A multitude of studies show the relevance of both inflectional paradigms (word form frequency dis... more A multitude of studies show the relevance of both inflectional paradigms (word form frequency distributions, i.e., inflectional entropy) and inflectional classes (whole class frequency distributions) for visual lexical processing. Their interplay has also been proven significant, measured as the difference between paradigm and class frequency distributions (relative entropy). Relative entropy effects have now been recorded in nouns, verbs, adjectives, and prepositional phrases. However, all of these studies used visual stimuli – either written words or picture-naming tasks. The goal of our study is to test whether the effects of relative entropy can also be captured in the auditory modality. Forty young native speakers of Romanian (60% female) living in Serbia as part of the Romanian ethnic minority participated in an auditory lexical decision task. Stimuli were 168 Romanian verbs from two inflectional classes. Verbs were presented in four forms: present and imperfect 1st person sin...
Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies
Article prediction is a task that has long defied accurate linguistic description. As such, this ... more Article prediction is a task that has long defied accurate linguistic description. As such, this task is ideally suited to evaluate models on their ability to emulate native-speaker intuition. To this end, we compare the performance of native English speakers and pre-trained models on the task of article prediction set up as a three way choice (a/an, the, zero). Our experiments with BERT show that BERT outperforms humans on this task across all articles. In particular, BERT is far superior to humans at detecting the zero article, possibly because we insert them using rules that the deep neural model can easily pick up. More interestingly, we find that BERT tends to agree more with annotators than with the corpus when inter-annotator agreement is high but switches to agreeing more with the corpus as inter-annotator agreement drops. We contend that this alignment with annotators, despite being trained on the corpus, suggests that BERT is not memorising article use, but captures a high level generalisation of article use akin to human intuition.
Non-native speakers have been found to be penalized for their accent and grammatical errors. Howe... more Non-native speakers have been found to be penalized for their accent and grammatical errors. However, little is known about whether and how accent and grammaticality interact to influence native listeners' perception of non-native speech, and whether listeners' personality plays a role in this. We examined these questions by relating the acceptability scores of 40 English speech stimuli rated by 60 British listeners (30 female; 30 unfamiliar with accented speech) to factors including speech accent (British vs. Polish), grammaticality (well-formed vs. error-filled) and listener personality. The results suggest that non-native accent "protects" the speaker from being penalized for grammatical errors unless the listener has a certain personality profile: compared to non-native accented speech, the acceptability ratings of native accented speech showed a larger decrease when grammatical errors were present, yet listeners who were more conscientious, extravert or agreea...
Dividing attention by loading working memory is an effective method of probing the declarative an... more Dividing attention by loading working memory is an effective method of probing the declarative and procedural underpinnings of linguistic knowledge. The current study explores WM/DA effects on four specific L2 domains: morphology (aspect versus case), syntax and semantics (collocations). We contrast performance of 68 learners of Polish as a foreign language (L1 Chinese) in a grammaticality judgement task in baseline and divided attention conditions. We found corroborating evidence for heightened dependency on declarative memory in early (A2) L2 learners across all linguistic categories and in both experimental conditions. While the lexical judgement of more advanced (B1) learners also proved vulnerable to secondary-task interference, introduction of additional cognitive load had a positive effect on participants' grammatical judgement, yielding more accurate responses in the aspect category. These results point towards the need for pursuing a new line of inquiry into a potential...
We examined how language supports the expression of temporality within sentence boundaries in Eng... more We examined how language supports the expression of temporality within sentence boundaries in English, which has a rich inventory of grammatical means to express temporality. Using a computational model that mimics how humans learn from exposure we explored what the use of different tense and aspect (TA) combinations reveals about the interaction between our experience of time and the cognitive demands that talking about time puts on the language user. Our model was trained onn-grams extracted from the BNC to select the TA combination that fits the context best. It revealed the existence of two different sub-systems within the set of TA combinations, a “simplex” one that is supported lexically and is easy to learn, and a “complex” one that is supported contextually and is hard to learn. The finding that some TA combinations are essentially lexical in nature necessitates a rethink of tense and aspect as grammatical categories that form the axes of the temporal system. We argue that t...
Implicit sequence learning is an integral part of human experience, yet the nature of the mechani... more Implicit sequence learning is an integral part of human experience, yet the nature of the mechanisms underlying this type of learning remains a matter of debate. In the current study we provide a test for two accounts of implicit sequence learning, i.e., one that highlights sequence learning in the absence of any motor responses (with suppressed eye-movements) and one that highlights the relative contribution of the motor processes (i.e., eye movements) to learning. To adjudicate between these accounts and determine whether a motor response is a requisite process in sequence learning, we used anticipation measures to compare performance on the standard oculomotor serial reaction time (SRT) task and on a version of the SRT task where the eyemovements were restricted during the learning phase. Consistent with the response-based account, our results demonstrated an increased proportion of correct anticipations in the standard SRT task compared to the restricted-movement task.
Where a licence is displayed above, please note the terms and conditions of the licence govern yo... more Where a licence is displayed above, please note the terms and conditions of the licence govern your use of this document. When citing, please reference the published version. Take down policy While the University of Birmingham exercises care and attention in making items available there are rare occasions when an item has been uploaded in error or has been deemed to be commercially or otherwise sensitive.
Jos od studije Mejera i Svanevelta, mnoga istraivanja potvrđuju postojanje facilitatornog efekta... more Jos od studije Mejera i Svanevelta, mnoga istraivanja potvrđuju postojanje facilitatornog efekta u obradi date reci kada joj prethodi semanticki ili asocijativno povezana rec. U ovom istraivanju ispitivan je uticaj jacine asocijativne povezanosti konteksta i mete na vizuelno prepoznavanje reci u srpskom jeziku. Osnovni cilj bio je da se utvrdi kognitivna relevantnost tvrednosti i zajednicke informacije (MI) kao mera jacine asocijativne veze. Rezultati su pokazali da zajednicka informacija utice na brzinu obrade vizuelno prikazanih stimulusa: sa povecanjem kolicine zajednicke informacije koju neki bigram poseduje, skracuje se vreme reakcije u zadatku leksicke odluke. Sa druge strane, nije potvrđena kognitivna relevantnost t-vrednosti kao mere jacine asocijativne veze. Kljucne reci: t-vrednost, mere jacine asocijativne povezanosti reci, zajednicka informacija
A central goal of typological research is to characterize linguistic features in terms of both th... more A central goal of typological research is to characterize linguistic features in terms of both their functional role and their fit to social and cognitive systems. One longstanding puzzle concerns why certain languages employ grammatical gender. In an information theoretic analysis of German noun classification, Dye et al. (2017) enumerated a number of important processing advantages gender confers. Yet this raises a further puzzle: If gender systems are so beneficial to processing, what does this mean for languages that make do without them? Here, we compare the communicative function of gender marking in German (a deterministic system) to that of prenominal adjectives in English (a probabilistic one), finding that despite their differences, both systems act to efficiently smooth information over discourse, making nouns more equally predictable in context. We examine why evolutionary pressures may favor one system over another, and discuss the implications for compositional account...
is chapter proposes that the stable coexistence of regular and irregular patterns can be understo... more is chapter proposes that the stable coexistence of regular and irregular patterns can be understood in terms of a trade-o between the opposing communicative pressures imposed by predictability and discriminability. On this view, irregularity is not 'defective' or 'anomalous'. Instead, irregular formations exhibit an enhanced discriminability that brings them into maximal conformance with precepts like the 'one form-one meaning principle' , while allowing them to act as attractors within a larger system. sonversely, regularity is neither 'optimal' nor 'normative'. Regular patterns serve to facilitate predictability within a system. In order for regular items to perform this function, it must be possible to assign partially attested paradigms that exhaust the variation in the system. We suggest that a correlation between lexical neighbourhoods and patterns of co-lled cells bootstraps this analogical process.
Over the last decades, a growing body of evidence on the mechanisms governing lexical storage, ac... more Over the last decades, a growing body of evidence on the mechanisms governing lexical storage, access, acquisition and processing has questioned traditional models of language architecture and word usage based on the hypothesis of a direct correspondence between modular components of grammar competence (lexicon vs. rules), processing correlates (memory vs. computation) and neuro-anatomical localizations (prefrontal vs. temporo-parietal perisylvian areas of the left hemisphere). In the present chapter, we explore the empirical and theoretical consequences of a distributed, integrative model of the mental lexicon, whereby words are seen as emergent properties of the functional interaction between basic, language-independent processing principles and the language-specific nature and organization of the input. From this perspective, language learning appears to be inextricably related to the way language is processed and internalized by the speakers, and key to an interdisciplinary understanding of such a way, in line with Tomaso Poggio's suggestion that the development of a cognitive skill is causally and ontogenetically prior to its execution (and sits "on top of it"). In particular, we discuss conditions, potential and prospects of the epistemological continuity between psycholinguistic and computational modelling of word learning, and illustrate the yet largely untapped potential of their integration. We use David Marr's hierarchy to clarify the complementarity of the two viewpoints. Psycholinguistic models are informative about how speakers learn to use language (interfacing Marr's levels 1 and 2). When we move from the psycholinguistic analysis of the functional operations involved in language learning to an algorithmic description of how they are computed, computer simulations can help us explore the relation between speakers' behavior and general learning principles in more detail. In the end, psycho-computational
Studies of the structure and function of paradigms are as old as the Western grammatical traditio... more Studies of the structure and function of paradigms are as old as the Western grammatical tradition. The central role accorded to paradigms in traditional approaches largely reflects the fact that paradigms exhibit systematic patterns of interdependence that facilitate processes of analogical generalization. The recent resurgence of interest in word-based models of morphological processing and morphological structure more generally has provoked a renewed interest in paradigmatic dimensions of linguistic structure. Current methods for operationalizing paradigmatic relations and determining the behavioral correlates of these relations extend paradigmatic models beyond their traditional boundaries. The integrated perspective that emerges from this work is one in which variation at the level of individual words is not meaningful in isolation, but rather guides the association of words to paradigmatic contexts that play a role in their interpretation.
Proceedings of the 28th International Conference on Computational Linguistics, 2020
While lexico-semantic elements no doubt capture a large amount of linguistic information, it has ... more While lexico-semantic elements no doubt capture a large amount of linguistic information, it has been argued that they do not capture all information contained in text. This assumption is central to constructionist approaches to language which argue that language consists of constructions, learned pairings of a form and a function or meaning that are either frequent or have a meaning that cannot be predicted from its component parts. BERT's training objectives give it access to a tremendous amount of lexico-semantic information, and while BERTology has shown that BERT captures certain important linguistic dimensions, there have been no studies exploring the extent to which BERT might have access to constructional information. In this work we design several probes and conduct extensive experiments to answer this question. Our results allow us to conclude that BERT does indeed have access to a significant amount of information, much of which linguists typically call constructional information. The impact of this observation is potentially far-reaching as it provides insights into what deep learning methods learn from text, while also showing that information contained in constructions is redundantly encoded in lexicosemantics.
In this paper we present the Widrow-Hoff rule and its applications to language data. After contex... more In this paper we present the Widrow-Hoff rule and its applications to language data. After contextualizing the rule historically and placing it in the chain of neurally inspired artificial learning models, we explain its rationale and implementational considerations. Using a number of case studies we illustrate how the Widrow-Hoff rule offers unexpected opportunities for the computational simulation of a range of language phenomena that make it possible to approach old problems from a novel perspective.
A multitude of studies show the relevance of both inflectional paradigms (word form frequency dis... more A multitude of studies show the relevance of both inflectional paradigms (word form frequency distributions, i.e., inflectional entropy) and inflectional classes (whole class frequency distributions) for visual lexical processing. Their interplay has also been proven significant, measured as the difference between paradigm and class frequency distributions (relative entropy). Relative entropy effects have now been recorded in nouns, verbs, adjectives, and prepositional phrases. However, all of these studies used visual stimuli – either written words or picture-naming tasks. The goal of our study is to test whether the effects of relative entropy can also be captured in the auditory modality. Forty young native speakers of Romanian (60% female) living in Serbia as part of the Romanian ethnic minority participated in an auditory lexical decision task. Stimuli were 168 Romanian verbs from two inflectional classes. Verbs were presented in four forms: present and imperfect 1st person sin...
Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies
Article prediction is a task that has long defied accurate linguistic description. As such, this ... more Article prediction is a task that has long defied accurate linguistic description. As such, this task is ideally suited to evaluate models on their ability to emulate native-speaker intuition. To this end, we compare the performance of native English speakers and pre-trained models on the task of article prediction set up as a three way choice (a/an, the, zero). Our experiments with BERT show that BERT outperforms humans on this task across all articles. In particular, BERT is far superior to humans at detecting the zero article, possibly because we insert them using rules that the deep neural model can easily pick up. More interestingly, we find that BERT tends to agree more with annotators than with the corpus when inter-annotator agreement is high but switches to agreeing more with the corpus as inter-annotator agreement drops. We contend that this alignment with annotators, despite being trained on the corpus, suggests that BERT is not memorising article use, but captures a high level generalisation of article use akin to human intuition.
Non-native speakers have been found to be penalized for their accent and grammatical errors. Howe... more Non-native speakers have been found to be penalized for their accent and grammatical errors. However, little is known about whether and how accent and grammaticality interact to influence native listeners' perception of non-native speech, and whether listeners' personality plays a role in this. We examined these questions by relating the acceptability scores of 40 English speech stimuli rated by 60 British listeners (30 female; 30 unfamiliar with accented speech) to factors including speech accent (British vs. Polish), grammaticality (well-formed vs. error-filled) and listener personality. The results suggest that non-native accent "protects" the speaker from being penalized for grammatical errors unless the listener has a certain personality profile: compared to non-native accented speech, the acceptability ratings of native accented speech showed a larger decrease when grammatical errors were present, yet listeners who were more conscientious, extravert or agreea...
Dividing attention by loading working memory is an effective method of probing the declarative an... more Dividing attention by loading working memory is an effective method of probing the declarative and procedural underpinnings of linguistic knowledge. The current study explores WM/DA effects on four specific L2 domains: morphology (aspect versus case), syntax and semantics (collocations). We contrast performance of 68 learners of Polish as a foreign language (L1 Chinese) in a grammaticality judgement task in baseline and divided attention conditions. We found corroborating evidence for heightened dependency on declarative memory in early (A2) L2 learners across all linguistic categories and in both experimental conditions. While the lexical judgement of more advanced (B1) learners also proved vulnerable to secondary-task interference, introduction of additional cognitive load had a positive effect on participants' grammatical judgement, yielding more accurate responses in the aspect category. These results point towards the need for pursuing a new line of inquiry into a potential...
We examined how language supports the expression of temporality within sentence boundaries in Eng... more We examined how language supports the expression of temporality within sentence boundaries in English, which has a rich inventory of grammatical means to express temporality. Using a computational model that mimics how humans learn from exposure we explored what the use of different tense and aspect (TA) combinations reveals about the interaction between our experience of time and the cognitive demands that talking about time puts on the language user. Our model was trained onn-grams extracted from the BNC to select the TA combination that fits the context best. It revealed the existence of two different sub-systems within the set of TA combinations, a “simplex” one that is supported lexically and is easy to learn, and a “complex” one that is supported contextually and is hard to learn. The finding that some TA combinations are essentially lexical in nature necessitates a rethink of tense and aspect as grammatical categories that form the axes of the temporal system. We argue that t...
Implicit sequence learning is an integral part of human experience, yet the nature of the mechani... more Implicit sequence learning is an integral part of human experience, yet the nature of the mechanisms underlying this type of learning remains a matter of debate. In the current study we provide a test for two accounts of implicit sequence learning, i.e., one that highlights sequence learning in the absence of any motor responses (with suppressed eye-movements) and one that highlights the relative contribution of the motor processes (i.e., eye movements) to learning. To adjudicate between these accounts and determine whether a motor response is a requisite process in sequence learning, we used anticipation measures to compare performance on the standard oculomotor serial reaction time (SRT) task and on a version of the SRT task where the eyemovements were restricted during the learning phase. Consistent with the response-based account, our results demonstrated an increased proportion of correct anticipations in the standard SRT task compared to the restricted-movement task.
Where a licence is displayed above, please note the terms and conditions of the licence govern yo... more Where a licence is displayed above, please note the terms and conditions of the licence govern your use of this document. When citing, please reference the published version. Take down policy While the University of Birmingham exercises care and attention in making items available there are rare occasions when an item has been uploaded in error or has been deemed to be commercially or otherwise sensitive.
Jos od studije Mejera i Svanevelta, mnoga istraivanja potvrđuju postojanje facilitatornog efekta... more Jos od studije Mejera i Svanevelta, mnoga istraivanja potvrđuju postojanje facilitatornog efekta u obradi date reci kada joj prethodi semanticki ili asocijativno povezana rec. U ovom istraivanju ispitivan je uticaj jacine asocijativne povezanosti konteksta i mete na vizuelno prepoznavanje reci u srpskom jeziku. Osnovni cilj bio je da se utvrdi kognitivna relevantnost tvrednosti i zajednicke informacije (MI) kao mera jacine asocijativne veze. Rezultati su pokazali da zajednicka informacija utice na brzinu obrade vizuelno prikazanih stimulusa: sa povecanjem kolicine zajednicke informacije koju neki bigram poseduje, skracuje se vreme reakcije u zadatku leksicke odluke. Sa druge strane, nije potvrđena kognitivna relevantnost t-vrednosti kao mere jacine asocijativne veze. Kljucne reci: t-vrednost, mere jacine asocijativne povezanosti reci, zajednicka informacija
A central goal of typological research is to characterize linguistic features in terms of both th... more A central goal of typological research is to characterize linguistic features in terms of both their functional role and their fit to social and cognitive systems. One longstanding puzzle concerns why certain languages employ grammatical gender. In an information theoretic analysis of German noun classification, Dye et al. (2017) enumerated a number of important processing advantages gender confers. Yet this raises a further puzzle: If gender systems are so beneficial to processing, what does this mean for languages that make do without them? Here, we compare the communicative function of gender marking in German (a deterministic system) to that of prenominal adjectives in English (a probabilistic one), finding that despite their differences, both systems act to efficiently smooth information over discourse, making nouns more equally predictable in context. We examine why evolutionary pressures may favor one system over another, and discuss the implications for compositional account...
is chapter proposes that the stable coexistence of regular and irregular patterns can be understo... more is chapter proposes that the stable coexistence of regular and irregular patterns can be understood in terms of a trade-o between the opposing communicative pressures imposed by predictability and discriminability. On this view, irregularity is not 'defective' or 'anomalous'. Instead, irregular formations exhibit an enhanced discriminability that brings them into maximal conformance with precepts like the 'one form-one meaning principle' , while allowing them to act as attractors within a larger system. sonversely, regularity is neither 'optimal' nor 'normative'. Regular patterns serve to facilitate predictability within a system. In order for regular items to perform this function, it must be possible to assign partially attested paradigms that exhaust the variation in the system. We suggest that a correlation between lexical neighbourhoods and patterns of co-lled cells bootstraps this analogical process.
Over the last decades, a growing body of evidence on the mechanisms governing lexical storage, ac... more Over the last decades, a growing body of evidence on the mechanisms governing lexical storage, access, acquisition and processing has questioned traditional models of language architecture and word usage based on the hypothesis of a direct correspondence between modular components of grammar competence (lexicon vs. rules), processing correlates (memory vs. computation) and neuro-anatomical localizations (prefrontal vs. temporo-parietal perisylvian areas of the left hemisphere). In the present chapter, we explore the empirical and theoretical consequences of a distributed, integrative model of the mental lexicon, whereby words are seen as emergent properties of the functional interaction between basic, language-independent processing principles and the language-specific nature and organization of the input. From this perspective, language learning appears to be inextricably related to the way language is processed and internalized by the speakers, and key to an interdisciplinary understanding of such a way, in line with Tomaso Poggio's suggestion that the development of a cognitive skill is causally and ontogenetically prior to its execution (and sits "on top of it"). In particular, we discuss conditions, potential and prospects of the epistemological continuity between psycholinguistic and computational modelling of word learning, and illustrate the yet largely untapped potential of their integration. We use David Marr's hierarchy to clarify the complementarity of the two viewpoints. Psycholinguistic models are informative about how speakers learn to use language (interfacing Marr's levels 1 and 2). When we move from the psycholinguistic analysis of the functional operations involved in language learning to an algorithmic description of how they are computed, computer simulations can help us explore the relation between speakers' behavior and general learning principles in more detail. In the end, psycho-computational
Studies of the structure and function of paradigms are as old as the Western grammatical traditio... more Studies of the structure and function of paradigms are as old as the Western grammatical tradition. The central role accorded to paradigms in traditional approaches largely reflects the fact that paradigms exhibit systematic patterns of interdependence that facilitate processes of analogical generalization. The recent resurgence of interest in word-based models of morphological processing and morphological structure more generally has provoked a renewed interest in paradigmatic dimensions of linguistic structure. Current methods for operationalizing paradigmatic relations and determining the behavioral correlates of these relations extend paradigmatic models beyond their traditional boundaries. The integrated perspective that emerges from this work is one in which variation at the level of individual words is not meaningful in isolation, but rather guides the association of words to paradigmatic contexts that play a role in their interpretation.
Proceedings of the 28th International Conference on Computational Linguistics, 2020
While lexico-semantic elements no doubt capture a large amount of linguistic information, it has ... more While lexico-semantic elements no doubt capture a large amount of linguistic information, it has been argued that they do not capture all information contained in text. This assumption is central to constructionist approaches to language which argue that language consists of constructions, learned pairings of a form and a function or meaning that are either frequent or have a meaning that cannot be predicted from its component parts. BERT's training objectives give it access to a tremendous amount of lexico-semantic information, and while BERTology has shown that BERT captures certain important linguistic dimensions, there have been no studies exploring the extent to which BERT might have access to constructional information. In this work we design several probes and conduct extensive experiments to answer this question. Our results allow us to conclude that BERT does indeed have access to a significant amount of information, much of which linguists typically call constructional information. The impact of this observation is potentially far-reaching as it provides insights into what deep learning methods learn from text, while also showing that information contained in constructions is redundantly encoded in lexicosemantics.
In this paper we present the Widrow-Hoff rule and its applications to language data. After contex... more In this paper we present the Widrow-Hoff rule and its applications to language data. After contextualizing the rule historically and placing it in the chain of neurally inspired artificial learning models, we explain its rationale and implementational considerations. Using a number of case studies we illustrate how the Widrow-Hoff rule offers unexpected opportunities for the computational simulation of a range of language phenomena that make it possible to approach old problems from a novel perspective.
Proceedings of the 32nd International Workshop on Statistical Modelling (IWSM), 2017
Baayen, R. H., T. Sering, C. Shaoul, and P. Milin
The initial stage of language comprehension is... more Baayen, R. H., T. Sering, C. Shaoul, and P. Milin
The initial stage of language comprehension is a multi-label classi- fication problem. Listeners or readers, presented with an utterance, need to dis- criminate between the intended words and the tens of thousands of other words they know. We propose to address this problem by pairing a network trained with the learning rule of Rescorla and Wagner (1972) with a second network trained independently with the learning rule of Widrow and Hoff (1960). The first net- work has to recover from sublexical input features the meanings encoded in the language signal, resulting in a vector of activations over all meanings. The second network takes this vector as input and further reduces uncertainty about the in- tended meanings. Classification performance for a lexicon with 52,000 entries is good. The model also correctly predicts several aspects of human language com- prehension. By rejecting the traditional linguistic assumption that language is a (de)compositional system, and by instead espousing a discriminative approach (Ramscar, 2013), a more parsimonious yet highly effective functional characteri- zation of the initial stage of language comprehension is obtained.
Uploads
Papers by Petar Milin
The initial stage of language comprehension is a multi-label classi- fication problem. Listeners or readers, presented with an utterance, need to dis- criminate between the intended words and the tens of thousands of other words they know. We propose to address this problem by pairing a network trained with the learning rule of Rescorla and Wagner (1972) with a second network trained independently with the learning rule of Widrow and Hoff (1960). The first net- work has to recover from sublexical input features the meanings encoded in the language signal, resulting in a vector of activations over all meanings. The second network takes this vector as input and further reduces uncertainty about the in- tended meanings. Classification performance for a lexicon with 52,000 entries is good. The model also correctly predicts several aspects of human language com- prehension. By rejecting the traditional linguistic assumption that language is a (de)compositional system, and by instead espousing a discriminative approach (Ramscar, 2013), a more parsimonious yet highly effective functional characteri- zation of the initial stage of language comprehension is obtained.