Dabrowska 2008c

Download as pdf or txt
Download as pdf or txt
You are on page 1of 35

Questions with long-distance dependencies:

A usage-based perspective
EWA DA˛BROWSKA*

Abstract

Attested questions with long-distance dependencies (e.g., What do you


think you’re doing?) tend to be quite stereotypical: the matrix clause
usually consists of a WH word, the auxiliary do or did, the pronoun you,
and the verb think or say, with no other elements; and they virtually never
contain more than one subordinate clause. This has lead some researchers
in the usage-based framework (Da˛browska 2004; Verhagen 2005) to hy-
pothesise that speakers’ knowledge about such constructions is best ex-
plained in terms of relatively specific, low level templates rather than gen-
eral rules that apply ‘‘across the board’’. The research reported here was
designed to test this hypothesis and alternative hypotheses derived from
rule-based theories.

Keywords: Usage-based model; long-distance dependencies; unbounded


dependencies; acceptability judgment experiment; prototype
e¤ects.

1. Introduction
Questions and other constructions with long distance dependencies
(henceforth LDDs) have played an important role in the development of
syntactic theory, especially in the generative framework. Such structures

* This study was supported by the Arts and Humanities Research Council (grant number
AH/F001924/1). I would like to thank Adele Goldberg, Marcin Szczerbiński and two
anonymous referees for their comments on an earlier draft of this paper, and Mike Pin-
combe for help with data collection. Address for correspondence: School of English Lit-
erature, Language and Linguistics, University of She‰eld, She‰eld S10 2TN, United
Kingdom. E-mail: e.dabrowska@shef.ac.uk.

Cognitive Linguistics 19–3 (2008), 391–425 0936–5907/08/0019–0391


DOI 10.1515/COGL.2008.015 6 Walter de Gruyter
392 E. Da˛browska

are interesting because they exhibit a dependency between a ‘‘filler’’ in the


main clause and a ‘‘gap’’ in a subordinate clause, as in example (1). The
dependencies are frequently referred to as ‘‘unbounded’’, as, in principle,
there can be any number of clauses intervening between the filler and the
gap, as illustrated by (1d) and (1e).

(1) a. What will John claim that you did ? (Culicover 1997: 184)
b. Which problem does John know (that) Mary solved ?
(Ouhalla 1994: 72)
c. Whom do you believe that Lord Emsworth will invite ?
(Haegeman 1991: 342)
d. Who did Mary hope that Tom would tell Bill that he should
visit ? (Chomsky 1977: 74)
e. Which problem do you think (that) Jane believes (that) Bill
claims (that) Mary solved ? (Ouhalla 1994: 71)

It is noteworthy, however, that attested questions with LDDs are very


di¤erent from these constructed examples, as illustrated by the following
examples from the spoken part of the British National Corpus:

(2) a. And how do you think you’d spell classical like do you like clas-
sical music? (FMG 725)
b. Why’d why do you think why do you think it is that there
wasn’t that motivation? (FY8 201)
c. What is it and why do you think it looks like that? (JJS 882)
d. What do you think Brian’ll say? (KE1 256)
e. What did they say it meant? (KD0 622)

These ‘‘real life’’ LDD questions are much more stereotypical than the
sentences in (1). The textbook examples contain a variety of matrix sub-
jects and verbs and di¤erent auxiliaries; most of them also contain an
overt complementizer and two involve a dependency over more than one
intervening clause. In the corpus sentences, in contrast, the matrix subject
is usually you, the matrix verb think or say, and the auxiliary do; there are
no other elements in the matrix clause, no complementizer, and only one
complement clause. In fact, almost 70 percent of the LDD questions with
finite complement clauses in the spoken part of the BNC have the form
WH do you think S-GAP? or WH did NP say S-GAP?, where S-GAP is
a subordinate clause with a missing constituent. Most of the remaining
questions are minimal variations on these patterns: that is to say, they
contain a di¤erent matrix subject or a di¤erent verb or a di¤erent auxil-
iary or an additional element like an adverbial or complementizer. Only
Questions with long-distance dependencies 393

6 percent depart from the prototype in more than one respect (Da˛brow-
ska in press a; see also Da˛browska 2004 and Verhagen 2005).1
This has lead some researchers in the usage-based framework (Da˛b-
rowska 2004, Verhagen 2005) to hypothesise that speakers’ knowledge
about such constructions is best explained in terms of relatively specific,
low level templates—WH do you think S-GAP? and WH did you say S-
GAP?—rather than in terms of abstract rules and principles of the type
proposed by formal linguists (see, for example Cheng and Corver 2006;
Chomsky 1977; Levine and Hukari 2006).2 Declaratives with verb com-
plement clauses, in contrast, are much more varied—the main clauses
take di¤erent subjects, auxiliaries and verbs, and often contain additional
elements (Da˛browska in press a; Verhagen 2005)—as a result of which
language learners develop more general representations for this construc-
tion in addition to lexically specific templates for frequent combinations
such as I think S, I don’t think S, I mean S.
However, conclusions about speakers’ mental representations based on
the fact that a particular structure is rare or not attested at all in a corpus
are problematic, since this could be merely a result of sampling. Even
with a large and balanced corpus, sentences which are perfectly compati-
ble with speakers’ mental grammars may be unattested simply because
they are pragmatically implausible. In short, while restricted patterns of
usage are suggestive, they do not license strong conclusions about mental
representation: the observational data need to be corroborated by experi-
mental studies.
According to the usage-based proposals put forward by Da˛browska
and Verhagen, prototypical LDD questions are produced simply by in-
serting new material into the appropriate slots in a pre-existing template.
Non-prototypical LDD questions such as What does she hope she’ll get?
require additional work, since the speaker has to adapt the template—in
this case, substitute she for you and hope for think, and modify the auxil-
iary so that it agrees with the subject. To be able to do this, the speaker
would have to construct a proportional analogy such as the one in (3),

1. I am using the term ‘‘prototype’’ as it is usually used in linguistics: to refer to an ideal-


ised typical instance. Many natural categories are centred around a prototype, in the
sense that other instances are assimilated in the category on the basis of their perceived
similarity to it (cf. Lako¤ 1987; Langacker 1987). The properties of prototypical instan-
ces are thus shared by most other members of the category.
2. The term ‘‘construction’’ is used in this paper to refer to any grammatical pattern found
in any language: thus expressions such as ‘‘constructions with long-distance dependen-
cies’’ should not be taken as implying that such patterns necessarily have any mental
reality. I will use the terms ‘‘template’’ and ‘‘schema’’ to refer to speakers’ mental repre-
sentations of these patterns.
394 E. Da˛browska

where semantic structure is represented in CAPITALS and phonological


structure in italics:3
(3) YOU THINK SHE WILL GET SOMETHING: WHAT?
is to What do you think she’ll get?
as SHE HOPES SHE WILL GET SOMETHING: WHAT?
is to ???
To solve this problem, the speaker needs to establish correspondences
between the relevant parts of semantic and phonological structure:
YOU ¼ SHE, therefore the target expression will have the phonological
form corresponding to SHE, namely she, in place of the phonologi-
cal form corresponding to YOU, namely you, and so on. This requires
knowledge about linguistic categories (the speaker must know what can
be substituted for what), the internal structure of the source expression,
i.e., YOU THINK SHE WILL GET SOMETHING: WHAT?/What
do you think she’ll get? (so that he/she knows where to substitute it), and
about agreement (the auxiliary needs to agree with the new subject). A
listener or reader, of course, would have to use the phonological forms
of source and target, plus an understanding of the relationships between
their constituent parts, to construct a semantic representation of the
target.4
If speakers don’t have a ready-made template for non-prototypical
questions and have to extrapolate from existing knowledge in the manner
described above, such sentences will require extra e¤ort to produce and
understand (which should translate into longer processing times) and
should be judged to be less acceptable than more prototypical variants.
Both of these predictions can be tested experimentally; the present paper
is devoted to testing the second one.
An acceptability judgment experiment, of course, can only provide in-
direct evidence about sentence processing, and hence is less clearly rele-
vant to the subject of this special issue than, for example, a study which
compared reading times or the time taken to respond to a sentence. On

3. For ease of expositions, I have assumed that the source expression for the analogy is an
actual expression rather than the template; but this need not be the case.
4. Reliance on analogy and schema use are not as di¤erent as it might at first seem. As
Langacker points out, applying analogy requires the speaker to apprehend an abstract
commonality between the source and target forms; and the abstract commonality, of
course, is what would be captured by the schema (Langacker 2000: 60; see also Da˛brow-
ska 2008). Furthermore, repeated use of analogy will result in the ‘‘abstract commonal-
ity’’ being entrenched until it becomes a linguistic unit in its own right. The critical dif-
ference between the two processes, then, is whether the relevant knowledge is retrieved
from memory or created ‘‘on the fly’’.
Questions with long-distance dependencies 395

the plus side, an acceptability judgment study is easier to conduct (since it


doesn’t require any special apparatus) and hence it is a sensible first step
when investigating a syntactic construction. Furthermore, many linguists
would argue that it provides more useful evidence about the nature of
speakers’ underlying linguistic representations, or ‘‘competence’’ (cf.
Wasow and Arnold 2005), and hence, perhaps, will be less likely to be
dismissed as ‘‘mere performance’’.
Of course, judging the acceptability of a sentence is a type of perfor-
mance, and, like other types of performance, can be influenced by a vari-
ety of factors: plausibility, complexity, fatigue, mode of presentation, and
so on. This raises an obvious problem for an analyst trying to interpret
the results. The solution to the problem, however, is not—as some lin-
guists have suggested—to give up attempts to study speakers’ judgments
experimentally, but to control as many confounding factors as possible,
and be cautious in interpreting the results.

2. Experimental design
Da˛browska (2004) reports on a preliminary study showing that speakers
rate prototypical LDD questions such as Where do you think they sent the
documents? and What did you say the burglars stole? as more acceptable
than questions which had lexical subjects, a main verb other than think
or say, and an auxiliary other than do (e.g., Where will the customers
remember they sent the documents? What have the police revealed the
burglars stole?). There was no corresponding e¤ect for declaratives. It is
not clear, however, whether the di¤erence in speakers’ judgments was
due to the choice of subject, verb, or auxiliary, or some combination
of these factors. The experiment described in this paper was designed to
investigate how each of these three factors individually contributes to
speakers’ judgment. It will also examine two additional grammatical fac-
tors: the presence or absence of a complementizer and the number of
clauses intervening between the WH word and the gap, as well as the
e¤ects of plausibility and syntactic complexity.
In the experiment, native speakers of English completed a written ques-
tionnaire in which they were asked to rate the acceptability of LDD ques-
tions of varying degrees of prototypicality. There were seven experimental
conditions:
1. Prototypical LDD questions (WH Prototypical): These had the form
WH do you think S-GAP? or WH did you say S-GAP?;
2. LDD questions with lexical matrix subjects (WH Subject);
3. LDD questions with auxiliaries other than do (WH Auxiliary);
396 E. Da˛browska

4. LDD questions with matrix verbs other than think or say (WH Verb);
5. LDD questions with overt complementizers (WH Complementizer);
6. LDD questions with ‘‘very long’’ dependencies, i.e., with an addi-
tional complement clause (WH Long);
7. Unprototypical LDD questions (WH Unprototypical): These had a
lexical subject, an auxiliary other than do, and a main verb other than
think or say, an overt complementizer, and an additional complement
clause.
The questionnaire also contained two types of control sentences.
Grammatical controls were declarative versions of the LDD questions
constructed by replacing the WH word with a noun phrase or a preposi-
tional phrase (and adding a conjunction: see below). Ungrammatical
controls involved four types of structures: that trace violations (*That),
sentences involving a dependency reaching into a complex NP (*Com-
plexNP), negative sentences without do support (*Not), and negative sen-
tences with double tense marking (*DoubleTn). Examples of each type of
sentence are provided in Table 1; a complete list of all sentences used in
one version of the questionnaire is given in the Appendix.

3. Predictions
3.1. General rules
If speakers have the competence attributed to them by generative lin-
guists, and if their grammaticality judgments are a more or less direct re-
flection of this competence, then we could expect the following prediction
to hold:
Prediction 1: Grammatical sentences should receive ratings close to 5;
ungrammatical sentences should be rated about 1.

3.2. General rules þ processing and pragmatics


Prediction 1 is unrealistic, since it is well known that speakers’ judgments
are influenced by factors such as complexity and plausibility, just like any
other kind of performance. In particular, sentences involving filler-gap
dependencies are computationally more demanding since the filler must
be held in memory while the rest of the sentence is being processed (cf.
Frazier and Clifton 1989; Hawkins 1999; Kluender and Kutas 1993); and
sentences involving a dependency over more than one clause are particu-
larly di‰cult. Furthermore, since it is rather odd to assert what the ad-
dressee thinks or says (and perfectly natural to ask about these things),
we might expect an interaction between construction type and the lexical
Questions with long-distance dependencies 397

Table 1. Examples of sentences used in the experiment

Condition Example

1. WH Prototypical What do you think the witness will say if they don’t intervene?
2. WH Subject What does Claire think the witness will say if they don’t
intervene?
3. WH Auxiliary What would you think the witness will say if they don’t
intervene?
4. WH Verb What do you believe the witness will say if they don’t
intervene?
5. WH Complementizer What do you think that the witness will say if they don’t
intervene?
6. WH Long What do you think Jo believes he said at the court hearing?
7. WH Unprototypical What would Claire believe that Jo thinks he said at the court
hearing?
Grammatical controls
1. DE Prototypical But you think the witness will say something if they don’t
intervene.
2. DE Subject And Claire thinks the witness will say something if they don’t
intervene.
3. DE Auxiliary You would think the witness will say something if they don’t
intervene.
4. DE Verb So you believe the witness will say something if they don’t
intervene.
4. DE Complementizer So you think that the witness will say something if they don’t
intervene.
5. DE Long So you think Jo believes he said something at the court
hearing.
6. DE Unprototypical Claire would believe that Jo thinks he said something at the
court hearing.
Ungrammatical Controls
*That *What did you say that works even better?
*Complex NP *What did Claire make the claim that she read in a book?
*Not *Her husband not claimed they asked where we were going.
*DoubleTn *His cousin doesn’t thinks we lied because we were afraid.

properties of the matrix subject: specifically, speakers may assign low rat-
ings to declaratives with second person subjects, but accept the corre-
sponding interrogatives.
Taking processing demands and pragmatics into consideration, one
might make the following predictions:
Prediction 2: WH questions (WH) will receive lower ratings than the
corresponding declaratives (DE):
DE Protototypical > WH Prototypical
DE Subject > WH Subject
398 E. Da˛browska

DE Verb > WH Verb


DE Auxiliary > WH Auxiliary
DE Complementizer > WH Complementizer
DE Long > WH Long
DE Unprototypical > WH Unprototypical
Prediction 3: WH questions involving very long dependencies (WH
Long, WH Unprototypical) should receive lower ratings
than WH questions involving dependencies over just one
clause boundary (all other WH conditions). The corre-
sponding declaratives should be rated equally acceptable,
since they do not involve long distance dependencies.
WH Prototypical, WH Subject, WH Verb, WH Auxiliary,
WH Complementizer > WH Long, WH Unprototypical
Prediction 4: Declaratives with lexical subjects and complement-taking
verbs will receive higher ratings than declaratives with sec-
ond person subjects. There will be no corresponding e¤ect
for interrogatives.
DE Subject > DE Prototypical

3.3. Usage-based models


According to the usage-based hypothesis, speakers store lexically specific
templates corresponding to frequent combinations they have encountered
in their experience such as WH do you think S-GAP? and WH did you say
S-GAP? If this hypothesis is correct, prototypical LDD questions (i.e.,
those that fit one of these templates) should be rated as more acceptable
than non-prototypical questions. There should be no corresponding dif-
ferences in the acceptability of declaratives, or the relevant di¤erences
should be smaller.
It was also hypothesised that speakers construct non-prototypical LDD
questions on analogy with prototypical questions, by adding elements or
substituting a di¤erent item in a particular position in a template. Since
some elements may be more easily substitutable than others, di¤erent
substitutions may result in di¤erent degrees of acceptability. We know
from language acquisition research that children learn to substitute new
items into nominal slots relatively early; the ability to substitute verbs
into slots emerges later, and auxiliary substitutions are later still (Da˛b-
rowska and Lieven 2005; Lieven et al. 2003, Tomasello et al. 1997). The
most likely explanation for this finding is that nominals are autonomous
units which can be defined independently of the constructions they occur
Questions with long-distance dependencies 399

in, while verbs are non-autonomous in the sense that their descriptions
must make reference to the entities participating in the relationship they
designate—in other words, these entities are part of the verb’s profile (cf.
Langacker 1987: 298¤ ). Likewise, tensed auxiliaries, as grounding predi-
cations (cf. Langacker 1991: 193¤ ), conceptually presuppose the events
that they ground—designated by the verb plus its arguments. It follows
that nominals should be more easily substitutable than verbs, which in
turn should be more easily substitutable than auxiliaries.

Prediction 5: Prototypical LDD questions will receive higher ratings


than questions with lexical subjects, which will be more
acceptable than questions with matrix verbs other than
think or say; questions with auxiliaries other than do will
be judged least acceptable.
WH Prototypical > WH Subject > WH Verb > WH
Auxiliary

Finally, LDD questions containing overt complementizers and ‘‘very


long’’ dependencies are also less prototypical, in that both contain an ex-
tra element (the complementizer or an additional clause). It is not clear
whether inserting an extra element is more or less di‰cult than substitu-
tion; however, we can make the following predictions:

Prediction 6: LDD questions without overt complementizers will receive


higher ratings than questions with that:
WH Prototypical > WH Complementizer
Prediction 7: LDD questions with very long dependencies (i.e., depen-
dencies reaching over more than one intervening clause)
will be judged less acceptable than questions with shorter
dependencies, but more acceptable than unprototypical
questions (since the latter contain a very long dependency
as well as lexical substitutions).
WH Prototypical > WH Long > WH Unprototypical

4. Method
4.1. Stimuli
4.1.1. Experimental sentences. The experimental sentences were con-
structed by combining a ‘‘sentence stub’’ with a completion consisting of
either a complement clause and an adverbial clause or two complement
400 E. Da˛browska

clauses (see below). The stubs for the WH Prototypical condition were as
follows:
(4) a. What do you think . . .
b. Where do you think . . .
c. What did you say . . .
d. Where did you say . . .
The stubs for the non-prototypical sentences were constructed by
changing or adding the relevant element. Thus, in the WH Subject condi-
tion, you was changed to a proper name (and a third person ending was
added to the auxiliary so that it agreed with the subject); in the WH Aux-
iliary condition, do was changed to will or would; in the WH Verb condi-
tion, think was replaced with believe or suspect and say with claim or
swear; in the WH Complementizer condition, an overt complementizer
(that) was added after the verb; and in the WH Unprototypical condition,
all of the above changes were made.5
The completions for conditions 1–5 consisted of a four word comple-
ment clause followed by a four-word adverbial clause, e.g.,

(5) (What do you think) the witness will say


(stub) (first complement clause)
if they don’t intervene?
(adverbial clause)

The completions for sentences with very long dependencies (i.e., condi-
tions 6 and 7) consisted of a two-word complement clause followed by a
pronominal subject, verb, and a four-word prepositional phrase, e.g.,
(6) (What do you think) Jo believes
(stub) (first complement clause)
he said at the court hearing?
(second complement clause)
Thus, all experimental sentences without complementizers were 12
words long and contained three clauses, with seven words intervening be-
tween the WH word and the gap. Sentences with complementizers were
13 words long, with 8 words between the WH word and the gap.
There were two versions of the questionnaire, and two sets of comple-
tions. In version 1, the stubs were combined with completion set 1 in con-

5. Throughout this paper, I use the term non-prototypical to refer to questions which di¤er
in some respect from the prototypical instances of the construction, and the term unpro-
totypical to refer to questions which di¤er from the prototype in all relevant respects.
Questions with long-distance dependencies 401

ditions 1, 3, 5 and 7 and with completion set 2 in conditions 2, 4, and 6;


in version 2, the stubs were combined with completion set 1 in even-
numbered conditions and with completion set 2 in odd-numbered
conditions.

4.1.2. Grammatical controls. The grammatical control conditions were


constructed by supplying appropriate lexical material for the WH word in
the appropriate position in the clause; in addition, a conjunction (so, and,
or but) was added at the beginning of declaratives corresponding to inter-
rogatives with the auxiliary do: for example, the declarative counterpart
of
(7) What do you think the witness will say if they don’t intervene?
was
(8) But you think the witness will say something if they don’t intervene.
This was done so that the interrogative sentence and the corresponding
declarative control contained the same number of words; it also made the
declarative sentences sound more natural. (It is somewhat odd for a
speaker to assert what the addressee thinks or said; adding the conjunc-
tion makes the sentence pragmatically more plausible because it conveys
the impression that the speaker is either inferring the addressee’s beliefs
from his or her words, contrasting them with those of another person, or
clearing up an apparent misunderstanding).

4.1.3. Ungrammatical controls. The ungrammatical control conditions,


like the experimental sentences and the grammatical controls, contained
subordinate clauses. Half of the ungrammatical sentences were declara-
tive and the other half interrogative. That-trace sentences (*That) con-
tained an overt complementizer immediately before a gap in the subject
position:
(9) *What do you think that probably got lost during the move?
(10) *Who do you think that will turn up in the evening?
In Complex NP sentences (*ComplexNP) the matrix clause contained a
complement-taking noun (claim, fact, rumour, hypothesis) followed by a
complement clause with a gap, e.g.,
(11) *What did Claire make the claim that she read in a book?
(12) *Where did you discover the fact that the criminals put the car ?
In Negatives without do support (*Not), the matrix clause contained a
negated verb but no auxiliary, with tense being marked on the main verb:
402 E. Da˛browska

(13) *Her husband not claimed they asked where we were going.

Finally, in declaratives with double tense/agreement marking (*Dou-


bleTn), the matrix clause contained a third person subject and a negated
verb, and agreement was marked on the auxiliary as well as on the main
verb:

(14) *The girl doesn’t remembers where she spent her summer holidays.
The same ungrammatical controls were used in both versions of the
questionnaire. There were four sentences in each condition, giving a total
of 16 ungrammatical controls.

4.1.4. Constructing the questionnaire. The test sentences were divided


into four blocks, each containing one sentence from each of the seven
experimental and eleven control conditions. The order of the sentences
within each block was random.
A full list of the sentences used in version 1 of the experiment is given
in the Appendix.

4.2. Participants
38 second and third year literature students from the School of English at
the University of Newcastle participated in the experiment. All were na-
tive speakers of English.

4.2.1. Procedure. Participants were asked to complete a written ques-


tionnaire and were given the following instructions:
The questionnaire is part of a study of speakers’ intuitions about En-
glish sentences. It is not an intelligence test or a grammar test.
Please indicate how acceptable/unacceptable you find each of the fol-
lowing sentences by choosing a number on a scale from 1 (very bad)
to 5 (fine). Read the sentences carefully, but do not spend too much
time thinking about them: we are interested in your initial reaction.
Do not go back and change your responses to earlier sentences.
The instructions were followed by two examples for which ratings were
provided:
(15) Will the girl who won the prize come to the party?
(16) Did the man who arrive by train is my cousin?
The first was given a rating of 5 and the second 1. These examples were
provided in order to anchor the participants’ ratings. Thus, in essence, the
Questions with long-distance dependencies 403

Table 2. Acceptability ratings for all conditions

Condition Mean Std. Deviation

WH Prototypical 4.31 0.63


WH Subject 4.25 0.59
WH Verb 3.93 0.71
WH Auxiliary 3.23 0.83
WH Complementizer 3.84 0.84
WH Long 3.85 0.76
WH Unprototypical 2.54 0.75
DE Prototypical 3.57 0.85
DE Subject 4.00 0.63
DE Verb 3.74 0.78
DE Auxiliary 3.49 0.66
DE Complementizer 3.53 0.79
DE Long 3.89 0.75
DE Unprototypical 3.14 0.90
*That 2.50 0.75
*DoubleTn 2.41 0.95
*ComplexNP 1.69 0.56
*Not 1.31 0.49

participants’ task was to decide whether the sentences in the questionnaire


were more like 1, more like 5, or in-between.
The questionnaires were distributed after a lecture and took 10–15
minutes to complete. Participants were randomly assigned to one version
of the questionnaire, with about half completing each version.

5. Results and discussion


5.1. Grammatical v. ungrammatical sentences
The mean acceptability ratings for all conditions are given in Table 2. For
clarity, the same information is presented visually in Figure 1, where the
bars corresponding to each experimental condition have been arranged
from highest to lowest. Although the mean rating for all grammatical
sentences combined (3.67) was considerably higher than for the ungram-
matical controls (1.98), there is no sharp contrast between grammati-
cal and ungrammatical sentences. What we have instead is a continuum
of acceptability ratings ranging from 4.3 for prototypical WH ques-
tions to 1.3 for negatives without do support, with the other sentence
types occupying various intermediate points. The four ungrammati-
cal sentence types cluster at the lower end of the continuum; however,
the acceptability ratings for unprototypical LDD questions were not
404 E. Da˛browska

Figure 1. Mean acceptability ratings for all conditions


Note: Grey bars correspond to questions; white bars correspond to declaratives
and striped bars to ungrammatical controls. Error bars represent 95 percent confi-
dence intervals.

significantly di¤erent from those for that-trace and double-tense sentences


(WH Unprototypical v. *that: tð37Þ ¼ 0:26, p ¼ 0:798; WH Unprototyp-
ical v. *DoubleTn: tð37Þ ¼ 0:70, p ¼ 0:486), although they were higher
than those for the other two ungrammatical control conditions (WH Un-
prototypical v. *Complex NP: tð37Þ ¼ 6:15, p < 0:001; WH Unprototyp-
ical v. *Not: tð37Þ ¼ 8:81, p < 0:001). Thus, the pure competence gram-
mar prediction that grammatical sentences will receive ratings close to 5
and ungrammatical sentences will be rated about 1 is clearly false.

5.2. Grammatical sentences


A preliminary analysis of the participants’ ratings for the grammatical
sentences was conducted using a construction (2)  prototypicality
(7)  version (2) ANOVA. The analysis revealed a significant main
e¤ect of prototypicality, F ð6; 216Þ ¼ 49:82, p < 0:001, hp2 ¼ 0:58, and a
prototypicality  construction interaction, F ð6; 216Þ ¼ 15:14, p < 0:001,
hp2 ¼ 0:30. No other e¤ect or interaction was significant. Since there was
no significant e¤ect of version, and no interaction between version and
any of the other factors, the results for the two versions were collapsed
in all further analyses.
Questions with long-distance dependencies 405

Table 3. Testing prediction 2

Prediction Mean SD t-test p value Prediction


value confirmed?

WH Prototypical 4.31 0.63 5.234 <0.001 77


<DE Prototypical 3.57 0.85
WH Subject 4.25 0.59 2.859 0.007 77
<DE Subject 4.00 0.63
WH Complementizer 3.84 0.84 2.149 0.038 77
<DE Complementizer 3.53 0.79
WH Verb 3.93 0.71 1.644 0.109 7
<DE Verb 3.74 0.78
WH Long 3.85 0.76 0.452 0.654 7
<DE Long 3.89 0.75
WH Auxiliary 3.23 0.83 1.950 0.059 7
<DE Auxiliary 3.49 0.66
WH Unprototypical 2.54 0.75 5.882 <0.001 3
<DE Unprototypical 3.14 0.90

Note: 3 indicates that a prediction has been confirmed; 7 indicates that a prediction has
not been confirmed; 77indicates a significant di¤erence in the opposite direction.

5.2.1. Processing cost of dependencies: Questions v. declaratives. Pre-


diction 2 was that LDD questions will receive lower ratings than the cor-
responding declaratives because they involve displaced constituents. This
prediction was not confirmed: there was no significant e¤ect of construc-
tion. Instead, as indicated above and shown in Table 3, we have an inter-
action between construction type and lexical content, with prototypical
LDD questions, questions with lexical subjects, and questions with overt
complementizers being judged significantly more acceptable than the cor-
responding declaratives. (Note that the significance levels reported in
Table 3 and elsewhere in this paper have not been corrected for multiple
comparisons: since the hypothesis tested predicts that all the relevant
comparisons should be significant, using the Bonferroni adjustment or
an equivalent method would not be appropriate.)
The interrogative sentences, particularly in the WH Prototypical condi-
tion, contained some frequent bigrams (what do, do you, you think): these
occur with a frequency of 6433, 27602, and 9901 respectively in the Brit-
ish National Corpus. This could be partly responsible for the fact that
interrogatives were rated as more acceptable than the corresponding
declaratives in most conditions. However, as shown in Table 4, there is
no strong relationship between the number of frequent bigrams and ac-
ceptability ratings of WH questions (i.e., questions containing more
high-frequency bigrams are not necessarily more acceptable) or the num-
406 E. Da˛browska

Table 4. Frequent bigrams in WH questions

Condition Mean Di¤erence Frequent


score* bigrams

WH Prototypical 4.31 0.74 what do


do you
you think
WH Complementizer 3.84 0.31 what do
do you
you think
WH Subject 4.25 0.25 what do
WH Verb 3.93 0.19 what do
do you
WH Long 3.85 0.04 what do
do you
you think
WH Auxiliary 3.23 0.26 you think
WH Unprototypical 2.54 0.60

* Di¤erence scores were computed by subtracting the rating of the declarative control from
the rating of the interrogative sentence.

ber of frequent bigrams and the advantage for questions over declaratives
(questions containing more high-frequency bigrams are not necessarily
better than the corresponding declaratives). Furthermore, bigram fre-
quency cannot explain the interaction between construction type and
verb or between construction type and complementizer (see below), since
the words immediately preceding and following the verb and the comple-
mentizer were the same in both the declarative and the interrogative var-
iants. In fact, for sentences with complementizers, bigram frequency
makes precisely the wrong predictions. In the WH Prototypical and DE
Prototypical conditions, the main clause verb (think or say) was followed
by the pronoun they or we or the determiner the, while in the WH Com-
plementizer and DE Complementizer conditions, it was followed by the
complementizer that. The mean frequency of the bigrams think the, think
they, think we, say the, say they, say we in the British National Corpus is
2539, while the mean frequency of the bigrams think that and say that is
9473. Thus, if acceptability ratings were simply a reflection of bigram fre-
quency, sentences with complementizers should receive higher ratings
than sentences without them. (The mean frequency of the bigrams that
they, that the and that we is even higher: 59333.) However, as we will see
in section 5.2.4, the ratings for WH questions with complementizers were
Questions with long-distance dependencies 407

considerably lower than for the prototypical variants, while there was no
di¤erence in the acceptability of the corresponding declaratives.

5.2.2. Processing cost of ‘‘very long’’ dependencies. According to pre-


diction 3, WH questions containing ‘‘very long’’ dependencies, i.e., de-
pendencies spanning two clause boundaries, should receive lower ratings
than questions containing dependencies across just one clause boundary,
whilst the corresponding declaratives should be rated equally acceptable,
since they do not involve a filler-gap dependency. This prediction was
evaluated by comparing the mean ratings for questions with ‘‘very long’’
dependencies (WH Long and WH Unprototypical) and the correspond-
ing declaratives and for questions containing dependencies over one
clause boundary (WH Prototypical, WH Subject, WH Verb, WH Auxil-
iary, WH Complementizer) and the corresponding declaratives. The rat-
ings were analyzed using a 2  2 ANOVA with the within-participants
factors of construction type (WH question, declarative) and complemen-
tation (1 clause, 2 clauses). This revealed a main e¤ect of complementa-
tion, F ð1; 37Þ ¼ 50:44, p < 0:001, hp2 ¼ 0:58, indicating that sentences
containing two complement clauses were judged as less acceptable than
sentences containing one complement clause and one adverbial clause.
This was qualified by a construction  complementation interaction,
F ð1; 37Þ ¼ 34:14, p < 0:001, hp2 ¼ 0:48 (see Figure 2). Post-hoc compari-
sons showed that WH questions with very long dependencies were judged
significantly worse than questions with long dependencies: tð37Þ ¼ 9:62,

Figure 2. Construction type  complementation interaction (All conditions)


408 E. Da˛browska

Table 5. Testing prediction 5 (Questions)

Prediction Mean SD t-test p value Prediction


value confirmed?

WH Prototypical 4.31 0.63 4.67 <0.001 3


>WH Long 3.85 0.76
WH Subject 4.25 0.59 3.63 0.001 3
>WH Long 3.85 0.76
WH Verb 3.93 0.71 0.62 0.539 7
>WH Long 3.85 0.76
WH Auxiliary 3.23 0.83 4.13 <0.001 77
>WH Long 3.85 0.76
WH Complementizer 3.84 0.84 0.11 0.910 7
>WH Long 3.85 0.76
WH Prototypical 4.31 0.63 14.17 <0.001 3
>WH Unprototypical 2.54 0.75
WH Subject 4.25 0.59 15.36 <0.001 3
>WH Unprototypical 2.54 0.75
WH Verb 3.93 0.71 11.29 <0.001 3
>WH Unprototypical 2.54 0.75
WH Auxiliary 3.23 0.83 5.73 <0.001 3
>WH Unprototypical 2.54 0.75
WH Complementizer 3.84 0.84 9.38 <0.001 3
>WH Unprototypical 2.54 0.75

p < 0:001. For declaratives, the corresponding di¤erence approaches sig-


nificance: tð37Þ ¼ 1:87, p ¼ 0:070. This suggests that processing di‰culty
had the predicted e¤ect on acceptability ratings.
However, comparisons of ratings for each of two ‘‘very long’’ depen-
dency questions with each of the ‘‘long’’ dependency conditions suggests
a slightly more complex picture. As shown in Table 5, WH Long items
were judged significantly worse than WH Prototypical and WH Subject
items, and unprototypical questions were judged significantly worse than
all other WH questions. However, there was no significant di¤erence be-
tween WH Long and WH Verb items or between WH Long and WH
Complementizer (indeed, the latter received slightly higher ratings), and
questions with a modal auxiliary were judged significantly worse than
WH Long sentences. Since the declarative versions of unprototypical
LDD questions were also judged to be less acceptable than the other de-
clarative sentences (see Table 6), the low acceptability ratings for the un-
prototypical sentences may be partially due to the lexical content of the
sentences. Thus, although the existence of very long dependencies may
have an e¤ect on acceptability, this can only account for some of the ob-
served di¤erences.
Questions with long-distance dependencies 409

Table 6. Testing prediction 3 (Declaratives)

Prediction Mean SD t-test p value Same as


value question?

DE Prototypical 3.57 0.85 3.10 0.004 7


DE Long 3.89 0.75
DE Subject 4.00 0.63 1.22 0.230 7
DE Long 3.89 0.75
DE Verb 3.74 0.78 1.45 0.156 3
DE Long 3.89 0.75
DE Auxiliary 3.49 0.66 3.89 <0.001 3
DE Long 3.89 0.75
DE Complementizer 3.53 0.79 3.68 0.001 7
DE Long 3.89 0.75
DE Prototypical 3.57 0.85 3.22 0.003 3
DE Unprototypical 3.14 0.90
DE Subject 4.00 0.63 6.74 <0.001 3
DE Unprototypical 3.14 0.90
DE Verb 3.74 0.78 4.31 <0.001 3
DE Unprototypical 3.14 0.90
DE Auxiliary 3.49 0.66 2.99 0.005 3
DE Unprototypical 3.14 0.90
DE Complementizer 3.53 0.79 3.17 0.003 3
DE Unprototypical 3.14 0.90

5.2.3. Pragmatics. Prediction 4 stated there should be an interaction


between construction type and the lexical properties of the matrix subject:
declaratives with lexical subjects should receive higher ratings than de-
claratives with second person subjects, but there should be no correspond-
ing di¤erence for questions. To test this prediction, a 2  2 ANOVA
with the within-participants factors of construction (WH question, de-
clarative) and subject (second person, lexical). The analysis showed that
the predicted interaction did indeed occur: F ð1; 37Þ ¼ 14:82, p < 0:001,
hp2 ¼ 0:29 (see Figure 3); the main e¤ect of construction was also signifi-
cant, F ð1; 37Þ ¼ 25:01, p < 0:001, hp2 ¼ 0:40. Further analysis confirmed
that declaratives with lexical subjects were judged more acceptable than
declaratives with second person subjects: tð37Þ ¼ 3:92, p < 0:001. The
ratings for interrogatives with lexical subjects were marginally lower
than for interrogatives with second person subject, but the di¤erence was
not statistically significant: tð37Þ ¼ 0:91, p ¼ 0:368.

5.2.4. Prototypicality e¤ects. According to the usage-based hypo-


thesis, any modification of the LDD template should result in lower
410 E. Da˛browska

Figure 3. Construction type  subject interaction

acceptability ratings for interrogative sentences but have no e¤ect (or a


much smaller e¤ect) on declaratives. In other words, usage-based theories
predict an interaction between construction type and lexical content.
As we have just seen, although there is a significant interaction between
construction type and the lexical properties of the subject, this is best in-
terpreted as reflecting pragmatic implausibility of second person declara-
tives with think and say. Substituting a lexical subject for you in questions
did not result in a significant reduction of acceptability, although there
was a small di¤erence in the predicted direction. This suggests that the
LDD template does not specify the subject—although it is also possible
that the processing cost of NP substitution is too small to be revealed by
an acceptability judgment task.
The relationship between construction type and the other four factors
(verb, auxiliary, complementizer, and the number of complement clauses)
was investigated by means of four additional 2  2 ANOVAs (see
Table 7) followed up by t-tests. All the interactions were as predicted
by the usage-based account. There was a significant interaction between
construction type and verb (see Figure 4): changing the matrix verb
in a LDD question results in significantly lower acceptability ratings
(tð37Þ ¼ 3:23, p ¼ 0:003); changing the verb in the corresponding declar-
ative, on the other hand, results in a slightly more acceptable sentence,
although the di¤erence is not statistically significant (tð37Þ ¼ 1:45,
p ¼ 0:155). There was also an interaction between construction type and
auxiliary (see Figure 5): replacing do or did with the modal auxiliary will
Questions with long-distance dependencies 411

Table 7. ANOVA results

E¤ect/Interaction Test statistics

Construction (WH, DE)  verb (think/say, believe/suspect/claim/swear)


Construction F ð1; 37Þ ¼ 8:28, p < 0:001, hp2 ¼ 0:34
Construction  verb F ð1; 37Þ ¼ 13:81, p ¼ 0:002, hp2 ¼ 0:27
Construction (WH, DE)  auxiliary (do/did, will/would )
Construction F ð1; 37Þ ¼ 7:30, p < 0:001, hp2 ¼ 0:17
Auxiliary F ð1; 37Þ ¼ 47:06, p < 0:001, hp2 ¼ 0:56
Construction  auxiliary F ð1; 37Þ ¼ 22:06, p < 0:001, hp2 ¼ 0:37
Construction (WH, DE)  complementizer (none, that)
Construction F ð1; 37Þ ¼ 21:15, p < 0:001, hp2 ¼ 0:36
Complementizer F ð1; 37Þ ¼ 13:29, p ¼ 0:001, hp2 ¼ 0:26
Construction  complementizer F ð1; 37Þ ¼ 6:68, p ¼ 0:014, hp2 ¼ 0:15
Construction (WH, DE)  complementation (1 clause, 2 clauses)
Construction F ð1; 37Þ ¼ 12:59, p ¼ 0:001, hp2 ¼ 0:25
Construction  complementation F ð1; 37Þ ¼ 5:82, p < 0:001, hp2 ¼ 0:42

Note: Only significant main e¤ects and interactions are listed in the table.

Figure 4. Construction type  verb interaction

or would made the interrogatives less acceptable (tð37Þ ¼ 8:26,


p < 0:001), while adding the same auxiliaries in declaratives had no e¤ect
on ratings (tð37Þ ¼ 0:56, p ¼ 0:578). The size of the interaction between
construction type and complementizer (Figure 6) was somewhat smaller,
although it was also in the predicted direction: adding an overt com-
plementizer had no e¤ect on declaratives (tð37Þ ¼ 0:44, p ¼ 0:666) but
412 E. Da˛browska

Figure 5. Construction type  auxiliary interaction

resulted in lower ratings for interrogatives (tð37Þ ¼ 3:77, p ¼ 0:001).6


Finally, there is a significant interaction between construction type and
complementation: adding a second complement clause between the filler
and the gap makes the interrogative less acceptable while the correspond-
ing declarative is more acceptable (see Figure 7; for pairwise compari-
sons, see Table 5). (Note that this analysis compares just the WH Proto-
typical and WH Long conditions and their declarative counterparts, i.e.,
sentences which are most closely matched lexically. The analysis in sec-
tion 5.2.2 contrasted questions with ‘‘very long’’ dependencies, i.e., WH
Long and WH Unprototypical, with questions containing dependencies
spanning just one clause, i.e., WH Prototypical, WH Subject, WH Verb,
WH Auxiliary, and WH Complementizer.)
The usage-based model adopted here also predicts a particular order in
the acceptability of non-prototypical LDD questions: specifically, LDD
questions with a lexical subject should be more acceptable than questions
with a non-prototypical matrix verb, which in turn should be more

6. We know from the psycholinguistic literature that overt complementizers facilitate pro-
cessing, presumably because they signal the presence of a subordinate clause and thus
help the processor to avoid garden path e¤ects: for instance, sentences with complement
clauses introduced by a complementizer are read faster than sentences without comple-
mentizers, even when the main clause verb has a strong preference for clausal comple-
ments (Trueswell et al. 1993, Holmes et al. 1989). Thus the presence of the complemen-
tizer e¤ect for questions provides strong evidence in favour of lexical storage of the
whole construction.
Questions with long-distance dependencies 413

Figure 6. Construction type  complementizer interaction

Figure 7. Construction type  complementation interaction (‘‘Prototypical’’ and ‘‘Long’’


sentences only)

acceptable than those with a non-prototypical auxiliary. As shown in


Tables 8 and 9, these predictions have also been confirmed. (Table 8
also shows the results of pairwise comparisons relevant for testing other
usage-based predictions for the sake of completeness.)
Although the model made no specific predictions about the relative size
of the e¤ect of the other manipulations, it is interesting to see how they
compare to lexical substitutions in the subject, verb and auxiliary slot.
414 E. Da˛browska

Table 8. Testing predictions 5, 6 and 7 (Questions)

Prediction Mean SD t-test p value Prediction


value confirmed?

WH Prototypical 4.31 0.63 0.91 0.368 7


>WH Subject 4.25 0.59
WH Subject 4.25 0.59 3.32 0.002 3
>WH Verb 3.93 0.71
WH Verb 3.93 0.71 4.96 <0.001 3
>WH Auxiliary 3.23 0.83
WH Prototypical 4.31 0.63 3.77 0.001 3
>WH Complementizer 3.84 0.84
WH Prototypical 4.31 0.63 4.69 <0.001 3
>WH Long 3.85 0.76
WH Long 3.85 0.76 10.51 <0.001 3
>WH Unprototypical 2.54 0.75

Table 9. Testing predictions 5, 6 and 7 (Declaratives)

Prediction Mean SD t-test p value Same as


value question?

DE Prototypical 3.57 0.85 3.92 <0.001 7


DE Subject 4.00 0.63
DE Subject 4.00 0.63 2.85 0.007 3
DE Verb 3.74 0.78
DE Verb 3.74 0.78 2.26 0.030 3
DE Auxiliary 3.49 0.66
DE Prototypical 3.57 0.85 0.44 0.666 7
DE Complementizer 3.53 0.79
DE Prototypical 3.57 0.85 3.10 0.004 7
DE Long 3.89 0.75
DE Long 3.89 0.75 8.44 <0.001 3
DE Unprototypical 3.14 0.90

As can be seen from Figure 1, the e¤ects of adding a complementizer and


of adding an additional complement clause are about the same as that of
changing the matrix verb. The mean acceptability ratings for WH Verb,
WH Long and WH Complementizer sentences were 3.93, 3.85, and 3.84
respectively; none of the di¤erences between these conditions was statisti-
cally significant (WH Verb v. WH Complementizer: tð37Þ ¼ 0:71,
p ¼ 0:484; WH Verb v. WH Long: tð37Þ ¼ 0:62, p ¼ 0:539; WH Com-
plementizer v. WH Long: tð37Þ ¼ 0:11, p ¼ 0:910).
Questions with long-distance dependencies 415

5.3. Why are ‘‘real life’’ LDD questions so stereotypical?

Why are questions with long distance dependencies so stereotypical? One


possible explanation is o¤ered by Verhagen (2005). Verhagen observes
that the propositional content of most complementation constructions is
expressed by the subordinate clause; the main clause normally just signals
epistemic stance (see also Thompson 2002), i.e., it invites the hearer to
adopt a particular subjective perspective on the object of conceptualiza-
tion. The greater the ‘‘distance’’ between the onstage conceptualizer (i.e.,
the subject of the main clause) and the ground (in the sense of Langacker
1987), the more di‰cult it is to construe the main clause as an epistemic
marker (as opposed to a prediction in its own right). Verhagen argues
that this distance is minimal when the conceptualizer is the first person
(in declaratives) or the second person (in interrogatives), when the verb
is relatively generic, and when there are no other elements qualifying the
verb; it follows that the matrix clause in LDD questions will normally
contain a second person subject, a relatively non-specific verb such as
think and say, and no additional constituents.
A di¤erent, but not necessarily incompatible, explanation for restric-
tions on questions and other constructions with long distance dependen-
cies is proposed by Goldberg (2006). Goldberg argues that di¤erences in
acceptability arising as a result of the use of di¤erent main clause verbs
can be explained by appealing to a general principle which she calls BCI,
which states that backgrounded constituents are islands. The gap in a
filler-gap dependency construction must occur within the ‘‘potential focus
domain’’; the constituent containing the gap cannot be backgrounded.
Since complements of factive verbs and manner of speaking verbs (as
well as complex NPs, sentential subjects, and presupposed adjuncts) are
backgrounded, they cannot participate in filler-gap constructions.
An experimental study by Ambridge and Goldberg (this issue) provides
some empirical support for this proposal. Participants in this experiment
completed two tasks. In the first task, they rated the acceptability of WH
questions with long distance dependencies (e.g., What did Jess think that
Dan liked?) and the corresponding declaratives (e.g., Daniele thought that
Jason liked the cake). In the second task they were presented with a ne-
gated sentence containing a verb complement clause (e.g., Maria didn’t
know that Ian liked the cake) and asked to judge to what extent it implied
the negation of the complement clause (Ian didn’t like the cake): this mea-
sured the extent to which speakers judged the information in the subordi-
nate clause to be presupposed, and thus backgrounded. The main finding
was that, as predicted by the BCI hypothesis, there was a very strong neg-
ative correlation between responses on the negation test and ‘‘di¤erence
416 E. Da˛browska

scores’’ computed by subtracting the acceptability rating of the questions


from those of the corresponding declaratives, and a weaker negative cor-
relation between responses on the negation test and acceptability ratings.
It remains to be seen whether BCI can also explain other restrictions
on LDD questions documented in this study: the fact that they strongly
disprefer main clause verbs other than think or say (not just factives
and manner-of-speaking verbs), auxiliaries other than do, and comple-
mentizers, and that in real life (as opposed to the examples found in the
linguistic literature) they virtually never involve a dependency spanning
more than one clause.
Thus we have two independent proposals explaining why particular
lexical variants of LDD questions may be preferred or dispreferred in us-
age. A central claim of usage-based approaches is that mental grammars
are shaped by usage patterns: it is thus not surprising that speakers de-
velop strong lexically specific templates for LDD questions and possibly
fail to develop more abstract representations of these constructions.

6. Conclusion
The most striking result of the experiment reported here is the existence of
strong prototypicality e¤ects for LDD questions. Prototypical instances
of this construction, i.e., those which fit one of the templates postulated
on the basis of corpus research (WH do you think S-GAP?, WH did you
say S-GAP?) were judged to be the most acceptable of all sentences. De-
partures from the prototype (use of a di¤erent auxiliary or verb in the
matrix clause, addition of a complementizer or an extra complement
clause) resulted in lower acceptability ratings. Crucially, there was no cor-
responding e¤ect on declaratives, so the di¤erences in grammaticality
cannot be attributed simply to the properties of the lexical items used in
the experiment. Acceptability also depended on the type of substitution
required: nominals are apparently easier to substitute than verbs, which
in turn were easier than auxiliaries.
The participants’ judgments were also influenced by pragmatic consid-
erations: declaratives with lexical subjects (DE Subject, e.g., So Steve said
the children could stay here when their father returns) were judged to be
more acceptable than declaratives with second person subjects (DE Pro-
totypical, e.g., So you said the children could stay here when their father
returns). This e¤ect, however, was fairly small in comparison with the
purely lexical e¤ects.
Adding an additional complement clause also reduced the acceptability
of the questions (and had the opposite e¤ect on declaratives). This could
be attributed to the greater processing demands posed by the increased
Questions with long-distance dependencies 417

syntactic distance between the filler and the gap, since the filler must be
held in working memory while the pre-gap part of the sentence is being
processed. However, questions with very long dependencies (with two
clause boundaries intervening between the filler and the gap) were not
consistently judged to be less acceptable than questions involving depen-
dencies across only one clause boundary; and the e¤ect of adding an ad-
ditional complement clause was no bigger than that of adding a comple-
mentizer or changing the matrix verb. Thus, appealing to prototypicality
e¤ects provides a more parsimonious explanation for these findings. This
is not to say that the processing demands of holding the filler in memory
have no e¤ect on processing—but the costs may be relatively small com-
pared the e¤ects of prototypicality.7
Interestingly, unprototypical LDD questions—those with a comple-
mentizer, an additional verb complement clause, a lexical subject, a
modal auxiliary, and a verb other than think or say in the main clause—
were judged to be just as bad as that-trace violations and sentences with
double tense/agreement marking (though better than sentences involving
‘‘extraction’’ out of a complex NP and negatives without an auxiliary).
This is consonant with the results of two acceptability experiments con-
ducted by Kluender and Kutas (1993). In their first experiment, in which
participants were required to provided speeded categorical acceptability
judgments, LDD questions were accepted only 54 percent of the time. In
the second experiment, participants rated acceptability on a scale from 1
to 40, and could take as much time as they wished to make the judgment.
The mean acceptability rating for LDD questions was 19—significantly
higher than that for WH island variations, but much lower than those
for Y/N questions containing complement clauses. Kluender and Kutas
conclude that the low acceptability of LDD questions is attributable to
the processing demands of holding the filler in working memory. How-
ever, since their stimuli were quite di¤erent from prototypical LDD ques-
tions (they all contained overt complementizers and non-prototypical
verbs; some also had lexical subjects or auxiliaries other than do), it could
also be explained by appealing to their unprototypicality.
So what does the presence of prototypicality e¤ects in acceptability
judgments about LDD questions—in particular, those due to the lexical

7. Acceptability judgment and ease of processing are of course two di¤erent things, al-
though they tend to be correlated: other things being equal, sentences which are di‰cult
to process tend to be judged less acceptable (cf. Fanselow and Frisch 2006; Frazier and
Clifton 1989; Kluender and Kutas 1993). Note, too, that unprototypical LDD questions
contain more dysfluencies than prototypical instances of the construction (Da˛browska
forthc.), which also suggests that they are more di‰cult to process.
418 E. Da˛browska

properties of the sentences, since these are easier to interpret—tell us


about speakers’ mental representations of this construction? It has long
been acknowledged, of course, that the choice of lexical items in a sen-
tence a¤ects speakers’ acceptability judgments. In the generative tradi-
tion, this is usually regarded as a confound: the presence of a particular
word can make an otherwise well-formed sentence unacceptable, which
could lead the analyst to draw incorrect conclusions about grammar; lex-
ical e¤ects, therefore, are something that should be ‘‘controlled for where
possible, discounted when encountered’’ (Featherston 2005: 702). In this
case, however, we are dealing with the opposite situation: LDD questions
are fully acceptable only with particular lexical content. This suggests
that they are more like a constructional idiom than a fully general con-
struction. In other words, questions with long-distance dependencies are
conventional units with an unusual form (a WH word at the beginning
of the main clause associated with a gap in a subordinate clause) and a
specialized meaning: the unit WH do you think S-GAP? is used to inquire
about the speakers’ opinion about the content of the subordinate clause
(What do you think he wants Q ‘‘In your opinion, what does he want?’’);
and WH did you say S-GAP? is used when addressee already gave the
speaker the relevant information but the speaker does not remember
(When did you say he came? Q ‘‘You’ve already told me when he came,
but please tell me again’’.) 8 The non-prototypical uses are rather like
what Moon (1998) calls ‘‘exploitations’’ of idioms exemplified by expres-
sions such as throw in the moist towelette (constructed on analogy with
throw in the towel ) or use an earthmover to crack a nut (cf. use a sledge-
hammer to crack a nut). Once all the lexical content has been changed
(cf. the WH Unprototypical condition), it is no longer possible to identify
the motivating construction, and hence the sentence is judged as un-
acceptable.9 Alternatively, one could argue that when processing unpro-
totypical LDD questions, speakers have to fall back on more abstract
schemas (the mental analogues of the WH question construction and the
complementation construction), and that the low acceptability ratings for

8. It is interesting to note that the CollinsCobuild English Language Dictionary (Sinclair


1987: 1519) lists the use of think in long-distance questions as a separate sense of the
verb.
9. With most idioms, substituting di¤erent lexical items for every word would result in ex-
pressions which are still judged acceptable, provided they make sense semantically: for
example, taking use a sledgehammer to crack a nut as the model, one could construct
perfectly acceptable phrases such as do a headstand to impress a neighbour and tickle an
earthworm to amuse the children. This is possible because speakers have fully general
schemas for the transitive and the infinitival construction.
Questions with long-distance dependencies 419

such sentences reflect the di‰culty of combining highly complex and


abstract schemas.
At this point one may wonder why the possibility that speakers have
lexically specific knowledge about LDD questions has not even been con-
sidered by most syntacticians in spite of the fact that these constructions
have been so intensively studied for several decades. The short answer is
that the possible existence of lexical e¤ects was not regarded as theoreti-
cally interesting—so although there are a number of experimental studies
of LDD constructions (see, for example, Cowart 1997; Frazier and Clif-
ton 1989; Kluender and Kutas 1993), to my knowledge, nobody has sys-
tematically investigated the e¤ect of lexical content on speakers’ linguistic
intuitions about them.10 A second reason is that linguists tend to rely on
their own intuitions—and linguists’ intuitions about LDD questions may
be systematically di¤erent from those of ordinary speakers. Da˛browska
(in press b) shows that linguists tend to judge unprototypical LDD ques-
tions as considerably more acceptable than that trace violations and sen-
tences with double tense marking, and not much worse than prototypical
questions. This could be a reflection of their theoretical commitments (the
belief that instances of ‘‘the same’’ construction should be equally gram-
matical), but it could also be a result of di¤erences in linguistic experi-
ence. Many linguists spend a considerable amount of time constructing
examples of the structures they are interested in and reading papers con-
taining such constructed examples.11 Since LDD questions have been the
object of very intensive research, it is likely that linguists (or at least lin-
guists who work on LDD constructions, or discuss them with their stu-
dents) have been exposed to more instances of this construction than
most ordinary language users, and, crucially, the instances they have en-
countered are much more varied, as demonstrated by the examples in (1).

10. There is some work on lexical e¤ects on acceptability judgments in basic argument
structure constructions: see Theakston 2004, Ambridge et al. in press. In both studies,
speakers judged argument structure violations with high frequency verbs (e.g., *I
poured you with water) as less acceptable than argument structure violations with low
frequency verbs (e.g., *I dribbled teddy with water); the authors explain this by appeal-
ing to the higher entrenchment of the pattern with the frequent verb. Ambridge et al.
(2008) also found that fully grammatical sentences with high frequency verbs were
judged slightly more acceptable than sentences with low-frequency verbs, although the
di¤erence was very small (for adults, 4.82 v. 4.76; the authors do not indicate whether
or not it was statistically significant).
11. Note that constructing examples for a linguistic paper (or for one’s students to analyse)
is a very di¤erent kind of activity from ordinary language use: it is conscious and delib-
erate, and relies on metalinguistic and/or general problem-solving abilities rather than
normal linguistic routines.
420 E. Da˛browska

As a result, they are much more likely to develop more general represen-
tations of these constructions, and accept unprototypical instances of
them.12

Received 11 April 2007 She‰eld University, UK


Revision received 18 December 2007

Appendix: List of sentences used in version 1 of the experiment


Experimental sentences
Prototypical LDD question (WH Prototypical)
What did you say the family should know before they go there?
What do you think they decided to do when they got home?
Where did you say they hid the treasure when they found out?
Where do you think the children could stay when their father returns?
LDD question with lexical subject (WH Subject)
What did Steve say we bought for Alice when we visited her?
What does Claire think the witness will say if they don’t intervene?
Where did Andy say the young man went after they found her?
Where does Paul think we put the documents after he saw them?
LDD question with a di¤erent verb (WH Verb)
What did you claim we bought for Alice when we visited her?
What do you believe the witness will say if they don’t intervene?
Where did you swear the young man went after they found her?
Where do you suspect we put the documents after he saw them?
LDD question with a di¤erent auxiliary (WH Auxiliary)
What will you say the family should know before they go there?
What would you think they decided to do when they got home?
Where will you say they hid the treasure when they found out?
Where would you think the children could stay when their father returns?
LDD question with an overt complementizer (WH Complementizer)
What did you say that the family should know before they go there?
What do you think that they decided to do when they got home?
Where did you say that they hid the treasure when they found out?

12. For other work suggesting that linguists’ judgments may be systematically di¤erent
from those of ordinary speakers, see Spencer 1973 and Bradac et al. 1980. See also
Hiramatsu 1999 and Snyder 2000 for experimental studies of ‘‘syntactic satiation’’, a
phenomenon whereby sentences which were initially judged ungrammatical become
increasingly acceptable as a result of repeated exposure.
Questions with long-distance dependencies 421

Where do you think that the children could stay when their father
returns?
LDD question with an additional subordinate clause (WH Long)
What did you say Eve claimed we bought during our first visit?
What do you think Jo believes he said at the court hearing?
Where did you say Mike swore he went after the evening performance?
Where do you think Phil suspects we were during the last war?
Unprototypical LDD question (WH Unprototypical)
What will Steve believe that Jo thinks they did with their old furniture?
What would Claire claim that Eve said they know about the whole a¤air?
Where will Andy suspect that Phil thinks they stayed during the school
holidays?
Where would Paul swear that Mike said they were during the afternoon
session?
Grammatical control sentences
‘‘Prototypical’’ declarative (DE Prototypical)
And you think the children could stay here when their father returns.
But you think they decided to do something when they got home
So you said the family should know everything before they go there.
So you said they hid the treasure somewhere when they found out.
Declarative with lexical subject (DE Subject)
And Claire thinks the witness will say something if they don’t intervene.
But Steve said we bought something for Alice when we visited her.
So Andy said the young man went home after they found her.
So Paul thinks we put the documents back after he saw them.
Declarative with a di¤erent verb (DE Verb)
And you swore the young man went home after they found her.
But you suspect we put the documents back after he saw them.
So you believe the witness will say something if they don’t intervene.
So you claimed we bought something for Alice when we visited her.
Declarative with a di¤erent auxiliary (DE Auxiliary)
You will say the family should know everything before they go there.
You will say they hid the treasure somewhere when they found out.
You would think the children could stay here when their father returns.
You would think they decided to do something when they got home.
Declarative with an overt complementizer (DE Complementizer)
And you said that the family should know everything before they go
there.
422 E. Da˛browska

But you think that the children could stay here when their father returns.
So you said that they hid the treasure somewhere when they found out.
So you think that they decided to do something when they got home.

Declarative with an additional subordinate clause (DE Long)


And you think Phil suspects we were here during the last war.
But you said Mike swore he went home after the evening performance.
So you said Eve claimed we bought something during our first visit.
So you think Jo believes he said something at the court hearing.
‘‘Unprototypical’’ Declarative (DE Unprototypical)
Andy will suspect that Phil thinks they stayed here during the school
holidays.
Claire would claim that Eve said they know everything about the whole
a¤air.
Paul would swear that Mike said they were here during the afternoon
session.
Steve will believe that Jo thinks they did something with their old
furniture.

Ungrammatical control sentences


that trace sentences (*that)
What did you say that will kill cockroaches but not ants?
What do you think that probably got lost during the move?
Who did you say that ate the spinach your mother cooked?
Who do you think that will turn up in the evening?
Sentences with extraction from a complex NP (*ComplexNP)
What did Claire make the claim that she read in a book?
What did Paul hear the rumour that I found in my garage?
Where did you discover the fact that the criminals put the car?
Where did you put forward the hypothesis that all the weapons were?
Negatives without do support (*Not)
Her husband not claimed they asked where we were going.
The manager not implied you knew what they were doing.
The teacher not suspected she remembered where that woman lived.
Your sister not believed I forgot what he had done.
Declaratives with double tense marking (*DoubleTn)
His cousin doesn’t thinks we lied because we were afraid.
The girl doesn’t remembers where she spent her summer holidays.
The mother doesn’t knows Julia was absent from school today
Your brother doesn’t believes the man is telling the truth.
Questions with long-distance dependencies 423

References

Ambridge, Ben and Adele E. Goldberg


this issue The island status of clausal complements: Evidence in favor of an informa-
tion structure explanation.
Ambridge, Ben, Julian M. Pine, Caroline F. Rowland and Chris R. Young
2008 The e¤ect of verb semantic class and verb frequency (entrenchment) on
children’s and adults’ graded judgments of argument structure overgenerali-
zation errors. Cognition.
Bradac, James J., Larry W. Martin, Norman D. Elliott and Charles H. Tardy
1980 On the neglected side of linguistic science: multivariate studies of sentence
judgment. Linguistics 18, 967–995.
British National Corpus, The, version 2 (BNC World)
2001 Distributed by Oxford University Computing Services on behalf of the BNC
Consortium. URL: http://www.natcorp.ox.ac.uk/.
Cheng, Lisa Lai-Shen and Norbert Corver (eds.)
2006 Wh Movement: Moving On. MIT Press.
Chomsky, Noam
1977 On wh-movement. In Formal Syntax, edited by Peter W. Culicover, Thomas
Wasow, and Adrian Akmajian. Academic Press, New York, 71–132.
Cowart, Wayne
1997 Experimental Syntax: Applying Objective Methods to Sentence Judgments.
Sage Publications, Thousand Oaks, CA.
Culicover, Peter W.
1997 Principles and Parameters: An Introduction to Syntactic Theory. Oxford Uni-
versity Press, Oxford.
Da˛browska, Ewa
2008 The e¤ects of frequency and neighbourhood density on adult native speak-
ers’ productivity with Polish case inflections: An empirical test of usage-based
approaches to morphology. Journal of Memory and Language 58, 931–951.
in press a Prototype e¤ects in questions with ‘unbounded’ dependencies.
in press b Naı̈ve v. expert competence: An empirical study of speaker intuitions.
2004 Language, Mind and Brain. Some Psychological and Neurological Con-
straints on Theories of Grammar. Edinburgh University Press, Edinburgh.
Da˛browska, Ewa and Elena Lieven
2005 Developing question constructions: Lexical specificity and usage-based oper-
ations. Cognitive Linguistics 16, 437–474.
Faneslow, Gisbert and Stefan Frisch
2006 E¤ects of processing di‰culty on judgments of acceptability. In Gradience in
Grammar, edited by G. Fanselow, C. Féry, M. Schlesewsky, and R. Vogel.
Oxford University Press, Oxford.
Featherston, Sam
2005 Universals and grammaticality: wh-constraints in German and English. Lin-
guistics 43, 667–711.
Frazier, Lyn and Charles Clifton, Jr.
1989 Successive cyclicity in the grammar and the parser. Language and Cognitive
Processes 4, 93–126.
Goldberg, Adele E.
2006 Constructions at Work. The Nature of Generalization in Language. Oxford
University Press, Oxford.
424 E. Da˛browska

Haegeman, Liliane
1991 Introduction to Government and Binding Theory. Basil Blackwell, Oxford.
Hawkins, John A.
1999 Processing complexity and filler-gap dependencies across grammars. Lan-
guage 75, 244–285.
Hiramatsu, Kazuko
1999 What syntactic satiation can tell us about islands. Papers from the Regional
Meetings, Chicago Linguistic Society 35, 141–151.
Holmes, V. M., L. Stowe and L. Cupples
1989 Lexical expectations in parsing complement-verb sentences. Journal of Mem-
ory and Language 28, 668–689.
Kluender, Robert and Marta Kutas
1993 Subjacency as a processing phenomenon. Language and Cognitive Processes
8, 573–633.
Lako¤, George
1987 Women, Fire and Dangerous Things. What Categories Reveal about the
Mind. Chicago University Press, Chicago.
Langacker, Ronald W.
1987 Foundations of Cognitive Grammar. Volume 1: Theoretical Prerequisites.
Stanford University Press, Stanford, CA.
1991 Foundations of Cognitive Grammar. Volume 2: Descriptive Application. Stan-
ford University Press, Stanford, CA.
2000 A dynamic usage-based model. In Usage-Based Models of Language, edited
by Michael Barlow and Suzanne Kemmer, 1–63. CSLI Publications, Stan-
ford, CA.
Levine, Robert D. and Thomas E. Hukari
2006 The Unity of Unbounded Dependency Constructions. CSLI Publications,
Stanford, CA.
Lieven, Elena V., Heike Behrens, Jennifer Speares, and Michael Tomasello
2003 Early syntactic creativity: A usage-based approach. Journal of Child Lan-
guage 30, 333–370.
Moon, Rosamund
1998 Fixed Expressions and Idioms in English: A Corpus-Based Approach. OUP,
Oxford.
Ouhalla, Jamal
1994 Introducing Transformational Grammar: From Rules to Principles and Pa-
rameters. Edward Arnold, London.
Sinclair, John (ed.)
1987 Collins Cobuild English Language Dictionary. London and Glasgow.,
HarperCollins.
Snyder, William
2000 An experimental investigation of syntactic satiation e¤ects. Linguistic In-
quiry 31, 575–582.
Spencer, N. J.
1973 Di¤erences between linguists and non-linguists in intuitions of
grammaticality-acceptability. Journal of Psycholinguistic Research 2, 83–
98.
Theakston, Anna L.
2004 The role of entrenchment in children’s and adults’ performance on gramma-
ticality judgment tasks. Cognitive Development 19, 15–34.
Questions with long-distance dependencies 425

Thompson, Sandra
2002 ‘‘Object complements’’ and conversation. Towards a realistic account.
Studies in Language 26, 125–164.
Tomasello, Michael, Nameera Akhtar, Kelly Dodson, and Laura Rekau
1997 Di¤erential productivity in young children’s use of nouns and verbs. Journal
of Child Language 24, 373–387.
Trueswell, John C., Michael K. Tanenhaus, and Christopher Kello
1993 Verb-specific constraints in sentence-processing—separating e¤ects of lexical
preference from garden-paths. Journal of Experimental Psychology: Learn-
ing Memory and Cognition 19, 528–553.
Verhagen, Arie
2005 Constructions of Intersubjectivity: Discourse, Syntax and Cognition. Oxford
University Press, Oxford.
Wasow, Thomas, and Jennifer Arnold
2005 Intuitions in linguistic argumentation. Lingua 115, 1481–1496.

You might also like