sampu?+in/ IDENT-IO[HI] *FINAL [u] *NONFINAL MID
o . [sampu?in]
p [sampo?in] *! *

Under this analysis, we could generalize *FINAL[u] to *FINALHIGH: a stem with a

final [e] that becomes [i] under suffixation would be underspecified (like /kalOs/); a stem
with final [i] would be specified [+high]. There would simply be many, many stems with
final i that would have to violate *FINALHIGH in unsuffixed form.

104
As for front vowels in final syllables, either words are always listed as having either [e] or [i], or some
are listed and the rest have their value filled in by some constraint(s).

137
 The analysis is not complete, however, because when we examine the data from
loanwords, it becomes apparent that there are regularities in the distribution of
exceptions. As with nasal substitution, the solution proposed will be the presence of low-
ranking constraints, which in this case are of some interest in themselves.

138
 4.4. Aggressive Reduplication

Before the loanword data are described, this section introduces the mechanism that is
invoked to explain them. I propose that, in all languages, speakers tend to construe
similar syllables (or other units) as being in correspondence (pseudoreduplicated). Such a
construal can result in the enhancement or preservation of internal similarity.
For example, in English there are sporadic examples of (often accidental) word-
internal similarity between feet or syllables that gets increased, resulting in lexical drift.
In (99) are shown some examples. Attestedness was verified by searching on the World-
Wide Web (using Altavista, www.altavista.com) for nonstandard spellings that reflect the
similarity-enhanced pronunciation. Clearly, some of the newer pronunciations are
widespread; others may be sporadic errors.

(99) Similarity enhancement in English105

Nonstandard hits Standard hits
orangutang 773 orangutan 6913
orangoutang 20 orangoutan 17
Okeefenokee 392 Okefenokee 2586
[ÇoUk«f«ÈnoUki]
smorgasborg 394 smorgasbord 17,228
Inuktituk 125 Inuktitut0 2569
sherbert106 about 1000 sherbet 7083
pompom107 2072 pompon 2066
Abu D(h)abu108 4 Abu Dhabi 21,234
Abi D(h)abi 4
asterist 12 asterisk 176,510
askerisk109 15

105
Of course, some of the hits may be from other languages in which the same lexical drifts and errors have
taken place (possibly for the same reasons), and from non-native writers of English.
106
4496 hits, but about ¾ (based on inspection of the first few dozen) were personal names.
107
This spelling appears in dictionaries.
108
Nonstandard spellings of Abu Dhabi were individually verified to ensure that they did refer to the city.

139
 Tagalog has a large vocabulary of words that have even more internal similarity
than orangutan or Inuktitut. These are the pseudoreduplicated words, which are generally
of the form CV-CVC or CVC-CVC; some pseudoreduplicated words also have
pseudoprefixes and pseudoinfixes. Some typical examples are given in (100).110

(100) Pseudoreduplicated words in Tagalog

CV-CVC
ki@ùkig ‘cleaning of ears’
gaga@d ‘mimicry’
pu@ùpog ‘pecking hard; repeated kissing’
CVC-CVC
bakba@k ‘peeled off’
damda@m ‘feeling’
saksa@k ‘stab wound’
CVC-a-CVC
busa@ùbos ‘slave’
sibasi@b ‘violent attack by animal’
pseudoprefixed (?u-, tu-, ku-, bu-, lu-, mu-, ti-, gi-, li-, ?ali-, bali-, sali-, and ja-)
bukadka@d ‘fully opened’
kulimli@m ‘overcast’
gipuspo@s ‘very dispirited’

109
The ratio of nonstandard to standard spellings of asterisk may seem low enough to be the result of
typographical errors or uninteresting perception errors. As a control against perception errors on the part of
the writer, I searched for asterisp and asperist, and found no hits. As a control against typographical errors,
I also searched for pages that had both the nonstandard spelling and the standard spelling, and found none.
110
Although I have not undertaken any statistical analysis, it is apparent from casual inspection of a
dictionary that there are far more pseudoreduplicated words than would be expected through random
phoneme combination. In addition, two occurrences of the same consonant within a root are very rare
except in pseudoreduplicated words. That is (modulo pseudoinfixation or medial a), two occurrences of the
any C within a root are allowed only if the two Cs are in the same syllabic position (onset or coda), and the
vowels of the Cs’ syllables are the same; if the Cs are codas, the onsets of their syllables must be the same,
and if they are onsets and the first C’s syllable has a coda, the second C’s syllable’s coda must be the same
as the first C’s.

In any case, whether or not pseudoreduplicated words form a definable, psychologically real, or
historically motivated class is of no consequence to the proposal here. The important characteristic of the
words I am calling pseudoreduplicated is only that they display a high degree of internal similarity.

140
 pseudoinfixed (-al-, -aR-, -ag-, or -a?-)
balusbo@s ‘spilling of grain from hole in container’
tagajta@j ‘mountain crest’
da?igdi@g ‘world’

Whatever the historical origin of these words, there are several reasons to call them
pseudoreduplicated synchronically. First, in Tagalog the minimal root is disyllabic, so if,
for example, saksa@k were reduplicated, it would be from a too-small root, (sak).
Pseudoreduplication might be a repair strategy for just such too-small roots, but there are
multiple pseudoreduplicating patterns, so letting just one pattern (say CVC-
reduplication) be the repair strategy would explain only a portion of the
pseudoreduplicated vocabulary. The rest would still have to be listed as-is in the lexicon.
Second, although Tagalog does have productive CV- reduplication, there is no productive
CVC- reduplication, nor are the pseudoprefixes and pseudoinfixes productive. And
finally, although many pseudoreduplicated roots have a mimetic flavor, there is no fixed
meaning associated with any of the pseudoreduplicating patterns—it would be strange to
posit a reduplicative morpheme when there does not seem to be any morphosyntactic
information associated with it.
Usually, the two halves of a pseudoreduplicated word behave independently. That
is, phonological phenomena apply transparently, even if the result is nonidentity between
the two halves. But over- and underapplication do occur sporadically, as if some
pseudoreduplicated words were being treated as productively reduplicated. I will discuss
five types of example, summarized in (101).

141
(101) Over- and underapplication in pseudoreduplicated roots
nasal substitution
productive reduplication most pseudoredup. handful of pseudoredup.
overapplies transparent overapplies
ku@ùlot kamka@m budbo@d
maù-Nu-Nulo@t ma-Namka@m ma-mudmo@d
‘lock of hair’, ‘hairdresser’ ‘usurpation’, ‘to usurp’ ‘sprinkling’, ‘to sprinkle’

intervocalic flapping
productive reduplication most pseudoredup. handful of pseudoredup.
transparent transparent overapplies
mag-da-Rasa@l di@ùRi Ru@@ùRok
‘will pray’ ‘loathing’ ‘acme’
underapplies
de@@ùde
‘baby bottle’

vowel raising
most productive redup. most pseudoredup. several pseudoredup.
transparent transparent underapplies
?aùbut-?a@ùbot dubdo@b goNgo@N
‘continuous’ ‘feeding a fire’ ‘gruntfish’

nasal assimilation
productive reduplication many pseudoredup. many pseudoredup.
underapplies transparent underapplies
mag-dunuùN-dunu@ùN-an danda@N diNdi@N
‘to engage in pedantry’ ‘toasting’ ‘wall’

glottal deletion
productive reduplication many pseudoredup. many pseudoredup.
underapplies transparent underapplies
?alat-?alat-a@n ?utot ?ig?ig
‘to make a little salty’ ‘flatulence’ ‘shaking’

First, recall from §2.2.1 that when nasal substitution applies to a productively
reduplicated word, it applies to both the base and the reduplicant, even though only the
reduplicant is adjacent to the triggering prefix: kulo@t ‘lock of hair’, ma@-Nu-Nulo@t
‘hairdresser’. In Wilbur’s (1973) and McCarthy and Prince’s (1995) terms, nasal

142
substitution overapplies.111 In most pseudoreduplicated words, only the first half
undergoes nasal substitution: kamka@m ‘usurpation’, ma-Namka@m ‘to usurp’. But I have
found one pseudoreduplicated root in which nasal substitution overapplies to the second
half in some derivatives, one in which it overapplies with an unproductive zero-prefix,
and one in which it overapplies with the unproductive prefix hiN-:

(102) Overapplication of nasal substitution

budbo@d ‘sprinkling’
?ipa-mudmo@d ‘to distribute to many individuals’
ma-mudmo@d ‘to scatter’
pa-mu-mudmo@d ~ pa-mu-mudbo@d ‘distribution of small quantities’
pam-budbo@d ‘used for sprinkling’
ba@ùbad ‘soak’
mama@d ‘softened by soaking’
hi-mulmo@l ‘plucking fine hairs’
bulbo@l ‘fine hair, feather’

Second is flapping. In the bulk of the native vocabulary, [d] and [R] are in
complementary distribution: [R] occurs intervocalically, [d] elsewhere, except that
sometimes root-initial [d] is retained despite prefixation with a vowel-final prefix.
Productive reduplication triggers flapping; that is, the constraints driving flapping are
obeyed despite the resulting nonidentity between base and reduplicant: mag-dasa@l ‘to
pray’, mag-da-Rasa@l ‘will pray’. Likewise, in most pseudoreduplicated words, flapping
applies transparently, both within roots and across morpheme boundaries: di@ùRi

111
Transparent application: a “rule” applies in all and only the expected environments, even though a
misidentity between base and reduplicant may result.

Overapplication: either the base or the reduplicant (but not both) is in the expected environment for a rule,
and the rule applies to both.

Underapplication: either the base or the reduplicant (but not both) is in the expected environment for a rule,
and the rule applies to neither.

143
‘loathing’; kadka@d ‘unfolded’, kadkaR-i@n ‘to unfurl; damda@m ‘feeling’, ma-Ramda@m-in
‘emotional’. But, there is one pseudoreduplicated word in which flapping underapplies,
de@ùde ‘baby bottle’, and two in which it overapplies, Rima@ùRim or dima@ùRim ‘nausea’,

Ru@ùRok ‘acme’. These words display stronger base-reduplicant identity than productively
reduplicated words.
Third is vowel raising. Exceptional nonfinal mid vowels are usually preserved
under productive reduplication: se@ùlos ‘jealousy’ pag-se-se@ùlos-an ‘jealousy of each
other’. Raising usually occurs in disyllabic productive reduplication, despite the resulting
misidentity: ?a@ùbot ‘overtaken’, ?aùbut-?a@ùbot ‘continuous’. That raising is often optional
in disyllabic reduplication (ha@ùlo ‘mix’, haùlu-ha@ùlo or haùlo-ha@ùlo ‘(frozen desert/drink)’)
may reflect a prosodic break comparable to the break within a compound rather than the
effect of reduplicative identity: the reduplicant and the base are each long enough to be a
prosodic word, and each has stress/length (if the reduplicant has a long penult, it bears
secondary stress; otherwise the reduplicant’s ultima bears secondary stress, even if
closed).
In pseudoreduplicated words, vowels usually diverge in height in order to obey
markedness constraints (dubdo@b, ‘feeding a fire’), but in a few words, both vowels are
mid, as in ko@ùkok ‘crow of rooster; chickie’, and goNgo@N ‘(gruntfish species)’ We could
say that in these words, *NONFINALMID underapplies. I have found no examples in
which *FINAL[u] overapplies (i.e., no words like *budbu@d).
Fourth is nasal assimilation. In Tagalog, a nasal usually agrees in place of
articulation with a following obstruent. This is true both root-internally and across clitic
boundaries. When productive disyllabic reduplication places a root-final nasal next to a
heterorganic root-initial stop, nasal assimilation underapplies: du@ùnoN ‘erudition’, mag-

144
dunuùN-dunu@ùN-an ‘to engage in pedantry’.112 In pseudoreduplicated words, nasal
assimilation often applies transparently, but often underapplies:113 danda@N ‘warming over
fire’ vs. diNdi@N ‘wall’.
Finally, glottal deletion: in Tagalog, a postconsonantal glottal stop is often
deleted.114 For example, when a verb ending in [?] syncopates, the glottal stop is deleted:
g-um-awa@? ‘to do (ActorFocus)’, gaw-i@n ‘to do (ObjectFocus)’ (instead of *gaw?-i@n).
Glottal stop is preserved, at least in careful speech, with most prefixation (?aba@N
‘watcher’, mag-?aba@N ‘to watch for’), and with productive reduplication: ?ala@t ‘salt’,
?alat-?alat-a@n ‘to make a little salty’.115 Root-internally, C? clusters are rare, and many
pseudoreduplicated words lack an expected glottal stop: ?utot ‘flatulence’. But, in about
half of relevant pseudoreduplicated words, glottal deletion underapplies: ?ig?ig
‘shaking’.
Thus, there is evidence that words that appear—phonologically—to be
reduplicated are sometimes treated as reduplicated, even in the absence of
morphosyntactic cues.

112
As with vowel raising, the failure of nasal assimilation to apply in the first nasal-obstruent cluster of
dunuùN-dunu@ùN-an may reflect a prosodic break between base and reduplicant rather than reduplicative
identity. The boundary between reduplicant and base would have to be sharper, though, than a clitic
boundary, where nasal assimilation is usual.
113
Although I have not performed a complete count, it appears that nasal assimilation underapplies at least
a third of the time.
114
Preconsonantal glottal stop seems always to be deleted/absent: at clitic boundaries, in productive
reduplication, and in pseudoreduplicated words.
115
Again, the lack of glottal deletion could reflect the strength of a boundary between base and reduplicant,
although glottal deletion is common at clitic boundaries. See fn. 112.

145
4.4.1. Analysis

I call the constraint driving morphosyntactically unmotivated reduplicative construals

REDUP (short for “Reduplicate”), and it penalizes every pair of syllables not in
correspondence with each other (to be more exact, REDUP penalizes a pair of syllables
when no correspondence relation is defined between the segments of those syllables). I
use “correspondence” in the sense of McCarthy and Prince (1995): an arbitrary relation
between segments that does not in itself require similarity; violable constraints require the
relation to have certain properties, and enforce similarity between segments that are in
correspondence. Matching Greek subscripts on syllables indicate that the representation
includes a correspondence relation between the segments of those syllables.

(103) REDUP
* Word where α ≠β
σα ... σβ

Two syllables within the same word must be in correspondence with each other.

For example, [ba]α[ba]α does not violate REDUP, because it has just one syllable pair, and
that pair is in correspondence; [ba]α[da]β violates REDUP once, because its one syllable
pair is not in correspondence. [ba]α[ba]α[da]β violates REDUP twice, because the syllable
da does not correspond to either of the ba syllables. Assuming that Correspondence is
transitive, we can also have words like [ba]α[ba]α[ba]α[ba]α, in which every syllable is in
correspondence with every other syllable (no violations of REDUP). The tableau in (104)
shows how a word with three syllables can violate REDUP three times, twice, or not at all.
Note that the quality of the correspondence relation is a separate matter—REDUP is
satisfied by the mere existence of correspondence between the two syllables, regardless
of how similar they are.

146
(104) Violations of REDUP for a 3-syllable input
/badaka/ REDUP
a [ba]α[da]β[ka]γ *(ba-da) *(ba-ka) *(da-ka)
b [ba]α[da]β[ka]β *(ba-da) *(ba-ka)
c [ba]α[da]β[ka]α *(ba-da) *(da-ka)
d [ba]α[da]α[ka]β *(ba-ka) *(da-ka)
e [ba]α[da]α[ka]α

The formulation of REDUP used here is somewhat arbitrary. Many of the English
examples in (99) seem to involve correspondence between feet rather than syllables (e.g.
[orang]α[utang]α, which could also be correspondence between nonadjacent syllables:
o[rang]αu[tang]α), and productive reduplication (in Tagalog as in other languages) can
involve foot-copying. Productive reduplication can also place into correspondence strings
that do not have the same prosodic shape, as in Ilokano pjan.-pja.no ‘pianos’ (also pii.-
pja.no, pi-p.ja.no; Hayes & Abad 1989): the reduplicant’s n is a coda, but the base’s is an
onset. If REDUP promotes the same correspondence structures that are found in
productive reduplication, it should be able to maximize correspondence over segments,
then, as well as over syllables and feet. For the case of Tagalog vowel height, however,
the definition in (103) is suitable.
Because REDUP promotes correspondence relations, the constraints governing
those relations proposed in McCarthy and Prince (1995) are also relevant. McCarthy and
Prince propose constraints that enforce similarity between input and output (IDENT-IO[F],
MAX-IO, DEP-IO, etc.—I abbreviate the set as CORR-IO) and between corresponding
syllables in the output (IDENT-BR[F], MAX-BR, DEP-BR, etc.—I abbreviate the set as
CORR-BR).116 IDENT-AB[F] constraints require that a segment in representation A and its

116
Because the examples here are from Tagalog, which has left-side reduplication, and because all the
examples considered here involve correspondence between just two syllables, I will refer to the first as the
reduplicant and the second as the base.

147
correspondent in representation B bear identical values of the feature F; MAX-AB
constraints require that every segment in A have a correspondent in B; and DEP-AB
constraints require that every segment in B have a correspondent in A.
The correspondence constraints interact with REDUP to (i) restrict which syllables
can be in correspondence and (ii) enhance the similarity of corresponding syllables. The
schematic factorial typology in (105) illustrates the interaction.

(105) Factorial typology of REDUP, CORR-IO, and CORR-BR

REDUP, CORR-BR >> CORR-IO

underlyingly dissimilar syllables correspond and are made identical
/bakpak/ REDUP CORR-BR CORR-IO
a . [bak]α[ bak]α *
b [bak]α[ pak]α *!
c [bak]α[pak]β *!
d [bak]α[bak]β *! *

REDUP , CORR-IO >> CORR-BR

underlyingly dissimilar syllables correspond but remain dissimilar
/bapa/ REDUP CORR-IO CORR-BR
a [bak]α[ bak]α *!
b . [bak]α[ pak]α *
c [bak]α[pak]β *!
d [bak]α[bak]β *! *

CORR-BR, CORR-IO >> REDUP

underlyingly dissimilar syllables cannot correspond
/bapa/ CORR-BR CORR-IO REDUP
a [bak]α[ bak]α *!
b [bak]α[ pak]α *!
c . [bak]α[pak]β *
d [bak]α[bak]β *! *

Because there are many CORR-BR and CORR-IO constraints, a language may
belong to different classes in this typology for different correspondence constraints—for
example, allowing a voiced and voiceless segment to correspond in an output, but
requiring correspondents to agree in sonority. The typology also becomes more

148
complicated when markedness constraints are included, as seen below. In particular, the
interplay of REDUP, correspondence constraints, and markedness constraints will show
that there is a difference between phonetically identical candidates like [ba]α[pa]a
(construed as reduplicated), the winner in the second tableau of (105), and [ba]α[pa]a (not
construed as reduplicated), the winner in the third tableau; the presence of internal
correspondence can be detected even when internal similarity is not enhanced.
There arises the question of why, if there is such a constraint as REDUP, there are
no languages in which all words are reduplicated. Such a language would be quite
inefficient—every word’s uniqueness point would be at the halfway mark, and the second
half of the word would serve no contrastive function. I cannot explain the mechanism that
prevents pathological grammars from arising, but it is clear that such a mechanism exists,
because it also prevents many other contrast-reducing constraints from rising to the top of
the grammar. For example, the silent language, in which *STRUC (Zoll 1993) dominates
all faithfulness constraints, does not exist. Similarly, Prince and Smolensky 1993 propose
constraints of the form *P/X that forbid X as a syllable nucleus; the less sonorous X is,
the more marked it is a nucleus: *P/[t]>> *P/[n] >> *P/[u] >> *P/[a]. But there is no
language in which all the *P/X except *P/[a] are undominated, requiring all syllable
nuclei to be [a].
Other authors have proposed constraints that encourage word-internal similarity.
MacEachern (1999) proposes a constraint BEIDENTICAL, which requires all segments of a
word to be identical; violations occur when two segments differ in a feature F and IDENT-
IO[F] outranks BEIDENTICAL. BEIDENTICAL differs from REDUP in that it is satisfied only
by full identity; BEIDENTICAL does not cause partial similarity enhancement or
preservation. Suzuki (1999) proposes a constraint that requires onsets of adjacent

149
syllables to be identical. Suzuki’s proposal differs from MacEachern’s in predicting that
being in the same syllable position is a prerequisite to becoming identical.
Walker (2000, to appear) proposes a family of constraints that require consonants
to enter into correspondence if they already share certain feature values. This constraint
family is similar to REDUP in that perfect identity is not required—only a correspondence
relation is required, and it is left to other constraints to enforce similarity (partial or total)
between the corresponding consonants. Walker’s proposal, which I will refer to as
Consonantal Correspondence does not predict that anything other than the consonants’
features (e.g., the consonants’ position in the syllable, the shape of the consonants’
syllables, vowels tautosyllabic to the syllables) should encourage correspondence.
Aggressive Reduplication and Consonantal Correspondence make largely
overlapping empirical predictions about consonantal similarity itself, with one
exception.117 Only Consonantal Correspondence can produce a system in which all
consonants that are similar to at least some degree become identical, and less-similar
consonants do not assimilate at all. For example, if {IDENT-BR[PLACE], IDENT-
BR[VOICE], CORRESPONDIFIDENTICALIN[PLACE],118 CORRESPONDIFIDENTICALIN[VOICE]}
>> {IDENT-IO[PLACE], IDENT-IO[VOICE]}>> CORRESPONDIFIDENTICALIN[SYLLABIC],
then /daba/ → [[da]α[da]α] , and /data/ → [[da]α[da]α] , but /dapa/ → [dapa] . In
Aggressive Reduplication, by contrast, if REDUP and the IDENT-BR[F] constraints are
ranked high enough to force the violations of IDENT-IO[PLACE] and IDENT-IO[VOICE] in
/daba/ → [[da]α[da]α] and /data/ → [[da]α[da]α] , then they must also require /dapa/ →
[[da]α[da]α] .

117
Factorial typologies for the two approaches were calculated using Hayes (1999).
118
This is not Walker’s notation.

150
 Aggressive Reduplication was discussed here because it will be employed to
explain the distribution of exceptions to vowel raising among loanwords. The following
section describes the loanword data and shows how Aggressive Reduplication could
account for them.

151
 4.5. Distribution of exceptions in the loanword vocabulary

As in the native vocabulary, there are exceptions of all kinds to vowel height
phonotactics in Tagalog loanwords. Exceptions are more numerous among the
loanwords, which come from languages that freely allow nonfinal mid vowels and final
[u]:

(106) Loanword stems with nonfinal mid vowels and final [u]
be@ùnta ‘sales’ (from Spanish venta)
kore@k ‘correct’ (from English correct)
?asu@l ‘blue’ (from Spanish azul)
?aùbaku@s ‘abacus’ (from English abacus)

Some mid-final loanword stems alternate, and some fail to alternate:

(107) Alternation in loanword stems

Alternating stems
sabo@n ‘soap’ sabun-a@n ‘to put soap on’
ata@ùke ‘attack’ ataki@ù-hin ‘to attack (object focus)’
go@@ùlpe ‘hit’ gulpi-hi@n ‘to hit (OF)’119
Nonalternating stems
ka@ùble ‘cable (message)’ kable-ha@n ‘to send a cable to’
mag-mane@ùho ‘to drive (AF)’ maneho@ù-hin ‘to drive (OF)’

Because vowel height within a bare stem is usually120 borrowed faithfully from
Spanish, it is of little interest—in other words, a nonfinal mid vowel is present because it

119
Occasionally a nonfinal mid vowel such as the o in go@@ùlpe in becomes high under suffixation. I know of
no cases in which this happens without the final mid vowel also raising. That fact lends to support to the
Aggressive Reduplication analysis of exceptions to vowel raising proposed here: although in most of the
examples seen here, the stem-final vowel resists raising in order to remain similar to the stem-penult vowel,
in go@@ùlpe the reverse happens—the stem-penult vowel and stem-final vowel remain similar by both raising.
“Double raising” cases like go@@ùlpe are not included in the statistical analysis because there are not enough
of them. but the prediction of Aggressive Reduplication would be that double raising, like nonraising, is
more likely when the stem ultima and stem penult are more similar.

152
was present in the Spanish or English word. What is of interest is whether or not a
loanword alternates when given a native suffix, since that can be determined only by the
Tagalog phonology. I constructed a database from English’s (1986) dictionary of all 488
Spanish and English loans with a mid vowel in the final syllable and one or more listed
suffixed derivatives.
As observed by Schachter and Otanes (1972), the best predictor that a loanword
stem will fail to alternate is the presence of a mid vowel in another syllable. As shown in
(108), only 6% of stems without a mid-vowel penult fail to raise (like tunel-an),121 but
32% of those with a mid-vowel penult fail to raise (like maneho-hin).122

(108) Effect of mid vowel in penult on probability of raising

13
100%
tunel-an
90% 30 11
80% maneho-hin
70%
5
60% fail to raise
50% 186 vary
40% gastus-in raise
59
30% betu-han
20%
10%
0%
mid vowel in penult no mid vowel in penult

120
though not always—still, there are not enough cases in which vowel height is nativized to investigate
what factors make such nativization probable.
121
The behavior of a stem’s derivatives is quite uniform (all raise, all vary, or all fail to raise), so, unlike in
the case of nasal substitution, it is possible to speak of stems that do or do not raise.
122
Statistical significance results are given in §4.11. All differences shown in bar charts are significant
except where otherwise noted.

153
 There are several possible explanations for why the presence of another mid
vowel discourages raising. First, perhaps the whole word is somehow marked as
contrastive for [high], since it contains one vowel with an unpredictable value of [high]
(the e in maneho). The final vowel would thus also be interpreted as contrastively (rather
than predictably) [-high], and so remain [-high] under suffixation.
A second explanation is that the presence of the nonfinal mid vowel (rare in
native words) marks the whole word as belonging to a foreign stratum, subject to a
different constraint ranking (see Itô and Mester 1995), in which Paradigm Uniformity
outranks the markedness constraints, preserving the [-high] quality of the vowel in the
bare stem even under suffixation. If this is the explanation, we would expect that other
markers of foreignness could be found that would also discourage alternation.
I examined several such predictors. Stress/length on a nonfinal closed syllable and
prepenultimate stress/length are both rare or nonexistent in the native vocabulary, but
neither one nor the other nor both was a predictor of nonalternation. I also examined
foreign distribution of [d] and [R] (in the native vocabulary, [R] is normally found
intervocalically and [d] elsewhere) as a predictor, but it had no effect on the likelihood of
alternation. Finally, I looked at overly large consonant clusters—initially, medially, or
finally—as predictors and found only a very small, weakly significant effect. Thus, a
nonfinal mid vowel’s serving as a cue to foreignness does not seem to be a good
explanation for why the presence of such a vowel discourages alternation.
A third possible mechanism by which the nonfinal mid vowel could discourage
alternation is vowel harmony. If a [-low] vowel must agree in [high] with a preceding
vowel (subject to *FINAL[u]), then the o in maneho would be prevented from raising
under suffixation:

154
(109) Vowel harmony as a mechanism for preventing alternation
/mag+lutO/ *FINAL[u] HARMONY *NONFINALMID
a .magluto *
b maglutu *!
/manehO+in/ *FINAL[u] HARMONY *NONFINALMID
c . manehohin *!
d manehuhin *!

If vowel harmony is the mechanism at work (in some probabilistic fashion),

certain factors might be expected to enhance the effect. First, agreement in backness
between target and trigger could encourage harmony (cf. Kaun 1995: agreement in height
encourages rounding harmony); and indeed, there is a strong effect, as shown in (110).

(110) Effect of matching backness between penult and ultima, given a mid penult.

100% 6
90% 3
80%
70% 24
60% fail to raise
50% vary
40% 48
2 raise
30%
20%
11
10%
0%
same mid vowel in different mid vowels
penult and ultima in penult and ultima
(e.g., todo 'all') (e.g., hero 'brand')

Second, proximity of trigger to target might increase the probability of vowel

harmony’s applying, and here again, there is a strong effect, as shown in (111).123

123
Aggressive Reduplication’s explanation for the proximity effect is that as in productive reduplication,
there are constraints (not discussed here) that prefer adjacency between reduplicant and base.

155
(111) Effect of proximity

100% 2
90% 0
30
80%
70%
5
60% fail to raise
50% 28 vary
40% raise
30% 59
20%
10%
0%
mid vowel in penult mid vowel in
(kamote 'sweet antepenult
potato') (ebakwet 'evacuate')

Third, among nonadjacent vowel pairs, the quality of the intervening vowel(s)
could have an effect—a high vowel could block harmony, by preventing the spread of [-
high]. There are, however, not enough relevant cases (stems with nonadjacent final and
nonfinal mid vowels that fail to raise) to test this prediction. Thus, vowel harmony fares
well as an explanatory mechanism. Still, I propose that Aggressive Reduplication is at
work, instead of or perhaps in addition to vowel harmony, because it makes an additional
correct prediction that vowel harmony cannot explain, as I will now demonstrate.

4.5.1. Aggressive Reduplication applied to the vowel raising

Recall that Aggressive Reduplication invokes a correspondence relationship between

syllables that are fairly similar, and can enhance or preserve similarity. In this case, a
pseudoreduplicative correspondence relationship is invoked between the two syllables
that contain mid vowels, because they are similar in both having mid vowels. If IDENT-

156
BR[HIGH] >> *NONFINALMID, raising of the second vowel under suffixation is prevented
(for greater visual clarity, lack of subscripts—instead of mismatched subscripts—is used
to indicate lack of correspondence relation, as in candidate e):

(112) Aggressive reduplication blocks vowel raising

/tonO + -an/ IDENT-IO IDENT-IO IDENT-BR REDUP *NONFINAL

[MANNER] [HI] [HI] MID
a . [to]α[no]αhan ** **
b [to]α[nu]αhan *! ** *
c [tu]α[nu]αhan124 *! **
d [to]α[to]αhan *! **
e tonu han ***! *

Candidate b in (112) fails because the vowels in the base and reduplicant fail to
match in height; c makes the vowels identical, but at the expense of changing an
underlying height specification; similarly, d makes the consonants identical at the
expense of changing various underlying manner features; and e fails because it is not
construed as reduplicated. Note that the above tableau assumes that IDENT-
BR[SONORANT] (along with other relevant IDENT-BR[F] constraints) is ranked low
enough to allow t and n to correspond.
This type of Aggressive Reduplication is a case of emergence of the unmarked
(McCarthy & Prince 1994): even if CORR-IO outranks CORR-BR, preventing
enhancement of internal similarity, REDUP can still make itself felt by setting up an

124
Candidates of this type do sometimes prevail. See fn. 119. Under the allomorph-listing approach argued
for in §4.7.2, this fact does not challenge the high ranking of CORR-IO constraints, because the listed form
being used is not the bare stem, but a separate, listed allomorph.

157
internal correspondence relation that preserves internal similarity—here by blocking
alternation.125
Agreement in backness encourages a reduplicative construal because, assuming
stochastic constraint ranking, sometimes IDENT-BR[BACK] will be ranked high enough to
prevent a reduplicative construal when the vowels do not agree in backness, as illustrated
in (113).

(113) A ranking that prevents correspondence between mismatched vowels

/donO + -an/ IDENT-IO IDENT-BR IDENT-BR REDUP *NONFINAL

[BACK] [BACK] [HI] MID
a . [do]α[no]αhan ** **
b [do]α[nu]αhan *! ** *
c donu han ***! *
/denO + -an/ IDENT-IO IDENT-BR IDENT-BR REDUP *NONFINAL
[BACK] [BACK] [HI] MID
d [de]α[no]αhan *! ** **
e [de]α[nu]αhan *! ** *
f [do]α[no]αhan *! ** **
g .denu han *** *

The cross-linguistic preference for reduplicative proximity explains the distance

effect.126 Thus, Aggressive Reduplication can also account for the predictions of vowel

125
Some casual data suggest that similar cases of similarity preservation through rule-blocking (rather than
outright enhancement) may exist in other languages: many English speakers feel that flapping of d is
almost obligatory in words like the proper name Frodo, but only optional in pseudoreduplicated dodo.
Similarly, Zulu allows either light or dark l, but pseudoreduplicated Lulu requires two light ls. Thanks to
Bruce Hayes for these observations.

In French, [] is usually found instead of [o] in nonfinal syllables (e.g., [ddy] ‘chubby’), but not
possible in baby-talk reduplicated words like [dodo] ‘beddie-bye’ (even though the source word, [d{mi{]
‘to sleep’, has []). Thanks to Roger Billerey for this observation.
126
This preference could be encoded in Alignment constraints that require, for example, the right edge of
the reduplicant to coincide with the left edge of the base.

158
harmony that were seen to be borne out above. But Aggressive Reduplication makes an
additional prediction: similarity between penult and ultima127 along any dimension—not
only vowel backness—should also encourage establishment of a reduplicative
correspondence relationship, and thus resistance to alternation. Section 4.6 shows that
this prediction is also correct, and §4.8 shows how differences in syllable similarity could
result in different probabilities of raising.
Aggressive Reduplication also predicts that in stems with a high-vowel penult,
similarity between penult and ultima should encourage raising. Unfortunately, because
nearly all non-mid-penult stems do raise, it is not possible to test this prediction.
Before moving on to §4.6, there is one problem with the rankings in (112) and
(113): how likely is the crucial ranking IDENT-BR[HIGH] >> *NONFINALMID? In
disyllabic reduplication of two-syllable stems ending in a mid vowel, the reduplicant
usually raises (?a@ùbot ‘reach; overtaken’, ?aùbut-?a@ùbot ‘one after the other’), although
this is not obligatory—nonraised pronunciations are common in many words, such as
haùlo? ‘mixture’, haùlo-ha@ùlo? ~ haùlu-ha@ùlo? ‘(drink made with shaved ice)’.128 The
prevalence of raising in disyllabic reduplication would suggest a strong tendency for
*NONFINALMID to outrank IDENT-BR[HIGH]. If this is the case, then IDENT-BR[HIGH]
should not have a noticeable tendency to prevent raising, even in words that are construed
as reduplicated. I have two possible explanations for this apparent contradiction.

127
There are not enough stems in which the correspondence relation that would block alternation would be
between the ultima and the antepenult (i.e., loanstems with three syllables or more and mid vowels in the
antepenult and ultima but not in the penult) to examine the effect of similarity between antepenult and
ultima.
128
It is unclear how lexically conditioned this optionality is. It could result from variability in the ranking
of *NONFINALMID vs. IDENT-BR[HIGH], or from variability in whether the reduplicant-base boundary
counts is strong enough to prevent raising (i.e., whether disyllabic-reduplicant-final counts as word-final).

159
 First, note that the reduplicative construals involved in blocking raising involve
single syllables ([to]α[do]α). Perhaps the IDENT-BR constraints involved in disyllabic
reduplication are different from those involved in CV reduplication. There is little
evidence for the ranking of IDENT-BR[HIGH] in CV reduplication of native words,
because native roots are at least one syllable long, and mid vowels are usually not found
in nonfinal syllables (so the syllable being copied would rarely have a mid vowel). There
are some exceptions, though (see (95)), as well as the systematic exception of the
transglottals, and in these cases, raising does not occur with CV reduplication:129

(114) Vowel non-raising in CV reduplication

te@ùkas ‘swindler’ ma-ne-ne@ùkas ‘swindler’
he@ùle ‘lullaby’ nag-he-he@ùle ‘is singing a lullaby’
lo?o@b ‘robbery’ pan-lo-lo?o@b ‘robbery’
man-lo-lo?o@b ‘burglar’
ma-no?o@d ‘to look on’ ma@-no-no?o@d ‘onlooker’

A possible interpretation is that IDENT-BR1SYLL[HIGH] >> *NONFINALMID >> IDENT-

BR2SYLL[HIGH]. This ranking would produce a strong tendency to resist raising in words
with mid penults and ultimas when they are construed as reduplicated.
A second possibility is that the lack of raising in CV reduplication reflects the fact
that the vowel being reduplicated is contrastively mid, whereas in disyllabic
reduplication, the final vowel of the base is predictably mid. Perhaps IDENT-BR[F]
constraints are sensitive to the whether F is contrastive (e.g., fully specified) in the base:
IDENT-BR[HIGH]CONSTRASTIVE >> *NONFINALMID >> IDENT-BR[HIGH]. A constraint like
IDENT-BR[HIGH]CONSTRASTIVE must have access to the reduplicant, the surface form of the

129
Similarly, raising usually does not occur (although it is often an optional variant) in CV reduplication of
loanwords with mid vowels in the initial syllable: e.g., dRo@wiN ‘drawing’; pag-do-dRo@wiN ‘act of drawing’.

160
base, and the underlying form of the base (if contrastiveness is encoded in the underlying
representation), but is otherwise no different from ordinary correspondence constraints.

161
 4.6. Similarity along other dimensions

In stems with mid vowels in both the penult and the ultima, similarity between the onset
consonants of the penult and the ultima should encourage nonraising. When both onsets
are simple (the majority case), we can simply compare the two consonants on various
features. (115) shows that when the penult and ultima onsets have the same place of
articulation, nonraising is more likely.130 The mechanism is the same as that behind the
matching-backness effect: the lo and to in pilo@ùto ‘pilot’ can correspond no matter what
the ranking of IDENT-BR(PLACE), but the bo and no of ?abo@ùno ‘fertilizer’ can
correspond only if REDUP outranks IDENT-BR(PLACE).

(115) Effect of onset place of articulation on rate of raising

100%
90% 15
80%
70% 12
60% 3 fail to raise
50% vary
40% 1
raise
30% 27
20% 8
10%
0%
same place different place
(piloto 'pilot') (abono
'fertilizer')

130
Note that the charts in this section compare all stems whose penult and ultima are similar along the
dimension under discussion to all stems whose penult and ultima are dissimilar along the dimension under
discussion. For example, in (115), the penult and ultima onsets of the words grouped with piloto must be
identical in place, but may differ in voicing or manner, and the syllables may differ in shape or vowel
quality; the penult and ultima onsets of the words grouped with abono must be different in place, but may
be different or identical along other dimensions.

162
 Identical onset manner also encourages nonraising. although the difference shown
in (116) is not significant (see §4.11):131

(116) Effect of onset manner on rate of raising

100%
90%
15
80%
12
70%
60% 3 fail to raise
50% 1 vary
40% raise
30% 24
20% 11
10%
0%
same manner different manner
(beto 'veto') (tsaperon
'chaperon')

Again, the mechanism is the same: the b and t of be@ùto ‘veto’ can correspond no
matter what the ranking of IDENT-BR[SONORANT], IDENT-BR[NASAL], or any other
IDENT-BR[MANNER] constraints. But the p and R of tSaùpeRo@n ‘chaperon’ can correspond
only if REDUP outranks various IDENT-BR[MANNER] constraints.
Voicing has no effect132 (the small difference in (117) is not significant):

131
The lack of a significant difference may be because “manner” is too crude a category. There are not
enough relevant tokens, however, to compare single-feature distinctions such as “same value for [nasal] vs.
different value for [nasal]”.
132
There were not enough stems in which both onsets were obstruents to examine obstruent voicing.

163
(117) Effect of onset voicing on rate of raising

100%
90%
80% 12 15
70%
60% fail to raise
3 1
50% vary
40% raise
30% 20
15
20%
10%
0%
same voicing different voicing
(epekto 'effect') (semento
'cement')

See §4.9 for a possible reason for the lack of a voicing effect.
When onsets match in shape (simple vs. complex), nonraising is also encouraged:

(118) Effect of onset shape on rate of raising

100% 3
90%
1
80% 27
70%
60% 4 fail to raise
50% vary
40% 22
raise
30% 37
20%
10%
0%
same shape (loko different shape
'insane') (preso 'prisoner')

Here the crucial constraints are MAX-BR and DEP-BR: correspondence between the
pRe@ù and the so of pRe@ùso ‘prisoner’ incurs a violation of DEP-BR.

164
 There are not enough cases in which both penult and ultima are closed to compare
coda consonants, but we can compare rhyme shape (open vs. closed), and again a match
promotes nonraising.

(119) Effect of rhyme shape on rate of raising

100%
90% 9
80% 21
1
70%
60% fail to raise
4
50% vary
40% 31 raise
30% 28
20%
10%
0%
same shape different shape
(doktor 'doctor') (tonto 'silly')

As for vowel length, recall that there are two basic types of stem in Tagalog: those
with a long, stressed penult and those with no long vowels and a stressed final syllable. In
stems with a long, stressed penult, length and stress shift to the right in the most common
suffixing constructions: he@ùRo ‘brand’, heRu@ù-han ‘to brand’. In stems with no long vowel,
stress shifts to the right in suffixed form: seRmo@n ‘sermon’, seRmun-a@n ‘to preach to’. We
might expect that stems of the second type would be more susceptible to a reduplicative
construal because there is no length difference between the vowels. Because final stress is
unusual in both Spanish and English, there are too few examples of final-stressed
loanstems for a significant difference, but the trend is in the predicted direction:

165
(120) Effect of vowel length on raising

100%
90% 22
80% 8
70% 5
60% fail to raise
50% 0 vary
40% raise
30% 50
9
20%
10%
0%
both short one long, one
(sermOn short (hE:ro
'sermon') 'brand')

Finally, we can look at the number of properties that the penult and ultima share
(i.e., the number of CORR-BR constraints whose ranking is irrelevant to whether a
reduplicated construal is possible, because they would not be violated), as a global
measure of similarity. With seven properties (onset place, onset manner, onset voicing,
onset shape, vowel backness, vowel length, and rhyme shape), stems can be grouped into
eight categories: those that share 0, 1, 2, 3, 4, 5, 6, or 7 of those properties (there were no
stems that shared all seven properties, so only seven categories are shown). The chart in
(121) shows that the more shared properties, the more likely a failure to raise.

166
(121) Effect of number of shared properties on raising

100%
90% 4
6
80% 1
70% 2 12 5
60% fail to raise
50% 1 12 3 vary
40% 18 raise
16 2
30%
20% 9
10% 2
0%
0 1 2 3 4 5 6
number of shared properties

To summarize: REDUP, interacting with CORR-BR, tends to discourage raising to

the extent that the final syllable is similar to a preceding syllable that also has a mid
vowel: the more similar the two syllables are, the fewer CORR-BR constraints are violated
by establishing a correspondence relation between the two. If a correspondence relation is
established, raising is less likely because it would violate IDENT-BR[HIGH].

167
 4.7. Representations

Chapter 2 assumed that all existing, potentially nasal-substituting words are listed to
some degree, whether they undergo nasal substitution or not. Listing all words provided a
three-way distinction among existing words that reliably substitute, existing words that
reliably fail to substitute, and new words, whose behavior should vary. This section
argues that for vowel raising, the three-way distinction should be achieved through a
different mechanism.

4.7.1. Separate entries for derivatives?

Separate lexical entries for all derived words (or, equivalently, separate sub-entries under
the stem’s entry) were appropriate for nasal substitution, because different derivatives of
the same stem often behave differently. In vowel raising, however, different suffixed
derivatives of the same stem nearly always133 behave the same way (all raise, or all fail to
raise). So, although occasional full listing may be a possibility in those rare stems whose
derivatives are not uniform, it is not likely the usual state of affairs.
If each stem’s raising behavior is uniform, then raising or nonraising should be
determined by some property of the stem’s own lexical entry; this property would then be
inherited by derived forms. This was the assumption in the analysis sketched above
(§4.3), which represented raising stems as having final vowels underspecified for [high],

133
I found three definite exceptions (out of 100 loanstems with multiple suffixed forms), do@ùble ‘double’,
lo@ùko ‘crazy person’, and ba@ùle ‘worth; I.O.U.’, although only for do@ùble does behavior actually differ
between suffixed stems—for lo@ùko and ba@ùle it differs between suffixed stems on the one hand and
disyllabically reduplicated stems on the other hand. There were also three possible exceptions: ro@ùljo ‘roll’,
tu@ùrno ‘lathe’, and je@ùlo ‘ice’, some of whose derivatives are pronounced variably, and some of whose
derivatives are listed as having only one pronunciation (for ro@ùljo, the difference is between suffixed stems
and reduplicant stems).

168
and nonraising stems as having final vowels specified [-high]. There is a problem,
however, with the analysis in §4.3: what do stems that have never occurred in suffixed
form yet look like? If they are underspecified, then they are identical in form to
underspecified stems that do have established suffixed forms, and should behave just like
them (always raising). But then how do nonraising stems come about? Novel suffixed
derivatives must have some freedom to raise or not raise (as determined by the stochastic
grammar). A three-way contrast is required among raising stems, nonraising stems, and
“undecideds” (stems whose suffixed form has not yet been established).

4.7.2. Environment-tagged allomorphs

A three-way contrast between raising stems, nonraising stems, and undecideds can be
achieved using environment-tagged allomorphs. Many Tagalog stems do seem to have
separate, listed allomorphs that are used in suffixal form, as demonstrated by the sporadic
phenomenon of syncope. Some stems (it is not predictable which134) undergo vowel
syncope under suffixation.135 The resulting consonant cluster can sometimes undergo
metathesis or other, unpredictable changes:

134
There is partial predictability, in that some stem shapes are always prevented from undergoing syncope:
stems with penultimate stress/length (e.g., bu@haj ‘life’) cannot syncopate, because the syncopated vowel
would be the one to which stress/length would have shifted under suffixation; and stems with a consonant
cluster between the penult and ultima (e.g., sampa@l ‘slap on the face’) cannot syncopate, because the result
would be a cluster of three consonants (*sampl-i@n)
135
At least under verbal suffixation. There are stems that syncopate in some constructions, but not others:
e.g., dati@N ‘arrival’, ka@-hi-natn-a@n ‘to be the outcome’, ka@-Ratn-a@n ‘possible result, ka-Ra-Ratn-a@n
‘expected time of menstrual period’, datn-a@n/datn-i@n ‘to arrive at’, but pa-RatiN-a@n/pa-RatiN-i@n ‘to have
(someone or something) sent; to have someone bribed’. It is possible that those constructions that shun the
syncopated allomorph have special prosodic requirements, or are separately listed, or that high-ranking
Paradigm Uniformity constraints enforce similarity to a related, unsuffixed form (e.g., pa-Rati@N ‘message
sent; bribe’).

169
(122) Syncope
syncope alone (many examples)
mag-biga@j ‘to give (AF)’ bigj-a@n ‘to give (IOF)’
mag-taki@p ‘to cover (AF)’ takp-a@n ‘to cover (LF)’
b-um-ili@ ‘to buy (AF)’ bilh-i@n ‘to buy (OF)’136
syncope plus consonant changes (few examples)
d-um-ati@N ‘to arrive at (AF)’ datn-a@n ‘to arrive at (LF)’
t-um-iNi@n ‘to look at (AF)’ tiNn-a@n ~ tign-a@n ‘to look at (LF)’
mag-tani@m ‘to plant (AF)’ tamn-a@n ‘to plant (LF)’
h-um-ali@k ‘to kiss (AF)’ halk-a@n ~ hagk-a@n ‘to kiss (OF)’

Just as a stem that undergoes syncope would have a syncopated nonfinal

allomorph137 in its lexical entry, so would a stem that fails to raise have a nonraised
allomorph, and a stem that raises would have a raised allomorph:138

(123) Suffixal allomorphs—sample partial lexical entries

‘give’ ‘basket’ ‘adobo’
/biga@j/_# /ba@ùsket/_# /?ado@ùbo/_#
/bigj/_X /basket/_X /?adobu@ù/_X

For stems like these that have an existing suffixal allomorph, high-ranking IDENT-
IO[HIGH] requires that the underlying height of the final vowel be faithfully parsed. We
need, in addition, a constraint that requires allomorphs to be context-appropriate:

136
The use of listed allomorphs helps explain why vowel-final stems that syncopate have a final [h], even
though the [h] is not needed to resolve hiatus. Listed suffixal allomorphs can also encode the exceptionality
of stems like kita@ ‘visible’, which has final [?] instead of [h] in suffixed form (paù-kita@ù?-an ‘showing
(something) to one another’).
137
Constraints against large consonant clusters would have to outrank MATCHCONTEXT (see below), to
prevent use of the syncopated allomorph in disyllabic reduplication (*bigj-bigaj—in this case, the
disyllabic requirement also is not met).
138
The lexical entries for the allomorphs need not be simple phoneme strings as shown here. They could
employ diacritics, cross-references to context-insensitive allomorphs, or some other device.

170
(124) MATCHCONTEXT

The context requirements [e.g., “__#”] of a morpheme in the input must not
contradict the context in which that morpheme’s output-correspondent segments
occur in the output.

For example, the candidate /b1i2g3a@4j5/_# + /-i6n7/ → [b1i2g3a@4j5i6n7] violates

MATCHCONTEXT, because the first morpheme in the input requires a nonfinal context, but
the output correspondent the last segment (j5) is not word-final. The tableau in (125)
illustrates faithful use of a suffixal allomorph.

(125) Faithful use of suffixal allomorphs

‘to make into adobo’ IDENT-IO MATCH *NONFINAL REDUP PU
[HIGH] CONTEXT MID
a. /?adobu/_X + /-in/ * ****** *
→ [?adobuhin]
b /?adobu/_X + /-in/ *! * ****** *
→ [?adobohin]
c /?adobo/_# + /-in/ *! ** ******
→ [?adobohin]
d /?adobo/_# + /-in/ *! * ** ******
→ [?adobuhin]
e /?adobo/_# + /-in/ *! * ** *****
→ [?a[do]α[bo]αhin]

When a stem has no listed suffixal allomorph, however, MATCHCONTEXT cannot

be satisfied. It cannot be the case that the speaker uses the word-final allomorph instead,
because then high-ranking IDENT-IO[HIGH] would always prevent raising, and listeners
would always add an unraised allomorph to the lexicons (i.e., no loanwords or other new
words would ever raise).
There is evidence in other languages that inflected words whose properties are
fully predictable from those of their stems (“regulars”) are usually not separately listed
(see §5.3.1)—and yet a distinction between existing regulars (not listed but always
regular) and novel words (not listed, behavior varies) is preserved. This suggests that

171
speakers must be able to reason about whether a listed form “should” exist or not (i.e.,
whether other speakers have a listed form): if a speaker’s lexical entry for a stem is strong
(i.e., she has heard it many times), and she has no lexical entry for the inflected form,
then probably none “should” exist, and the inflected form should be produced
synthetically, by inputting the stem and affixes to the grammar. But if the lexical entry
for the stem is weak, as in novel words, it is probable that a listed form exists for other
members of the speech community, and the speaker has simply never encountered it; in
that case, the speaker may feel free to construct potential listed forms.
This dissertation will not attempt to construct a model of how speakers decide
whether a listed form exists, or of how the speaker constructs possible listed forms for
novel words. This question is related to another that will not be modeled here: how
speakers reason from the amount of variation among derivatives of the same stem that for
nasal substitution, whole words must be listed, but for vowel raising, only context-tagged
allomorphs must be listed.139
Assuming that speakers can construct possible suffixal allomorphs for a novel
word, multiple candidates would satisfy the two highest-ranked constraints, and the
ranking of *NONFINALMID, CORR-IO, PU, and REDUP determines the winner:

139
This reasoning or some equivalent must take place to perpetuate the uniformity in behavior with respect
to vowel raising (and not impose uniformity in behavior with respect to nasal substitution). There may also
be a fundamental distinction between suffixal and non-suffixal environments in Tagalog: there are only two
suffixes (-in and -an), which play a variety of morphosyntactic roles. Suffixes condition stress and length
shifts, as well as syncope. By contrast, there are many prefixes; a single word may contain several prefixes;
and the only alternation triggered by prefixes is nasal substitution.

172
(126) Variability for constructed suffixal allomorphs
‘to gete’ (novel word) IDENT-IO MATCH *NON IDENT-BR REDUP PU
[HIGH] CONTEXT FINALMID [PLACE]
a /gete/_X + /-in/ ** ***
→ [getehin]
a /gete/_X + /-in/ ** * **
→ [[ge]α[te]αhin]
b /gete/_X + /-in/ *! * *** *
→ [getihin]
c /geti/_X + /-in/ *! ** ***
→ [getehin]
d /geti/_X + /-in/ * *** *
→ [getihin]
e /gete/_# + /-in/ *! ** ***
→ [getehin]

Given a mechanism by which speakers decide whether to construct a suffixal

allomorph, is environment-tagging really necessary? Without environment-tagging, stems
that raise would have two listed allomorphs (one unraised and one raised; markedness
constraints would select the best allomorph in each context), and stems that fail to raise
would have just one (unraised). The difference between stems that consistently fail to
raise, and novel stems (which would also have just one allomorph, and should behave
variably) would be that for familiar stems, the speaker knows not to entertain the
possibility that there exists a raised allomorph that she has simply never heard. The
reasoning procedure for determining whether or not to construct a raised allomorph
would have to involve the constraints in the grammar, so that a word’s phonological
properties (e.g., internal similarity) would contribute to the probability of constructing a
raised allomorph. Otherwise, since existence of a raised allomorph must always entail
raising, all novel words would have the same probability of raising. The remainder of this
chapter will continue to assume environment-tagged allomorphs, but with a theory of the

173
construction of lexical entries for inflected forms of novel words, this might not be
necessary.
Note that the phenomenon of syncope does not settle the question of whether or
not allomorphs are tagged for context. In most cases, markedness constraints could select
the correct allomorph (syncopated or not) for each context (suffixal or non-suffixal). For
example, the bigj allomorph of ‘give’ in (122) would be unsuitable word-finally, because
of its final consonant cluster; but when suffixation allows the gj cluster to straddle a
syllable boundary, *STRUC (a constraint against phonological material in the output—
Zoll 1993) would disprefer the bigaj allomorph. There is one type of case that might
support the idea of environment-tagging: when a stem ending in [?] syncopates, the
glottal stop is deleted (g-um-awa@? ‘to make, to do (actor focus)’, gaw-i@n ‘to make, to do
(object focus)’). A two-syllable minimal-word constraint (only clitics and some loans are
monosyllabic) could rule out gaw in unaffixed context, but gaw is not used even when a
prefix is present (e.g., mag-gawa@? ‘to manufacture’). It is possible that the minimal-word
constraint applies to post-prefix material, so a conclusive test would be a trisyllabic stem
ending in glottal stop that syncopates, but I have found none.
A final point concerns the possible difference between loanwords and native
words. The uniformity of raising under suffixation among native words (except
pseudoreduplicateds and transglottals) contrasts with the variability seen among
loanwords, even those with no mid vowel in the penult (in which cases the only
motivation for nonraising would be PU). It seems plausible that there is an additional
force against raising in loanwords: if bilinguals are the primary creators of suffixed forms
of loanstems, then trans-language correspondence constraints would tend to disprefer
raising. This dissertation will not develop a theory of trans-language correspondence
constraints, but their existence seems probable.

174
 4.8. Modeling raising

Aggressive Reduplication’s influence on the distribution of exceptions to vowel raising

can be explained in the model proposed above for nasal substitution: listed forms
generally prevail, but low-ranked constraints shape the lexical entries of new words in a
probabilistic fashion.
To summarize the model proposed for nasal substitution, as it would apply to
vowel-height alternations: when a stem undergoes suffixation for the first time, Paradigm
Uniformity constraints prefer nonraising (preserving identity to the final vowel of the
unsuffixed form). *NONFINALMID, however, prefers raising; if *NONFINALMID outranks
PU, the speaker raises the stem-final vowel, and the listener updates her lexicon
accordingly. REDUP and the CORR-BR constraints also influence the outcome, by
discouraging raising in stems that have a high degree of internal similarity and are
thereby susceptible to a reduplicative construal.
Sample tableaux in (127)-(129) illustrate how internal similarity affects the
chances of raising in a novel word. The tableau in (127) shows that in the case of
perfectly identical syllables, if REDUP >> *NONFINALMID, raising does not occur, even if
*NONFINALMID >> PU. Candidate d fails because it does not have reduplicated structure
(no subscripts). Candidate b fails because its two corresponding syllables are not identical
in height.

175
(127) Vowel height in a novel word: identical syllables
suffixed form of saklolo IDENT-IO IDENT-BR REDUP *NONFINAL PU
[HI] [HI] MID [HI]
a . /saklolo/_X + /-an/ ** **
→ saklolohan
b /saklolu/_X + /-an/ *! ** * *
→ sak[lo]α[lu]αhan
c /saklolu/_X + /-an/ ***! * *
→ sakloluhan

The tableau in (128) shows that in syllables that are fairly similar (in this case,
identical in place and manner, but differing in voice), if REDUP outranks *NONFINALMID
and the relevant CORR-BR constraints (here, IDENT-BR[VOICE]), raising is blocked
despite imperfect identity: candidate d wins despite its violation of IDENT-BR[VOICE].
Note that candidate g, in which the two syllables’ onsets are made identical, fails as long
as IDENT-IO[VOICE] >> REDUP.

(128) Vowel height in a novel word: similar syllables

suffixed form of todo ID-IO ID-BR ID-BR ID-IO ID-IO REDUP *NON PU ID-BR
[HI] [HI] [PL] [PL] [VOICE] FNL [HI] [VOICE]
MID
d . /todo/_X + /-an/ ** ** *
→ [to]α[do]αhan
e /todu/_X + /-an/ *! ** * * *
→ [to]α[du]αhan
f /todu/_X + /-an/ ***! * *
→toduhan
g /todo/_X + /-an/ *! ** ** *
→ [to]α[to]αhan

The tableau in (129) shows why when the syllables are less alike (in this case,
differing in place and manner), it is less likely that they will be construed as reduplicated:
there are more CORR-BR constraints that would have to be outranked by REDUP. In the
example shown here, the same ranking produces a reduplicative construal for todo, but a
nonreduplicative construal for ?estorbo: candidates h, and i violate IDENT-BR[PLACE] (as

176
well as DEP-BR, not shown); candidate l corrects the place misidentity, but violates
IDENT-IO[PLACE]. Since a reduplicative construal is impossible, *NONFINALMID chooses
the best nonreduplicated candidate, j.

(129) Vowel height in a novel word: dissimilar syllables

suffixed form of ID-IO ID-BR ID-BR ID-IO ID-IO REDUP *NON PU ID-BR
?estorbo [HI] [HI] [PL] [PL] [VCE] FNL [HI] [VCE]
MID
h /?estorbo/_X + /-an/ *! ***** ** *
→ ?es[tor]α[bo]αhan
i /?estorbu/_X + /-an/ *! * ***** * * *
→ ?es[tor]α[bu]αhan
j . /?estorbu/_X + /-an/ ****** * *
→ ?estorbuhan
k /?estorbo/_X + /-an/ ****** **!
→ ?estorbohan
l /?estorbo/_X + /-an/ *! ***** ** *
→ ?es[tor]α[do]αhan

These tableau only illustrate possible rankings that might occur on a given
occasion. Because IDENT-BR[PLACE] >> REDUP >> IDENT-BR[VOICE], consonants that
differed in place could not correspond, but consonants that differed in voice could. On
another occasion, a ranking might be generated that would prevent consonants that differ
in voice from corresponding (IDENT-BR[VOICE] >> REDUP), or would allow consonants
that differ in place to correspond (REDUP >> IDENT-BR[PLACE]). Similarly, whether or
not consonants that differ in manner can correspond depends on the relative ranking of
REDUP and IDENT-BR[MANNER] (a shorthand, like IDENT-BR[PLACE], for several IDENT-
BR[F] constraints); whether syllables that differ in onset or rhyme shape can correspond
depends on the ranking of REDUP versus MAX-BR or DEP-BR; whether vowels that differ
in backness can correspond depends on the ranking of REDUP versus IDENT-BR[BACK].
The key point is that there are many constraint rankings under which a word with
similar mid-vowel syllables will be construed as reduplicated, but fewer rankings under

177
which a word with less similar mid-vowel syllables will be construed as reduplicated.
todo will be construed as reduplicated—and so fail to alternate—whether IDENT-
BR[PLACE]>>REDUP or REDUP>>IDENT-BR[PLACE]. ?estorbo can be construed as
reduplicated only if REDUP >> IDENT-BR[PLACE]. Under stochastic constraint ranking, it
is thus more likely that a word like todo will fail to alternate.

178
 4.9. Learnability

In Chapter 2, it was argued that the rankings of the constraints involved in nasal
substitution were learnable from exposure to existing potentially nasal-substituted words;
this was possible because the patterns within nasal substitution (the voicing and place
effects) were found throughout the set of nasal-substituted words. Vowel height, by
contrast, is close to exceptionless within the native vocabulary (at least under
suffixation—see §4.4.1’s discussion of disyllabic reduplication), so very little
information about the relative ranking of, for example, REDUP and *NONFINALMID could
have been learned before the influx of the Spanish and English loanstems whose behavior
these constraints shaped.
Some information about the rankings of CORR-BR constraints can, however, be
learned from the reduplicative identity effects seen in productive reduplication (see
examples in §4.4). For example, the overapplication of nasal substitution
(/maN/+/REDCV/+/kulo@t/ → [ma@-Nu-Nulo@t] ‘hairdresser’) tells the learner that IDENT-
BR[NASAL] >> IDENT-IO[NASAL].140 The underapplication of nasal assimilation and

140
If the ranking IDENT-BR[NASAL], REDUP >> IDENT-IO[NASAL] is expected to occur occasionally,
nothing prevents inputs like /tanak/ from surfacing as [nanak] (the issue arises only for coronals; roots of
the form /pVm.../, /bVm.../, /kVN.../ and /gVN.../ are not attested). It must be assumed, then, as for the other
CORR-IO constraints, that REDUP is ranked low that it virtually never outranks IDENT-IO[NASAL]. By
transitivity, this means that REDUP can also never outrank IDENT-BR[NASAL]—that is, mid-penult stems
whose penult and ultima onsets differ in nasality should be no more likely to resist raising that are low-
penult or high-penult stems.

This is not the case, however: mid-penult stems whose penult and ultima onsets differ in nasality
have a 29.4% chance of resisting raising (compared to 44.9% for mid-penult stems whose penult and ultima
onsets match in nasality), whereas low-penult and high-penult stems have only a 6% chance of not raising.
Perhaps nasal substitution does not exhibit true overapplication. See Inkelas 2000 for an argument that
apparent overapplication in Tagalog nasal substitution really reflects Output-Output Correspondence.

179
glottal deletion141 (/mag/+/RED2syll/+/du@ùnoN/ → [mag-dunuùN-dunu@ùN-an] ‘to engage in
pedantry’; /RED2syll/+/?ala@t/ → [?alat-?alat-a@n] ‘to make a little salty’) means that IDENT-
BR[PLACE] >> IDENT-IO[PLACE] and DEP-BR142 >> *C? (or whatever the constraint is
that forbids postconsonantal glottal stop). There are no cases of reduplicative identity that
suggest a high ranking for IDENT-BR[VOICE], though, and this lack of evidence may
explain why voicing identity has no effect on rate of raising (see (117)).
There are other scattered sources of evidence for the rankings of CORR-BR
constraints, such as the fact that in disyllabic reduplication, the second syllable of the
reduplicant has a coda only if the base is just two syllables long (i.e., mag-ka-basag-
basa@g ‘to get thoroughly broken’ from ba@sag ‘break’, but mag-pa-baliù-baligta@d ‘to toss
and turn’ from baligta@d ‘upside-down’); we could say that TOTAL143 >> NOCODA >>
MAX-BR; this pushes the ranking of MAX-BR down.
Because the frequencies of all these types of evidence are not known, and in some
cases, such as nasal assimilation, the analysis itself is disputable (see fn. 112), this
chapter does not present simulations of learning like the one in §2.6. The somewhat
arbitrary grammar shown in (130), produces the rates of raising on novel words with mid-
vowel penults shown in ): greater internal similarity leads to a greater probability that
raising will be suppressed.

141
though see fnn. 112 and 115.
142
The glottal stop that *C? would delete is that of the base. The glottal stop of the reduplicant cannot be
deleted (to yield CORR-BR-satisfying alat-alat-a@n) because it would create an onsetless syllable, which is
prohibited in careful speech.
143
a binary constraint requiring the reduplicant to copy all of the base.

180
(130) Grammar used in simulation
constraint ranking value
MATCHCONTEXT 120
IDENT-IO[HIGH] 120
REDUP 108
IDENT-BR[HIGH] 110
IDENT-BR[BACK] 110
IDENT-BR[PLACE] 110
IDENT-BR[MANNER] 110
IDENT-BR[VOICE] 110
IDENT-BR[LENGTH] 110
MAX-BR 110
DEP-BR 110
*NONFINALMID 108
PU[HI] 106

(131) Rate of raising in novel words with mid penults, using the grammar in (130)

100%
90%
80%
70%
60%
fail to raise
50%
raise
40%
30%
20%
10%
0%
0 1 2 3 4 5 6 7
number of shared properties

The effect of internal similarity is not as sharp as in (121), but these are only the
rates of raising for novel words. In Chapter 3, small differences in rate of nasal
substitution in novel words became magnified as words were assimilated into the lexicon
(compare, for example, (57) and (83)).

181
 4.10. Chapter summary

This chapter has applied the model of lexical regularities developed in Chapter 2 to the
case of vowel raising in Tagalog, exceptions to which are found almost exclusively
among loanwords. The best predictor that a loanword would fail to raise under suffixation
is a mid vowel in the penult; it was argued that the mechanism preventing raising in these
cases is reduplicative correspondence between the penult and the ultima. This analysis
was supported by the finding that within mid-penult loanwords, similarity along other
dimensions between penult and ultima further increases the probability of nonraising
(because internal similarity favors a reduplicative construal).

182
 4.11. Appendix: statistical significance of influences on raising

To determine the statistical significance of the various claimed influences on raising, I

used contingency table analysis (see §2.2.2). To test whether the mid-vowel-in-penult
effect in (108) was significant, we can construct a table with the observed number of
stems with and without a mid vowel in the penult that raised or failed to raise,144 as in
(10), and a similar table with the “expected” values—the values that we would see if
raising and mid-vowel penult were independent of each other—as in (11).

(132) Raising and mid vowel in the penult: observed frequencies

raise don’t raise total
yes mid vowel in penult 59 30 89
no mid vowel in penult 186 13 199
total 245 43 288

(133) Raising and mid vowel in the penult: expected frequencies

raise don’t raise total
yes mid vowel in penult 75.712 13.288 89
no mid vowel in penult 169.288 29.712 199
total 245 43 288

The observed and expected values are quite different. It was expected that about
30 non-mid-penult stems would fail to raise, but only 13 did; it was expected that only
about 13 mid-penult stems would fail to raise, but 30 did. In other words, nonraising is
more common than expected among mid-penult stems, and less common than expected
among non-mid-penult stems.

144
Stems whose pronunciation varies were not included. The more rows and columns in a contingency
table, the more likely that the table of observed values will differ significantly from the table of expected
values. Using fewer rows and columns produces more conservative results.

183
 To test the significance of the differences between the observed and expected
values, we look at χ2. In this case, χ2 = 35.8. Given the number of rows and columns in
the table, the probability p that a χ2 value this big or bigger would be obtained by chance
is less than 0.0001. We can conclude that it is extremely likely that having a mid vowel in
the penult encourages nonraising. Fisher’s Exact Test also yields a p < 0.0001 that a table
with this degree of skew or higher could have arisen by chance if the two variables
(penult vowel and raising) were independent.
Significance measures for all proposed inhibitors of raising are summarized in
(134).

(134) Statistical significance of various inhibitors of vowel raising

χ2 Fisher’s exact test
mid (not high or low)vowel in penult χ2 = 35.756, p < 0.0001 p < 0.0001
matching backness χ2 = 32.508, p < 0.0001 p < 0.0001
mid vowel in penult (not antepenult) χ2 = 8.345, p = 0.0039 p = 0.0037
simple onset-same place χ2 = 3.250, p = 0.0714 p = 0.1012
simple onset-same manner χ2 = 1.107, p = 0.2928 p = 0.4268
onset shape χ2 = 7.331, p = 0.0068 p = 0.0066
rime shape χ2 = 4.178, p = 0.0433 p = 0.0705
vowel length χ2 = 1.676, p = 0.1654 p = 0.2552

The results for onset place and manner are not very impressive, probably because
the number of relevant observations is very small (remember, we are looking only at
stems with a mid vowel in the ultima and the penult, and with simple onsets in both
ultima and penult) and so the skew would have to be very great to get a satisfactorily
small value for p. The lack of significance for vowel length may also reflect the number
of observations: because penultimate stress is so common in English and Spanish, most
of the English and Spanish loanstems have penultimate stress/length; there were only 17
stems in which both vowels were short.

184
 5. Alternatives to Encoding Lexical Regularities in the Grammar
The preceding chapters have developed a model in which speakers’ apparent knowledge
of lexical regularities is encoded directly into the grammar, by constraints whose ranking
is learned through exposure to the lexicon: constraints that many words violate become
lower-ranked than constraints that few words violate. Although these constraints are
ranked low enough to be irrelevant in the production and perception of common, existing
words (for which only the requirement that listed words be faithfully used matters), they
come into play in the production of novel words and in rating their acceptability. As
discussed in Chapter 0, however, there are other ways to model behavior that appears to
reflect knowledge of lexical regularities. This chapter will consider some of those
alternatives. Section 5.1 discusses the possibility of encoding lexical regularities in a
separate perception grammar; §5.2 discusses the possibility of letting lexical regularities
emerge from the lexicon itself, using associative memory; and §5.3 discusses the dual-
mechanism model.

5.1. A separate module

An alternative to encoding lexical regularities in the same grammar that maps inputs to
outputs (the production grammar) is to encode them in a separate perception grammar.
The perception grammar would be responsible for recognizing words, and for generating
acceptability judgments.145
One advantage of having separate production and perception grammars is that it
could explain the disparity between speakers’ low rate of nasal substitution on novel

145
Similarly, rather than a grammar specifically for perception, the language system might contain a
module that lists lexical regularities in some form and is available for use in a variety of tasks.

185
words (in the production grammar, the constraints inhibiting nasal substitution usually
outrank the constraints promoting nasal substitution) and listeners’ high acceptability
ratings for certain novel substituted words (the perception grammar directly reflects
lexical frequencies, and for some obstruents, substitution is more frequent—and therefore
more acceptable—than nonsubstitution). The production grammar would still have to
encode nasal substitution, however (although the experimental results in §2.3.2.1 do not
provide evidence that the production grammar must include the patterns within nasal
substitution). And the perception grammar would have to somehow assign high ratings to
correctly produced existing words, even if they went against the prevailing pattern.146
But the account in §2.8.3 of acceptability judgments solves the problem of the
production/perception disparity without resorting to separate grammars: acceptability
ratings for novel substituted words can be high because the listener must consider the
possibility that the word in question, although novel to her, is not novel for her
interlocutor. As shown in (71), the acceptability ratings generated by the single-grammar
model were close to those produced by experimental participants.
Can the separate-grammars approach account for the assimilation of new words?
The single-grammar model used in Chapter 3 depended on Bayesian reasoning on the
part of the listener to give an advantage to substituted pronunciations of novel words such
that they were more likely to become listed in the lexicon than unsubstituted
pronunciations. A separate-grammars approach could achieve the same result by having
listeners use acceptability judgments to determine whether or not to add a pronunciation

146
This is not to say that an existing word that goes against the patterns of the lexicon must be rated as high
as an existing word that does not, but rather that an existing word that goes against the patterns must be
ranked higher than a novel word that goes against the patterns.

186
to the lexicon.147 For example, if a listener hears unsubstituted novel mampupuntol, the
perception grammar will assign it a low acceptability rating (because most p-stems in the
lexicon do substitute), and this low rating would inhibit adding mampupuntol to the
lexicon. Substituted novel mamumuntol, on the other hand, would receive a high
acceptability rating and thus be likely to be added to the lexicon.
The single-grammar model in Chapter 3 also relied on both speakers and listeners
to ensure that novel words with certain stem-initial obstruents have a higher probability
of becoming listed as substituted than novel words with other stem-initial obstruents. The
separate-grammars approach would rely solely on the listener, which does not seem
problematic.
The use of separate production and perception grammars, then, is workable, but
offers no empirical advantages over the use of a single grammar. The separate-grammars
model is not simpler than the single-grammar model: lexical regularities still must be
learned from the lexicon and stored. Moreover, there is duplication between the two
grammars: both perception and production grammars must encode at least nasal
substitution, if not the regularities within it.

5.2. Associative memory

It is possible that discrete knowledge of lexical regularities is not present anywhere in the
mind: behavior that appears to reflect such knowledge could emerge directly from the
lexicon itself. For example, in order to decide how to produce a novel word, the speaker
would not consult the grammar, but rather would select one or more similar, existing
words—perhaps the words that are activated first by feeding the novel word into an

147
A possible mechanism: the probability of adding a pronunciation to the lexicon as a single word is equal
to the acceptability of that pronunciation.

187
associative network (see, e.g., Rumelhart & McClelland 1986, Daugherty & Seidenberg
1994). The speaker would then apply the behavior of the existing words (or, if the
existing words disagree, perhaps the majority pattern) to the novel word.
In the case of nasal substitution, a novel p-stem word would tend
disproportionately to activate existing p-stem words, whereas a novel g-stem word would
tend to activate existing g-stem words. As a result, a speaker would substitute novel p-
stem words at a higher rate than novel g-stem words, and thus the behavior of novel
words would tend to match that of existing words. There would have to be some
additional bias against nasal substitution in the system to reproduce the experimental
result in §2.3.2 that the rate of substitution on novel words was much lower than the rate
of substitution among existing words.
Acceptability ratings would be derived similarly: the closer a novel word is to a
randomly selected (similar) existing word or group of words, the more acceptable it
would be. For example, in order to rate the acceptability of a nasal-substituted novel p-
stem word, the novel word could be compared to the first several existing words that
were activated by feeding the novel word into an associative network. The activated
existing words would be likely to derive from p-stems, and thus would be likely to
substituted. Because many of the activated existing words would be substituted, nasal
substitution on the novel stem would receive a high acceptability rating. A novel g-stem
word, on the other hand, would be likely to activate g-stem words, which are unlikely to
be substituted, and so substitution on the novel word would receive a low acceptability
rating. This idea is not refuted by the experimental data for nasal substitution in §2.3.3.1.
We would still need a mechanism that allows new words to become listed as
nasal-substituted despite a low initial rate of substitution. This could be accomplished by
having the listener use Bayesian reasoning as in §2.8, but using the comparison-to-

188
existing-words method rather than the grammar to estimate P([output] | /input/)—
assuming, still, some mechanism that keeps the rate of substitution lower on novel words
than the lexicon alone would dictate.
Even with a mechanism to prevent alternation in new words, and preserving
Bayesian reasoning, the very idea of comparison to existing words becomes problematic
in the case of vowel raising. As argued in Chapter 4, it is apparent from the distribution
of exceptions to vowel raising among loanwords that an important factor in determining
whether or not a novel word will undergo raising is the degree of similarity between the
word’s penult and ultima. How would a “similar existing word” be chosen when deciding
whether to apply vowel raising to a novel word? We would need a novel word like geke,
whose penult and ultima are identical except in onset voicing, to activate existing words
whose penult and ultima are similar to the same degree, such as todo. This means that the
criteria for similarity cannot involve merely shared segments or features, but would need
to include “internal similarity score” as a possible dimension of similarity.
Even if the lexicon could be structured in such a way as to allow words to activate
other words with similar internal similarity scores (whether through explicit encoding of
internal similarity, through computing similarity scores afresh when necessary, or by
some other mechanism), there remains a problem: exceptions to vowel raising under
suffixation are found almost exclusively among recent loanwords (I found only one
exception in the native vocabulary, pa-dede-hin ‘to give a baby a bottle’), although
failure to raise in disyllabic reduplication is fairly common, perhaps because of the
prosodic boundary between the reduplicant and the base (e.g. haùlo? ‘mixture’, haùlo-
ha@ùlo? ~ haùlu-ha@ùlo? ‘(drink made with shaved ice)’).148

148
See §4.4.1.

189
 Perhaps today new loanwords’ behavior could be determined by analogy to
existing loans, but what determined the behavior of the first loans? The existing words
activated by any then-novel word would have displayed raising; differences in probability
of raising among early loans could not have come from the lexicon. The model of vowel
raising presented in Chapter 4 avoids this problem by having differences in probability of
raising come from the grammar, not from the lexicon. The constraints responsible for the
differences (REDUP, CORR-IO, and CORR-BR) are universal, and their ranking can be
learned from facts other than vowel raising itself (see §4.9).

5.3. The dual mechanism model

The dual mechanism model (Pinker & Prince 1994) combines associative memory with a
traditional output grammar. The output grammar is responsible for productive
morphology and phonology; lexical regularities emerge from associative memory. Pinker
and Prince proposed the dual-mechanism to account for the behavior of English past
tense: in the majority (typewise) of verbs, the past tense is formed by adding the suffix -
ed, whose allomorphs [t], [d], and [«d] are predictably distributed (as in [lUk] ‘look’,
[lUk-t] ‘looked; [bEg] ‘beg’, [bEg-d] ‘begged’; [Qd] ‘add’, [Qd-«d] ‘added’). There are
quite a few irregular verbs (many of which are highly frequent), whose past tenses are
irregular (e.g., [sIN] ‘sing’, [sQN] ‘sang’; [titS] ‘teach’, [tt] ‘taught’). The irregulars are
patterned, in the sense that often irregulars whose past tense is formed in the same way
share other characteristics. For example, many of the verbs whose past tense is formed by
changing the vowel [I] to [Q] have a velar nasal in the coda and an alveolar in the onset
(sing, ring, sink, shrink, drink).
Pinker and Prince propose that when a verb has a listed past tense (this is true of
irregulars and perhaps some very frequent regulars), that past tense is used, but when a

190
verb lacks a listed past tense, the grammar supplies the regular suffix and chooses the
correct allomorph. Speakers sometimes supply irregular past-tense forms for novel words
(Bybee & Moder 1983), and their probability of doing so is influenced by the novel
word’s resemblance to existing irregulars (Prasada & Pinker 1993); these facts are
attributed to the effects of associative memory: the process of checking the lexicon to see
if a word has a listed past tense form activates the past tense forms of similar words, and
may result in the coining of an irregular past tense form.149
Pinker and Prince seem to conceive of the difference between regulars and
irregulars as twofold: (i) irregulars’ past tense forms must always be listed, whereas
regulars’ past tense forms are usually150 not listed and must be synthesized; and (ii)
patterns in the distribution of irregulars (such as the [IN]/[QN] pattern) exist only in the
lexicon, whereas the regular pattern (add -ed) comes from the grammar. But only (i) is
crucial: the evidence for a qualitative difference between regulars and irregulars can be
explained solely in terms of the difference between listed and synthesized forms.
The remainder of this section goes through several pieces of evidence for a
qualitative difference between regulars and irregulars, attempting to explain them in
terms of the model proposed in Chapter 2, with the assumption that regulars generally
lack a listed past tense (why this should be so is returned to at the end of the section). The
stochastic grammar for English past tense would have very high-ranking USELISTED and
faithfulness constraints (to ensure that listed pasts are used, and faithfully so), as well as a
large group of constraints Xpresent / Xtpast (“a verb stem of the form X in the present tense

149
This raises again the question from §2.5 and §4.7 of a three-way distinction: existing irregulars vs.
existing regulars vs. novel words that may be treated as irregular or regular. See §4.7.2.
150
See below for the conditions that can lead to listing of regulars.

191
should be of the form Xt in the past tense”),151 XINpresent / XΘNpast, X¨INYpresent / X¨ΘNYpast,
etc.152 Note that these constraints are of varying degrees of specificity, so a given verb
would be subject to more than one. Some of these present/past constraints are ranked
high (because there is much evidence for them), others low (such as XeIpresent / XEdpast,
exemplified only by say/said).

5.3.1. Evidence for a qualitative difference between irregulars and regulars

One piece of evidence for a difference between irregulars and regulars is Ullman’s (1999)
study of acceptability judgments for the past-tense forms of existing words. Ullman
found that the acceptability of irregular pasts depended on both the frequency of the past
itself and the frequency of the verb stem. Acceptability judgments for regular pasts,
however, depended only on the frequency of the stem. The interpretation is that only
irregulars have a listed past tense: without a separate lexical entry to reflect its frequency,
a regular past’s acceptability must rely solely on information in the stem’s lexical entry.
With the assumption that (most) regulars lack a listed past tense, the result is also
easy to interpret in the model proposed here. Under the view of acceptability adopted in
§2.8.3, a word’s acceptability is a function of its probability of being pronounced (Hayes
& MacEachern 1998, Hayes to appear). The probability of a particular pronunciation
depends, in turn, on two factors: what the set of available inputs is likely to be, and which
input-output pair the grammar is likely to choose. The frequency of an irregular past like
sang affects its acceptability because frequency largely determines Listedness, which in

151
There might be separate constraints for the three allomorphs of -ed, or just one that interacts with
markedness constraints to produce the correct allomorph in each situation.
152
See Albright & Hayes (1999) for how such constraints can be synthesized, and their relative rankings
learned, on the basis of evidence from the lexicon.

192
turn affects the likelihood that /sΘN/ would be available as an input. For example, if /sΘN/
is available, the high ranking of USELISTED and faithfulness constraints will almost
always make /sΘN/→ [sΘN] the optimal candidate. If /sang/ is not available, then
/sIN/+past→ [sΘN] is still a reasonable candidate (it satisfies XINpresent / XΘNpast), but so
are /sIN/+past→ [sINd] , /sIN/+past→ [sÃN] , and others, so the probability of getting the
output [sΘN] (and thus its acceptability) is reduced. For a regular verb with no listed
past, however, only synthesized candidates can be under consideration—the probability
of retrieving a listed past is always zero, no matter what the frequency.
Ullman also found that acceptability ratings of irregulars depended on their
“neighborhood size” (the number of similar stems whose past tense is formed in the same
way), whereas acceptability ratings of regulars did not. The dual-mechanism
interpretation is that regular pasts are unaffected by neighborhood size because they are
generated by a rule of the grammar, which is not sensitive to how many words follow it.
The explanation for Ullman’s finding in the model proposed here is that neighborhood
size affects the acceptability of irregular pasts because it determines (during learning) the
ranking of the past/present constraints that those pasts obey. For example, because
sing/sang is in a large neighborhood, the constraint XINpresent / XΘNpast is high-ranked,
increasing the production probability of every candidate with the output [sΘN] . Why no
neighborhood effect for regulars? It may be that in English, the general constraint Xpresent /
Xtpast is ranked so high that it swamps the effects of more specific constraints like
Xkpresent / Xktpast. Albright (1998, 1999), found that in assigning novel Italian verbs to
conjugation classes and rating their acceptability, judges were indeed sensitive to
neighborhoods within the default (regular) pattern. This may be because the particular
facts of Italian do not lead to a one single constraint for the regulars that is strong enough

193
to swamp the effect of the others; the fact that Albright’s search for neighborhoods was
more exhaustive may also have played a role.
Qualitative difference also exist in producing past-tense forms. Prasada, Pinker,
and Snyder (1990) found that speaker’s speed in producing irregular past tense forms
depended on the frequencies of both the past tense form itself and the verb stem. The
speed of producing regulars, on the other had, depended only on the frequency of the
stem. This finding makes sense in the model proposed here, because producing an
irregular past tense involves both retrieving it from memory and applying the grammar—
the frequency of the listed past would affect the speed of retrieving it. But in producing a
regular, there is never a listed past tense to affect the computation.153
A final qualitative difference between regulars and irregulars is in priming:
Stanners et al. (1979) found that irregular pasts prime their stems somewhat, but that
regulars prime their stem as well as the stem itself (in an all-visual-priming task). The
interpretation is that because irregular pasts are listed separately, recognizing them only
weakly activates the related entry for the stem. But recognizing a regular past requires
accessing the stem itself, because there is no separate, listed past. In the model presented
here, recognizing a regular past requires looking for the stem that, when run through the
grammar, would produce the right result. Recognizing an irregular past, on the other
hand, would not require activating the stem as thoroughly. If the grammar operates by
whittling away the set of candidates, starting by eliminating those that violate the highest-
ranked constraints, then once a listed irregular past was found, candidates synthesized

153
I assume, as the dual-mechanists do, that searching the lexicon and applying the grammar can apply in
parallel—if the grammar were applied only after the search of the lexicon was complete, regulars would
always be slower than the slowest irregular, because the speaker would have to search the entire lexicon
before concluding that no listed form existed and moving on to applying the grammar to synthesized inputs.
In my case, the grammar could work on evaluating the synthesized input-output pairs while waiting to see
what listed inputs might be available.

194
from the stem and an affix would be eliminated from consideration (because they violate
top-ranked USELISTED), and so the period during which the stem was activated would be
brief.
To summarize, the qualitative difference between regulars and irregulars can be
reduced to the difference between having a listed past-tense form (irregulars), and lacking
one (regulars). The English past tense may not be a case that argues strongly for putting
constraints that capture lexical regularities into the grammar (e.g., XINpresent / XÃNpast), but
neither is it an argument for keeping lexical regularities out of the grammar. The next
section considers the reasons for the difference in listedness between regulars and
irregulars.

5.3.2. Why are regular pasts not listed?

The account of listener behavior in §2.8 proposed that when a listener hears a word for
which she has no lexical entry, she must guess whether or not her interlocutor might have
been using a lexical entry unfamiliar to the listener (as opposed to concatenating some
familiar morphemes). If the listener guesses that the speaker was using a lexical entry, the
listener begins to build one herself. Every time the listener guesses that some speaker was
using a lexical entry for this word, she strengthens her own entry. In order to guess
whether or not the speaker was using a lexical entry, the listener applies Bayes’ Law: all
else being equal, the probability that the speaker was using a lexical entry is proportional
to the probability that the utterance the listener heard would have occurred if the listener
had been using a lexical entry. Similarly, the probability that the speaker was creating a
synthetic form is proportional to the probability that the utterance heard would have been
produced if the speaker had been using a synthesized input.

195
 When a listener hears a past tense form like said whose probability of being
produced by synthesis is low (the XeIpresent / XEdpast constraint is not very high-ranked),
she is likely to conclude that it must have come from a listed form and update her lexicon
accordingly. When she hears a regular past like jumped, on the other hand, she is likely to
conclude that it was produced by synthesis (because Xpresent / Xtpast is ranked so high) and
not add anything to her lexicon.
Regular pasts can become stored under certain circumstances. Because the
probability of obtaining a regular result from synthesis is never 100%, the listener may
occasionally guess that a regular past was listed and add it to her lexicon. If this listed
guess happens enough times for a particular past, that past can develop a strong lexical
entry. One way to produce many incidents in which the listener guesses that a word is
listed, even if such guesses are improbable, is simply for the word to be highly frequent.
There is indeed evidence that high-frequency regulars have a tendency to become stored:
Stemberger and MacWhinney (1986) found that error rates in forming the past tense of
regular verbs were lower for verbs with high-frequency past-tense forms; Baayen,
Dijkstra, and Schreuder (1997), found faster reaction times in a lexical-decision task for
high-frequency regular noun plurals in Dutch than for low-frequency noun plurals, even
holding constant the frequency of the singular form;154 Sereno and Jongman (1997) found
that for English regular noun plurals, reaction time was also correlated with frequency of
the inflected form. When frequency of the inflected form has an effect on behavior, the
interpretation is that the inflected form must be listed (i.e., if the inflected form were not
listed, behavior would depend solely on the frequency of the stem).

154
Baayen et al. did not find a frequency effect for Dutch verbs.

196
 Being a regular in a strong irregular neighborhood should also encourage the
formation of a lexical entry. For example, blink (past tense blinked, not *blunk or
*blank), violates the constraint XINX / XANX. Because a large neighborhood gives XINX /
XANX a high ranking (see Albright & Hayes 1999), the irregular synthesized candidate
/blINk/ + past→ [blΘNk] makes the regular synthesized candidate /blINk/ + past→
[blINkt] less of a sure winner than it would be for most regulars. This decreases the
probability of obtaining [blINkt] from synthesis, and thus makes the listener more likely
to guess that the word is listed. Ullman and Pinker 1991 found evidence that past-tense
forms like blink are indeed stored—their frequency influences their acceptability ratings.
Finally, the regular members of past-tense doublets (such as dived/dove—many
speakers are unsure which is the correct past-tense form of dive, and both are common)
have a tendency to become listed. This is because the presence of the strong competing
candidate /doUv/ → [doUv] reduces the likelihood that [daIvd] would be the optimal
output if the input /daIvd/ were not available. When the listener hears [daIvd] , then, she is
more likely to guess that it was listed. Ullman and Pinker (1990), found that acceptability
ratings regular (and irregular) members of doublets correlated with their frequency.
To summarize, the difference in listedness between regulars and irregulars need
not depend on a prior qualitative difference between the two. Rather, given a grammar
that tends to produce regular outputs for synthesized inputs, listener reasoning will
prevent most regulars from becoming listed. This difference in listing then leads to the
apparent qualitative difference between regulars and irregulars discussed in §5.3.1.

197
 6. Summary
The preceding chapters have proposed a model of grammar to account for the effect of
lexical regularities in speaking, listening, and the evolution of the lexicon. The grammar
is a basic OT grammar, but with stochastic constraint ranking. Reliably high-ranked
constraints ensure the stable behavior of listed words, but variably ranked subterranean
constraints come in to play for novel words. Boersma’s (1998) Gradual Learning
Algorithm (which was designed to handle free variation) was shown to be capable of
learning a grammar of this type through exposure to rates of lexical variation.
Candidates in this model consist of input-output pairs (rather than outputs that all
share the same input), so for both speakers and listeners, single-lexical-entry inputs
compete with synthesized inputs composed of strings of morphemes. In particular, in
order to form acceptability judgments and to decide whether and how to update her
lexicon, a listener must guess whether her interlocutor has used a listed word or has
synthesized a new word.
When a speaker utters a novel, morphologically complex word, only synthesized
input-output pairs are available. In the case of nasal substitution, the grammar that the
Gradual Learning Algorithm learned produces a low rate of nasal substitution when only
synthesized candidates are available. But when a listener hears a novel word, she cannot
be certain that the word was novel for her interlocutor; she must take into account the
chance that the pronunciation she heard could have come from a listed input. By
performing this reasoning, the model was able to emulate the experimental finding of
high acceptability for nasal-substituted novel words despite the low productivity of nasal
substitution on novel words.

198
 The low rate of nasal substitution on novel words also produced a challenge for
the assimilation of new words into the lexicon: if nonsubstitution is the majority
pronunciation, why does it not always win out? Why do some words eventually become
listed as substituted? The answer given was that in assimilating new words into the
lexicon (i.e., gradually developing lexical entries for them that are nasal-substituted or
not), Bayesian listener reasoning produces a bias in favor of nasal-substituted
pronunciations such that they have a disproportionately good chance of being added to
the lexicon. A computer simulation confirmed that high rates of nasal substitution in
assimilated words can be obtained despite low initial rates of nasal substitution when the
words are new.

199
 References

The American Heritage Larousse Spanish dictionary. (1986) Boston, Houghton Mifflin.
Ethnologue: Languages of the World (1996). 13th edition. Barbara Grimes and Joseph
Grimes, editors. Dallas TX, Summer Institute of Linguistics.
Handbook of the International Phonetic Association: A Guide to the Use of the
International Phonetic Alphabet. (1999) Cambridge, England, Cambridge
University Press.
Albright, Adam (1998). Phonological subregularities in productive and unproductive
inflectional classes: Evidence from Italian. MA thesis, UCLA.
Albright, Adam (1999). “The default is not a unitary rule.” Paper presented at the Annual
Meeting of the Linguistic Society of America in Los Angeles.
Albright, Adam and Bruce Hayes (1999). “An Automated Learner for Phonology and
Morphology.” Manuscript, UCLA.
Anttila, Arto (1997). “Deriving Variation from Grammar.” In Variation, Change and
Phonological Theory. Frans Hinskens, Roeland van Hout, and W. Leo Wetzels,
editors. Amsterdam, Benjamins: 35-68.
Archangeli, Diana and Terence Langendoen (1997). Optimality Theory: An Overview.
Malden MA, Blackwell.
Archangeli, Diana, Laura Moll, and Kazutoshi Ohno (1998). “Why not *NC8?” To appear
in the proceedings of the 34th annual meeting of the Chicago Linguistic Society.
Aronoff, Mark (1976). Word formation in generative grammar. Cambridge MA, MIT
Press.
Baayen, R. Harald, Ton Dijkstra, and Robert Schreuder (1997). “Singulars and plurals in
Dutch: Evidence for a parallel dual-route model.” Journal of Memory and
Language 37: 94-117.
Baroni, Marco (1997). The representation of prefixed forms in the Italian lexicon:
evidence of intervocalic [s] and [z] . MA thesis, UCLA.
Bellwood, Peter (1979). Man's conquest of the Pacific : the prehistory of Southeast Asia
and Oceania. New York, Oxford University Press.
Benua, Laura (1998). Transderivational Identity. PhD dissertation, University of
Massachusetts, Amherst.
Berkley, Deborah Milam (1994). “The OCP and Gradient Data.” Studies in the Linguistic
Sciences 24: 59-72.

200
Berko, Jean (1958). “The Child's Learning of English Morphology.” Word 14: 150-177.
Blake, Frank Ringgold (1925). A grammar of the Tagalog language, the chief native
idiom of the Philippine Islands. New Haven CT, American Oriental Society.
Bloomfield, Leonard and Alfredo Viola Santiago (1917). Tagalog texts with grammatical
analysis. Urbana IL, University of Illinois.
Boersma, Paul (1998). Functional phonology : formalizing the interactions between
articulatory and perceptual drives. The Hague, Holland Academic Graphics.
Boersma, Paul and Bruce Hayes (1999). “Empirical tests of the Gradual Learning
Algorithm.” Manuscript, University of Amsterdam and UCLA.
Bybee, Joan L. (1985). Morphology : a study of the relation between meaning and form.
Amsterdam, Benjamins.
Bybee, Joan and Carol Lynn Moder (1983). “Morphological Classes as Natural
Categories.” Language 59: 251-270.
Bybee, Joan and Dan Slobin (1982). “Rules and Schemes in the Development and Use of
the English Past Tense.” Language 58: 269-285.
Carrier, Jill Louise (1979). The interaction of morphological and phonological rules in
Tagalog : a study in the relationship between rule components in grammar. PhD
dissertation, Massachusetts Institute of Technology.
Cho, Taehong (to appear). “The specification of intergestural timing and gestural
overlap.” UCLA Working Papers in Phonology 4.
Cohn, Abigail and John McCarthy (1998). “Alignment and Parallelism in Indonesian
Phonology.” Working Papers of the Cornell Phonetics Laboratory 12: 53-137.
Crosswhite, Katherine (1996). Positionality and cyclicity in Chamorro phonology. MA
thesis, UCLA.
Crosswhite, Katherine (1998). Segmental vs. Prosodic Correspondence in Chamorro.
Phonology 15: 281-316.
Crosswhite, Katherine (1999). Vowel Reduction in Optimality Theory. PhD dissertation,
UCLA.
Daugherty, Kim and Mark Seidenberg (1994). “Beyond Rules and Exceptions: A
Connectionist Approach to Inflectional Morphology.” In The Reality of Linguistic
Rules. Susan Lima, Roberta Corrigan, and Gregory Iverson, editors. Amsterdam,
Benjamins: 353-88.
De Guzman, Videa (1978). “A Case for Nonphonological Constraints on Nasal
Substitution.” Oceanic Linguistics 17: 87-106.

201
Dempwolff, Otto (1969). Vergleichende Lautlehre des austronesischen Wortschatzes.
Nendeln, Kraus Reprint.
Dixon, Robert (1977). A Grammar of Yidiø. Cambridge, Cambridge University Press.
English, Leo James (1986). Tagalog-English dictionary. Manila, Congregation of the
Most Holy Redeemer. Distributed by (Philippine) National Book Store.
Forster, Kenneth and Susan Chambers (1973). “Lexical Access and Naming Time.”
Journal of Verbal Learning & Verbal Behavior 12: 627-635.
French, Koleen Matsuda (1988). Insights into Tagalog: Reduplication, Infixation, and
Stress from Nonlinear Phonology. Dallas TX, Summer Institute of Linguistics and
University of Texas at Arlington.
Frisch, Stefan (1996). Similarity and Frequency in Phonology. Dissertation,
Northwestern University.
Frisch, Stefan (to appear). “Emergent phonotactics and judgments of well-formedness.”
University of Alberta Papers in Experimental and Theoretical Linguistics 6.
Frisch, Stefan, Michael Broe, and Janet Pierrehumbert (1996). “Similarity and
phonotactics in Arabic.” Manuscript, Northwestern University.
Frisch, Stefan and Bushra Zawaydeh (to appear). “The psychological reality of OCP-
Place in Arabic.” Language.
Hale, Mark and Charles Reiss (1998). “Formal and Empirical Arguments Concerning
Phonological Acquisition.” Linguistic Inquiry 29: 656-83.
Halle, Morris (1959). The Sound Pattern of Russian. The Hague, Mouton.
Hammond, Michael (1999). “English stress and cranberry morphs.” Paper presented at
the Annual Meeting of the Linguistic Society of America in Los Angeles.
Hayes, Bruce (1999). OTSoft. Software package,
http://www.humnet.ucla.edu/humnet/linguistics/people/hayes/otsoft/.
Hayes, Bruce (to appear). “Gradient Well-formedness in Optimality Theory.” In
Conceptual Studies in Optimality Theory. Joost Dekkers, Frank van der Leeuw,
and Jeroen van de Weijer, editors.
Hayes, Bruce and May Abad (1989). “Reduplication and Syllabification in Ilokano.”
Lingua 77: 331-374.
Hayes, Bruce and Margaret MacEachern (1998). “Quatrain Form in English Folk Verse.”
Language 74: 473-507.
Hayes, Bruce and Tanya Stivers (1996). “The Phonetics of Postnasal Voicing.”
Manuscript, UCLA.

202
Ingram, David (1974) “Fronting in Child Phonology.” Journal of Child Language 1: 233-
41.
Inkelas, Sharon (2000). “Infixation obviates backcopying in Tagalog.” Paper presented at
the Annual Meeting of the Linguistic Society of America in Chicago.
Inkelas, Sharon, Orhan Orgun, and Cheryl Zoll (1997). “The Implications of Lexical
Exceptions for the Nature of Grammar.” In Derivations and Constraints in
Phonology. Iggy Roca, editor. New York, Oxford University Press: 393-418.
Itô, Junko and Armin Mester (1995). “Japanese Phonology.” In The Handbook of
Phonological Theory. John Goldsmith, editor. Cambridge MA, Blackwell: 817-
838.
Itô, Junko, Armin Mester, and Jaye Padgett (1995). “Licensing and Underspecification in
Optimality Theory.” Linguistic Inquiry 26: 571-613.
Kager, René (1999). Optimality theory. Cambridge, Cambridge University Press.
Kaun, Abigail Rhoades (1995). The Typology of Rounding Harmony: An Optimality
Theoretic Approach. PhD dissertation, UCLA.
Kenstowicz, Michael (1997). “Uniform Exponence: Exemplification and Extension.”
University of Maryland Working Papers in Linguistics 5: 139-155.
Lapoliwa, Hans (1981). A Generative Approach to the Phonology of Bahasa Indonesia.
Canberra, Department of Linguistics, Research School of Pacific Studies,
Australia National University.
MacEachern, Margaret R. (1999). Laryngeal Cooccurrence Restrictions. New York,
Garland.
McCarthy, John and Alan Prince (1993). “Generalized Alignment.” Yearbook of
Morphology: 79-153.
McCarthy, John and Alan Prince (1994). “Optimality in Prosodic Morphology: the
emergence of the unmarked.” In Proceedings of the North East Linguistic Society
24: 333-379.
McCarthy, John and Alan Prince (1995). “Faithfulness and Reduplicative Identity.”
Manuscript, University of Massachusetts, Amherst and Rutgers University.
Newman, John (1984). “Nasal Replacement in Western Austronesian: An Overview.”
Philippine Journal of Linguistics 15-16: 1-17.
Newman, Stanley (1944). Yokuts language of California. New York, Johnson Reprint
Corp.
Ohala, John and Carol Riordan (1980). “Passive Vocal Tract Enlargement during Voiced
Stops.” Report of the Phonology Laboratory, University of California, Berkeley 5:
78-88.

203
Pater, Joseph (1996). “Austronesian Nasal Substitution and Other *NC8 Effects.”
Manuscript, McGill University.
Pater, Joseph (1999a). “The comprehension/production dilemma and the development of
receptive competence.” Manuscript, University of Alberta.
Pater, Joseph (1999b). “Generality and restrictiveness in constraint formulation:
Austronesian nasal substitution and child consonant harmony.” Handout from a
talk given at the University of Massachusetts, Amherst.
Pierrehumbert, Janet (1993). “Dissimilarity in Arabic Verbal Roots.” In Proceedings of
the North East Linguistics Society 23: 367-381.
Pinker, Steven and Alan Prince (1994). “Regular and Irregular Morphology and the
Psychological Status of Rules of Grammar.” In The Reality of Linguistic Rules.
Susan Lima, Roberta Corrigan, and Gregory Iverson, editors. Amsterdam,
Benjamins: 321-51.
Prasada, Sandeep and Steven Pinker (1993). “Generalisation of regular and irregular
morphological patterns.” Language and Cognitive Processes 8: 1-56.
Prasada, Sandeep, Steven Pinker, and William Snyder (1990). “Some evidence that
irregular forms are retrieved from memory but regular forms are rule generated.”
Paper presented at the Annual Meeting of the Psychonomic Society in New
Orleans. As cited in Pinker & Prince 1994.
Prince, Alan and Paul Smolensky (1993). Optimality theory: constraint interaction in
generative grammar. Technical reports of the Rutgers University Center for
Cognitive Science TR-2.
Ramos, Teresita and Maria Lourdes Bautista (1986). Handbook of Tagalog verbs:
inflections, modes, and aspects. Honolulu, University of Hawaii Press.
Ross, Kie (1996). Floating Phonotactics: Variability in Reduplication and Infixation of
Tagalog Loanwords. MA thesis, UCLA.
Rubenstein, Herbert, Lonnie Garfield, and Jane Millikan (1970). “Homographic Entries
in the Internal Lexicon.” Journal of Verbal Learning and Verbal Behavior 9: 487-
494.
Rumelhart, David and James McClelland (1986). “On Learning the Past Tenses of
English Verbs.” In Parallel Distributed Processing: Explorations in the
Microstructure of Cognition. Volume II: Psychological and Biological Models.
David Rumelhart, James McClelland and the PDP Research Group, editors.
Cambridge MA, MIT Press: 216-217.
Schachter, Paul and Fe Otanes (1972). Tagalog Reference Grammar. Berkeley,
University of California Press.

204
Sereno, Joan and Allard Jongman (1997). “Processing of English inflectional
morphology.” Memory and Cognition 25: 425-437.
Smolensky, Paul (1996a). “The Initial State and 'Richness of the Base' in Optimality
Theory.” Technical Report JHU-CogSci-96-4. Cognitive Science Department,
Johns Hopkins University.
Smolensky, Paul (1996b). “On the Comprehension/Production Dilemma in Child
Language.” Linguistic Inquiry 27: 720-731.
Stanners, Robert, James Neiser, William Hernon, and Roger Hall (1979). “Memory
representation for morphologically related words.” Journal of Verbal Learning
and Verbal Behavior 18: 399-412.
Stemberger, Joseph and Brian MacWhinney (1986). “Frequency and the lexical storage
of regularly inflected forms.” Memory and Cognition 14: 17-26.
Steriade, Donca (1987). “Locality conditions and feature geometry.” In Proceedings of
the North East Linguistics Society 17: 595-617.
Steriade, Donca (1995). “Underspecification and Markedness.” In The Handbook of
Phonological Theory. John Goldsmith, editor. Cambridge MA, Blackwell: 115-
174.
Steriade, Donca (1996) “Paradigm Uniformity and the Phonetics-Phonology Boundary.”
To appear in Papers in Laboratory Phonology. Michael Broe and Janet
Pierrehumbert, editors.
Steriade, Donca (1999). “Lexical Conservatism in French Adjectival Liaison.” In Formal
Perspectives on Romance Linguistics. Selected Papers from the 28th Linguistic
Symposium on Romance Languages. J.-Marc Authier, Barbara Bullock, and Lisa
Reed, editors. Amsterdam, Benjamins: 243-70.
Suzuki, Keiichiro (1999). “Identity ? similarity: Sundanese, Akan, and tongue twisters.”
Paper presented at the Annual Meeting of the Linguistic Society of America in
Los Angeles.
Tesar, Bruce (1998). “An Iterative Strategy for Language Learning.” Lingua 104: 131-
145.
Ullman, Michael and Steven Pinker (1990). “Why do some verbs not have a single past
tense?” Paper presented at the 15th Annual Boston University Conference on
Language Development. As cited in Pinker & Prince 1994.
Ullman, Michael and Steven Pinker (1991). “Connectionism versus symbolic rules in
language: The English past tense as a case study.” Paper presented at the Spring
Symposium of the American Association for Artificial Intelligence. As cited in
Pinker & Prince 1994.
Ullman, Michael T. (1999). “Acceptability ratings of regular and irregular past-tense
forms: Evidence for a dual-system model of language from word frequency and

205
 phonological neighbourhood effects.” Language and Cognitive Processes 14: 47-
67.
Walker, Rachel (2000). Long-Distance Consonantal Identity Effects. Paper presented at
the West Coast Conference on Formal Linguistics in Los Angeles.
Walker, Rachel (to appear). “Consonantal Correspondence.” University of Alberta Papers
in Experimental and Theoretical Linguistics 6.
Wilbur, Ronnie Bring (1973). The phonology of reduplication. Bloomington IN, Indiana
University Linguistics Club.
Zimmer, Karl E. (1969). “Psychological Correlates of Some Turkish Morpheme Structure
Conditions.” Language 45: 309-321.
Zoll, Cheryl (1993). “Directionless Syllabification and Ghosts in Yawelmani.” Transcript
of talk given at ROW-1, Rutgers University.
Zorc, R. David (1972). “Current and Proto Tagalic Stress.” Philippine Journal of
Linguistics 3: 43-57.
Zorc, R. David (1983). “Proto Austronesian Accent Revisited.” Philippine Journal of
Linguistics 14: 1-24.

206