A joint model of word segmentation and phonological variation for English word-final /t/-deletion

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 1508–1516, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics A joint model of word segmentation and phonological variation for English word-final /t/-deletion Benjamin B¨orschinger1,3 and Mark Johnson1 and Katherine Demuth2 (1) Department of Computing, Macquarie University (2) Department of Linguistics, Macquarie University (3) Department of Computational Linguistics, Heidelberg University {benjamin.borschinger, mark.johnson, katherine.demuth}@mq.edu.au Abstract Word-final /t/-deletion refers to a common phenomenon in spoken English where words such as /wEst/ “west” are pro- nounced as [wEs] “wes” in certain con- texts. Phonological variation like this is common in naturally occurring speech. Current computational models of unsu- pervised word segmentation usually as- sume idealized input that is devoid of these kinds of variation. We extend a non-parametric model of word segmenta- tion by adding phonological rules that map from underlying forms to surface forms to produce a mathematically well-defined joint model as a first step towards han- dling variation and segmentation in a sin- gle model. We analyse how our model handles /t/-deletion on a large corpus of transcribed speech, and show that the joint model can perform word segmentation and recover underlying /t/s. We find that Bi- gram dependencies are important for per- forming well on real data and for learning appropriate deletion probabilities for dif- ferent contexts.1 1 Introduction Computational models of word segmentation try to solve one of the first problems language learn- ers have to face: breaking an unsegmented stream of sound segments into individual words. Cur- rently, most such models assume that the input consists of sequences of phonemes with no pro- nunciation variation across different occurrences of the same word type. In this paper we describe 1The implementation of our model as well as scripts to prepare the data will be made available at http://web.science.mq.edu.au/~bborschi. We can’t release our version of the Buckeye Corpus (Pitt et al., 2007) directly because of licensing issues. an extension of the Bayesian models of Gold- water et al. (2009) that incorporates phonologi- cal rules to “explain away” surface variation. As a concrete example, we focus on word-final /t/- deletion in English, although our approach is not limited to this case. We choose /t/-deletion be- cause it is a very common and well-studied phe- nomenon (see Coetzee (2004, Chapter 5) for a review) and segmental deletion is an interesting test-case for our architecture. Recent work has found that /t/-deletion (among other things) is in- deed common in child-directed speech (CDS) and, importantly, that its distribution is similar to that in adult-directed speech (ADS) (Dilley et al., to ap- pear). This justifies our using ADS to evaluate our model, as discussed below. Our experiments are consistent with long- standing and recent findings in linguistics, in par- ticular that /t/-deletion heavily depends on the im- mediate context and that models ignoring context work poorly on real data. We also examine how well our models identify the probability of /t/- deletion in different contexts. We find that models that capture …

Content truncated to 500 words