Studying the Effect of Input Size for Bayesian Word Segmentation on the Providence Corpus

Proceedings of COLING 2012: Technical Papers, pages 325–340, COLING 2012, Mumbai, December 2012. Studying the effect of input size for Bayesian Word Segmentation on the Providence Corpus Benjamin B¨orschinger1,3 Katherine Demuth2 Mark Johnson1 (1) Department of Computer Science, Macquarie University (2) Department of Linguistics, Macquarie University (3) Department of Computational Linguistics, Heidelberg University benjamin.borschinger@mq.edu.au, katherine.demuth@mq.edu.au, mark.johnson@mq.edu.au ABSTRACT Studies of computational models of language acquisition depend to a large part on the input available for experiments. In this paper, we study the effect that input size has on the per- formance of word segmentation models embodying different kinds of linguistic assumptions. Because currently available corpora for word segmentation are not suited for addressing this question, we perform our study on a novel corpus based on the Providence Corpus (Demuth et al., 2006). We find that input size can have dramatic effects on segmentation performance and that, somewhat surprisingly, models performing well on smaller amounts of data can show a marked decrease in performance when exposed to larger amounts of data. We also present the data-set on which we perform our experiments comprising longitudinal data for six children. This corpus makes it possible to ask more specific questions about computational models of word segmentation, in particular about intra-language variability and about how the performance of different models can change over time.1 KEYWORDS: word segmentation, language acquisition, resources, Bayesian modelling, unsu- pervised learning. 1The corpus and the code to run the experiments is available at http://web.science.mq.edu.au/ ~bborschi/. 325 1 Introduction Segmenting a stream of sounds into discrete words is one of the first tasks that children acquiring their native language have to tackle. Computational models of word segmentation enable us to study this problem in a controlled and detailed manner, allowing for example for an examination of the usefulness of different kinds of cues or different learning strategies. Just as important as the actual models, however, is the adequacy of the input used to evaluate them — if we are interested in answering questions about human language acquisition, the data we evaluate our models on needs to be comparable to what children are likely to have access to. To this end, several datasets of phonemically transcribed child directed speech (CDS) have been constructed in several languages, ranging from English to Italian, Polish, Sesotho and Chinese (Brent and Cartwright, 1996; Gervain and Erra, 2012; Boruta and Jastrzebska, 2012; Johnson, 2008a; Johnson and Demuth, 2010). In addition to cross-linguistic variation, however, adequate computational models also need to handle language-internal variation along several dimensions, a topic that has so far received little interest. In this paper, we look at a basic point of variation, namely the actual size of the input to the learner. The longer children are exposed to language, the more data they are exposed to and the better at their language they become, something one would expect from adequate models of language acquisition as well. We run our experiments on a novel dataset that contains longitudinal data for six children from the Providence …

Content truncated to 500 words