Section: New Results

Modelling the acquisition of linguistic categories by children

Participants : Benoît Crabbé, Luc Boruta, Isabelle Dautriche.

This task breaks in two sub-tasks: acquisition of phonemic categories, and acquisition of syntactic categories.

Although we are only able to distinguish between a finite, small number of sound categories – i.e., a given language's phonemes – no two sounds are actually identical in the messages we receive. Given the pervasiveness of sound-altering processes across languages – and the fact that every language relies on its own set of phonemes – the question of the acquisition of allophonic rules by infants has received a considerable amount of attention in recent decades. How, for example, do English-learning infants discover that the word forms [kæt] and [kat] refer to the same animal species (i.e. cat), whereas [kæt] and [bæt] (i.e. cat $\sim$ bat) do not? What kind of cues may they rely on to learn that [sı$\eta$kı$\eta$] and [$\theta$ı$\eta$kı$\eta$] (sinking $\sim$ thinking) can not refer to the same action? The work presented in this dissertation builds upon the line of computational studies initiated by [90] , wherein research efforts have been concentrated on the definition of sound-to-sound dissimilarity measures indicating which sounds are realizations of the same phoneme. We show that solving Peperkamp et al.'s task does not yield a full answer to the problem of the discovery of phonemes, as formal and empirical limitations arise from its pairwise formulation. We proceed to circumvent these limitations, reducing the task of the acquisition of phonemes to a partitioning-clustering problem and using multidimensional scaling to allow for the use of individual phones as the elementary objects. The results of various classification and clustering experiments consistently indicate that effective indicators of allophony are not necessarily effective indicators of phonemehood. Altogether, the computational results we discuss suggest that allophony and phonemehood can only be discovered from acoustic, temporal, distributional, or lexical indicators when—on average—phonemes do not have many allophones in a quantified representation of the input. This subtask has seen the Phd defense of Luc Boruta whose Phd thesis : "Indicators of allophony and phonemehood" was successfully defended in September 2012.

As for syntactic categorization, the task is concerned with modelling and implementing psychologically motivated models of language treatment and acquisition. Contrary to classical Natural Language Processing applications, the main aim was not to create engineering solutions to language related tasks, but rather to test and develop psycholinguistic theories. In this context, the study was concerned with the question of learning word categories, such as the categories of Noun and Verb. It is established experimentally that 2-year-old children can identify novel nouns and verbs. It has been suggested that this can be done using distributional cues as well as prosodic cues. While the plain distributional hypothesis had been tested quite extensively, the importance of prosodic cues has not been addressed in a computational simulation. We provided a formulation for modelling this hypothesis using unsupervised and semi-supervised forms of Bayesian learning (EM) both offline and online. This activity started with the master thesis of A. Gutman and has seen this year the start of a new Phd student : I. Dautriche.