Modeling Dynamical Linguistic Development in Early Bilinguals: a Computational Approach

The proposed project involves the design and implementation of a novel computational model of simultaneous acquisition of two languages from birth, or bilingual first language acquisition (BFLA). More specifically, we explore computationally the theory of Phonological Bootstrapping (Christophe et al. 1994, Christophe and Dupoux 1996, Morgan and Demuth 1996, Christophe et al. 1997), working within a bilingual context. Phonological Bootstrapping is a two-part phonological/acoustic-phonetic analysis outlining how infants start to acquire mental representations of word forms in the lexicon and early syntactic representations of their native language(s). According to the phonological bootstrapping model, children initially construct pre-lexical representations on the basis of cues to abstract symbols or linguistic objects (words or categories) that come readily available in the input from perceptual properties associated with that symbol/object. In order to find word forms and to also detect word boundaries for ‘bootstrapping’ into syntax, various sources of perceptual information are argued to be exploited as cues, including prosody, statistical analyses, and general purpose analogy mechanisms. To the degree that syntactic structures are a projection of lexical properties, researchers suggest that the child’s initial representations contain prosodically segmented units that are identifiable for each language and roughly correspond to syntactic units. In the proposed BFLA scenario the primary linguistic input is bilingual, and the lexicon(s) of content words and syntactic structures to be built are in two languages. In first language acquisition research, the concept of "bootstrapping" has been effectively used to explain how children utilize correlations between different aspects of language to infer structure. In our proposed computational approach, an adaptable analogical search for structures across primary linguistic inputs is employed. The core architecture is largely based on mechanisms from the Mitchell (1993) and Hofstadter (1995) program of Copycat, a cognitive computer model of ‘high-perception.’ Copycat has had success as a model of perception and analogy in various complex sequence domains (e.g., music, numbers, and letters), but to date it has not been applied to language learning tasks. We introduce a preliminary computational model, known as Babycat, to begin to test the BFLA hypotheses, and to serve as a springboard for a more advanced version of our current prototype. (Specifications are outlined in Section 4.) To calibrate and also evaluate the experiments that will be implemented in the proposed model, realworld bilingual language data are incorporated into this project from the CHILDES database, most notably data from an original longitudinal study (Pérez-Bazán 2002) of five Spanish-English bilingual infants, ages 0;8 to 3;0 years, and their caregivers.