TULCON 7: Crude Phonotactics: Evidence of UG in Conditional Random Field weights

• published: 11 Mar 2014
• views: 10

Abstract: In The sound pattern of English (1968), Noam Chomsky and Morris Halle defined the goal of linguistics as the search for the properties required to both describe any natural language and define all such possible languages and grammars. The study of language acquisition, which deals with the mechanisms and stages of acquiring a target language, has long been held a key into uncovering these universal properties, at least with respect to the areas of Phonology and Syntax (Fikkert 2007). The idea is by knowing how a child constructs the grammar, the structure of the final, adult grammar is implicit. One of the approaches from acquisition has been Uniformity of Learning (Gleitman and Newport 1999), where striking similarities across all children learning different languages is argued as evidence for a universal grammar, since each child would each receive different input, yet learn the language at the same rate and in the same patterns as their peers. In this talk, I explore phonetic output from five different children (three English, two French) from the Childes database (http://childes.psy.cmu.edu/browser/index.php?url=PhonBank-Phon/), where I employ a Conditional Random Field statistical model (CRF) to probe the systematicity of each child's production much in the spirit of Hayes and Wilson (2008), as its state transition weights represent each child's phonotactics. By comparing these weights for each child specific CRF model learned for the first several months of each child's earliest transcript, the results show strong similarities among the children's phonotactics, regardless of language or dialect. Additionally, by using the target adult form of each child's output as a sample of input each child received, I compared the weights of these adult inputs, and found that there was no significant similarity among these weights between languages. Using the weights as estimations of the phonotactics of each grammar, I argue that these results lend evidence to the existence of universal properties (at least for Phonology) under the Uniformity of Learning.