Improving Word Segmentation by Simultaneously Learning Phonotactics

CoNLL 2008: Proceedings of the 12th Conference on Computational Natural Language Learning, pages 65–72 Manchester, August 2008 Improving Word Segmentation by Simultaneously Learning Phonotactics Daniel Blanchard Computer & Information Sciences University of Delaware dsblanch@udel.edu Jeffrey Heinz Linguistics & Cognitive Science University of Delaware heinz@udel.edu Abstract The most accurate unsupervised word seg- mentation systems that are currently avail- able (Brent, 1999; Venkataraman, 2001; Goldwater, 2007) use a simple unigram model of phonotactics. While this sim- plifies some of the calculations, it over- looks cues that infant language acquisition researchers have shown to be useful for segmentation (Mattys et al., 1999; Mattys and Jusczyk, 2001). Here we explore the utility of using bigram and trigram phono- tactic models by enhancing Brent’s (1999) MBDP-1 algorithm. The results show the improved MBDP-Phon model outper- forms other unsupervised word segmenta- tion systems (e.g., Brent, 1999; Venkatara- man, 2001; Goldwater, 2007). 1 Introduction How do infants come to identify words in the speech stream? As adults, we break up speech into words with such ease that we often think that there are audible pauses between words in the same sentence. However, unlike some written lan- guages, speech does not have any completely reli- able markers for the breaks between words (Cole and Jakimik, 1980). In fact, languages vary on how they signal the ends of words (Cutler and Carter, 1987), which makes the task even more daunting. Adults at least have a lexicon they can use to rec- ognize familiar words, but when an infant is first born, they do not have a pre-existing lexicon to consult. In spite of these challenges, by the age of c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. six months infants can begin to segment words out of speech (Bortfeld et al., 2005). Here we present an efficient word segmentation system aimed to model how infants accomplish the task. While an algorithm that could reliably extract orthographic representations of both novel and fa- miliar words from acoustic data is something we would like to see developed, following earlier re- searchers, we simplify the problem by using a text that does not contain any word boundary markers. Hereafter, we use the phrase “word segmentation” to mean some process which adds word boundaries to a text that does not contain them. This paper’s focus is on unsupervised, incre- mental word segmentation algorithms; i.e., those that do not rely on preexisting knowledge of a par- ticular language, and those that segment the cor- pus one utterance at a time. This is in contrast to supervised word segmentation algorithms (e.g., Teahan et al., 2000), which are typically used for segmenting text in documents written in languages that do not put spaces between their words like Chinese. (Of course, unsupervised word segmen- tation algorithms also have this application.) This also differs from batch segmentation algorithms (Goldwater, 2007; Johnson, 2008b; Fleck, 2008), which process the entire corpus at least once be- fore outputting a segmentation of the …

Content truncated to 500 words