home

Description

Various recent threads of research within Linguistics are giving shape to the new field of Cognitive Computational Phonology, at the intersection of Linguistics, Psycholinguistics and Machine Learning:

The SFL lab in Paris is starting a research group on Cognitive Computational Phonology, with a focus on computational models of the child acquisition of phonotactics and phonology. To start off, we are organizing this informal one-day workshop, that brings together researchers working on the child acquisition of phonology from one of the three perspectives of psycholinguistics, learnability and linguistics.

Date and venue

The workshop will meet on Saturday November 10th 2012, at the following location:

Laboratoire Structures Formelles du Langage (UMR 7023)

UPS Pouchet

59/61 rue Pouchet

75017 Paris, France

Salle de Conferences (008), ground floor

Click here for a map and information about public transportation.

Program

Abstracts

Emmanuel Dupoux (LSCP, ENS)

Quantitative modeling of early phonological development

The past 40 years of psycholinguistic research has shown that infants learn their first language at an impressive speed. During the first year of life, even before they start to talk, infants converge on the basic building blocks of the phonological structure of their language. Yet, the way in which they are achieving this early phonological acquisition is counterintuitive from the viewpoint of linguistic theories. For instance, they learn phonotactics at an age where they know very few words and have not yet converged on the phoneme set of their language. We show that a modeling approach based on machine learning algorithms and speech technology applied to large speech databases can help to make sense of this developmental pattern. First, we argue that because of acoustic variability, phonemes cannot be acquired directly from the acoustic signal; only highly context dependent and talker dependent phones or phones fragments can be extracted in a bottom-up way. Second, words cannot be acquired directly from the acoustic signal either, but a small number of protowords or sentence fragments can be extracted on the basis of repetition frequency. Third, these two kinds of protolinguistic units can interact with one another in order to converge with more abstract units. The proposal is therefore that the different levels of the phonological system are acquired in parallel, through increasingly more precise approximations. This accounts for the largely overlapping development of lexical and phonological knowledge during the first year of life.

Naomi Yamaguchi (UMR 5596; CNRS & Université Lumière Lyon 2)

A phonological and frequential model of consonant acquisition in French

(slides available here)

This work focuses on consonantal acquisition of monolingual French-speaking children. Its aim is to show that the use of distinctive features and their associated principles (feature hierarchy, markedness avoidance, feature economy) captures the path of consonantal acquisition in French. The data of this dissertation consist of spontaneous longitudinal productions (during 16 and 28 months) of two French- speaking children. Analysis of the data reveals two main stages in the acquisition path of consonantal contrasts. Each of these stages relies on the intervention of a principle associated with distinctive features. The first stage captures the isolated acquisition of contrasts between consonants. The order of the acquisition of contrasts is guided by the feature hierarchy principle, which is expressed by feature robustness: the more robust a feature is, the faster it will be acquired. The acquisition of a feature also implies the acquisition of its two values through the intervention of the avoidance of markedness principle: the unmarked value of a feature will be acquired before the marked value. The second stage of the feature acquisition path consists in the distribution of a feature - acquired in an isolated way - throughout the whole system. This distribution is guided by the economy feature principle: the more a feature participates in the system economy, the more rapidly it will diffuse. In order to extract the relevant information from the child’s input that allows us to express the realisation of each principle into the language, we designed feature frequency calculations. We established a link between the expression of the hierarchy, markedness avoidance and economy principles and the different feature frequencies in child-directed speech. By approaching consonantal acquisition as contrast acquisition within an entire system, we were able to model the consonantal acquisition path based on distinctive features and their associated principles, paralleling it with work on the structure of adult sound inventories.

Sophie Wauquier (UMR 7023, Universite de Paris 8)

Templates as a learning strategy: The case of French

(slides available here)

No matter the input language, children's productions seem to begin (around the 50 word-stage) with a few systematic structural shapes built on the basis of a small inventory of features. According to the template hypothesis (Macken, 1992; Vihman, 2001), children at this stage are indeed matching their incomplete segmental content onto a defined number of prosodic (syllabic?) positions. Based on new data consisting of 12 monthly recordings from six monolingual French children (from a mean age of 19 months), we chart the reorganizational processes that lead from individual templates to the emergence of more adult-like structures and stable segmental, syllabic and prosodic representations. We argue that templatic activity provides a cognitive learning strategy. In the sense that it allows babies to provide a structural (phonological) response by extracting relevant linguistic structures from the linear input of a specific language. It helps them to be linguistically active in a period when the necessary grammatical and lexical knowledge is still missing. Templates facilitate production and help babies to interpret their own output as a structured input; this in turn supports a looping learning mechanism: the stabilization of their own phonological representations is reinforced.

Adam Albright (Department of Linguistics, MIT)

Evidence for feature-level generalization of phonological alternations

In order to learn alternations like [foʊt]~[foʊdi] and [seɪp]~[seɪbee], learners must identify which segments are affected, as well as the relation between surface alternants. In Optimality Theory, this minimally involves identifying the faithfulness constraints regulating the relevant changes, and demoting them below markedness constraints motivating the alternation. Analysts typically assume that faithfulness constraints regulate individual featural changes (IDENT[±voi]), rather than segmental correspondences (“FAITH(t~d)”, “FAITH(p~b)”). Therefore, multiple segment pairs may provide evidence for the same constraint ranking: t~d and p~b both require demotion of IDENT(±voi). This leads to an interesting prediction, especially in frameworks employing gradual constraint demotion on absolute numeric scales (Boersma 1997; Jaeger 2007): the speed of learning a segmental alternation x~y should depend not only on the frequency of x~y, but also on the total frequency of all pairs sharing the same featural change.

To test this prediction, we ran an Artificial Grammar experiment in which subjects implicitly learned consonant alternations in noun paradigms. In the artificial language, obstruent-final items always exhibited voicing or continuancy alternations. Among labials, p~f alternations outnumbered p~b, while among coronals, t~d outnumbered t~s. Overall, voicing alternations (OO-IDENT(±voi) violations) outnumbered continuancy alternations (OO-IDENT(±cont)). If learning t~d alternations facilitates learning p~b due to shared OO-IDENT(±voi) violations, we expect subjects to generalize p~b alternations more than p~f, despite their lower frequency. The results (58 subjects, Amazon MTurk) show that in a familiarity task, subjects erroneously judge that they have previously seen novel items in proportion to the frequency of the segmental alternation during training. This indicates that subjects correctly learned that continuancy alternations were more common among labials, and used this knowledge when guessing about lexicality (Schütze 2005). However, when selecting plurals for novel nouns in a generalization task, subjects preferred voicing alternations for both places. Thus, as predicted, exposure to voicing in coronals facilitated generalization of voicing in labials.

We model generalization with a grammar of weighted constraints, in which a single OO-IDENT(±voi) constraint applies to both labials and coronals. The preference for voicing alternations across both places is modeled by faster demotion of OO-IDENT(±voi) relative to OO-IDENT(±cont). Crucially, models that learn surface segment mappings (p~b, t~d) and lack featural faithfulness ([–voi]~[+voi]) cannot capture this facilitation effect. The fact that voicing alternations are nonetheless applied more readily to coronals is attributed to the fact that OO-IDENT competes with distinct markedness constraints: *VtV, *VpV. Since labials are rarer during training, the latter is promoted more slowly, and subjects frequently select non-alternating p~p.

Paula Fikkert (Dutch Department, Radboud University Nijmegen)

Where do phonological asymmetries come from?

In this talk I’ll provide an overview of various types of asymmetries that have been attested in various studies in our lab. We have investigated Place of Articulation, Manner of Articulation and Voicing. In many (though not all) experiments – from discrimination studies to word learning and word recognitions studies – asymmetries have been found, where a change from one direction to the other is noticed, but not the reverse. Investigations of production data show asymmetries as well: before a contrast is acquired children often seem to collapse sound categories and only use one form for various different target forms. Even when a contrast is in principle acquired, children still make errors, which show the same asymmetries, i.e. errors do not seem to be random. These results raise a number of questions, which will be addressed in this talk: Where do asymmetries come from? Are they part of universal phonology or the result of learning? What is the relationship between perception and production? What is the role of the lexicon in acquisition of phonology?

Alejandrina Cristia (Neurobiology of Language Department, Max Planck Institute for Psycholinguistics)

Beyond the state-of-the-art in infant speech perception

(slides available here)

The finding that perception begins to specialize to the ambient language(s) even in the first year of life spawned an ever-growing mass of research on the infant precursors of linguistic knowledge. To those who are not players in this active field, gaining a clear overview of what exactly babies know (or not) may be difficult. In this talk, I will first give a brief overview of the main take-home messages regarding infant speech perception. This overview will draw from both classical findings and state-of-the-art techniques, in order to provide a bird’s eye view on the phonological structure that infants at different ages impose on acoustic signals. Second, I will describe the greatest challenges still facing this research. Methodological limitations, conceptual vagueness, and the complexity of causal links remain inescapable when trying to understand early language acquisition. Moreover, ignoring these problems makes results from collaborations between infant researchers and other scientists misleading at best. Finally, I will outline some recent trends, and some ongoing projects, that seek to ameliorate the situation. These involve the development of novel techniques and the reinforcement of meta-analytic approaches that would allow more reliable empirical conclusions; the application of stricter definitions of theoretical knowledge; and the creation of large, public, multimodal databases.

Giorgio Magri (UMR 7023; CNRS/Universite de Paris 8)

The EDRA model of the child acquisition of phonotactics

(slides available here)

Nine-month-old infants already react differently to licit versus illicit sound combinations (Jusczyk et al. 1993), thus displaying knowledge of the target adult phonotactics. Children must thus rely on a remarkably efficient phonotactics learning strategy. What could it look like? According to the OT error-driven ranking model, the learner starts from a restrictive initial ranking, is trained on a stream of licit forms from the target adult language, and instantaneously slightly reranks the constraints whenever a mistake is made on the current piece of data. This learning model has been endorsed by the OT acquisition literature (Pater & Barlow 2003, Boersma & Levelt 2001, etcetera) because of its cognitive plausibility: it predicts a sequence of rankings that can be matched with child acquisition paths, thus modeling the observed acquisition gradualness; it relies on surface phonology without requiring any knowledge of morphology, that plausibly develops later than phonotactics; and it does not impose unrealistic memory requirements, as it only looks at a piece of data at the time without keeping track of previously seen data. Yet, error-driven ranking algorithms have been dismissed by the OT computational literature (Prince & Tesar 2004, Hayes 2004, etcetera) as algorithmically too weak: the behavior of the model depends on the stream of data, so that the model feels like a leaf in the wind of data, with little guarantees about the quality of its final grammar. In particular, with little guarantees concerning the restrictiveness of the final grammar, namely its ability to correctly rule out illicit forms, despite the fact that the algorithm is only trained on licit forms. Towards a reconciliation of these two acquisition and computational perspectives, I will present some initial but encouraging results on restrictiveness of error-driven ranking algorithms, that suggest that OT might have special formal properties that make it ideally suited in order to boost the algorithmic strength of this learning scheme.

Registration

The workshop is open and there is no registration fee. If you would like to attend, please email one of the two organizers below by November 3rd, 2012:

For any further information, please contact one of the workshop organizers.

Funding

This workshop is financially supported by the SFL UMR 7023 and by the Fyssen Research Foundation.