Modeling Acquisition of Word Structure with Lexicalized Grammar Learning

This paper introduces a framework for learning structure in natural languages, and reports results from a simple application of it to learning word-syntax of an agglutinative language in an unsupervised manner. Arguably, the learning environment of children acquiring languages provides more information—by means of linguistic interaction and extralinguistic information present in the learning setting— than the information provided to an unsupervised learner. However, completely unsupervised learning methods can still provide insights into how children acquire language, at least, (i) by setting a lower bound on what is learnable, (ii) by identifying type and quantity of cues in the input that is useful for successful learning, (iii) by testing different learning methods, algorithms and frameworks on the basis of how successful they are in learning from the data available to children and how well they match with the available data from developmental psycholinguistics. In this paper, we will first describe the general learning framework based on learning a lexicalized grammar, Categorial Grammar (CG, Ajdukiewicz, 1935; Bar-Hillel, 1953), then we will present our morphological learner in more detail, followed by the results obtained on testing the learner on learning morphology of Turkish from child directed speech from CHILDES database. The learning algorithm uses techniques similar to unsupervised morphology learning systems such as Goldsmith (2001) and Creutz and Lagus (2007). However, this study tries to model human language acquisition more closely by using data from child directed speech and not assuming avialability of the complete data to the learner. Another major difference of this study is the emphasis on the structure learning. The model presented here learns a lexicalized word-grammar, which has similarities to other lexicalized grammar learners (e.g., Villavicencio, 2002; Zettlemoyer & Collins, 2005; Yao, Ma, Duarte, & Çöltekin, 2009).