Unsupervised Word Segmentation in Context

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 2326–2334, Dublin, Ireland, August 23-29 2014. Unsupervised Word Segmentation in Context Gabriel Synnaeve and Isabelle Dautriche LSCP, DEC ENS Ulm, Paris, France gabriel.synnaeve@gmail.com isabelle.dautriche@gmail.com Benjamin B¨orschinger Institut f¨ur Computerlinguistik Universit¨at Heidelberg, Heidelberg, Germany. benjamin.boerschinger@gmail.com Mark Johnson Department of Computer Science Macquarie University, Sydney, Australia mark.johnson@mq.edu.au Emmanuel Dupoux LSCP, DEC EHESS, Paris, France emmanuel.dupoux@gmail.com Abstract This paper extends existing word segmentation models to take non-linguistic context into ac- count. It improves the token F-score of a top performing segmentation models by 2.5% on a 27k utterances dataset. We posit that word segmentation is easier in-context because the learner is not trying to access irrelevant lexical items. We use topics from a Latent Dirichlet Allocation model as a proxy for “activities” contexts, to label the Providence corpus. We present Adaptor Grammar models that use these context labels, and we study their performance with and without context annotations at test time. 1 Introduction and Previous Works Segmentation of the speech stream into lexical units plays a central role in early language acquisition. Because words are generally not uttered in isolation, one of the first task for infants learning a language is to extract the words that make up the utterances they hear. Experimental research has shown that infants are able to segment fluent speech into word-like units within the first year of life (Jusczyk and Aslin, 1995). How does this ability emerge? There is evidence that infants use a broad array of linguistic cues to perform word segmentation (e.g., phonotactics (Jusczyk et al., 1993a), prosodic information (Jusczyk et al., 1993b), statistical regularities (Saffran et al., 1996)). Past experimental and modeling research on speech segmentation has mainly focused on linguistic cues, treating them as independent from other non- linguistic cues naturally occurring in the child learning environment. Yet, language appears in context and is constrained by the events occurring in the daily life of the child. For example, during an eating event one is most likely to speak about food, while during a zoo-visit event, people are more likely to talk about the animals they see. Activity contexts may provide a natural structure to speech that would be readily be accessible to children. A recent study using dense recordings of a single child’s language development (Roy et al., 2006) showed that words appearing in specific activity contexts are learned faster (Roy et al., 2012). Relatedly, Johnson et al. (2010) showed that Adaptor Grammars (AGs) performed better on a segmentation task when the model has access to a hand-annotated set of objects present in the environment, that it can use to learn simultaneously word-object associations (see also (Frank et al., 2009)). This supports the view that integrating multiple sources of information, linguistic and non-linguistic, can improve learning. Following this idea, we posit that information from the broader context in which a word has been uttered may simplify the learning problem faced by the child. In particular, our hypothesis postulates that speech …

Content truncated to 500 words