Combining multiple information types in Bayesian word segmentation

Proceedings of NAACL-HLT 2013, pages 117–126, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Combining multiple information types in Bayesian word segmentation Gabriel Doyle and Roger Levy Department of Linguistics University of California, San Diego La Jolla, CA 92093, USA {gdoyle,rlevy}@ucsd.edu Abstract Humans identify word boundaries in continu- ous speech by combining multiple cues; exist- ing state-of-the-art models, though, look at a single cue. We extend the generative model of Goldwater et al (2006) to segment using sylla- ble stress as well as phonemic form. Our new model treats identification of word boundaries and prevalent stress patterns in the language as a joint inference task. We show that this model improves segmentation accuracy over purely segmental input representations, and recov- ers the dominant stress pattern of the data. Additionally, our model retains high perfor- mance even without single-word utterances. We also demonstrate a discrepancy in the per- formance of our model and human infants on an artificial-language task in which stress cues and transition-probability information are pit- ted against one another. We argue that this dis- crepancy indicates a bound on rationality in the mechanisms of human segmentation. 1 Introduction For an adult speaker of a language, word segmen- tation from fluid speech may seem so easy that it barely needed to be learned. However, pauses in speech and word boundaries are not well cor- related (Cole & Jakimik, 1980), word boundaries are marked by a conspiracy of partially-informative cues (Johnson & Jusczyk, 2001), and different lan- guages mark their boundaries differently (Cutler & Carter, 1987). This makes the problem of unsuper- vised word segmentation acquisition, whether by a computational model or an infant, a daunting task. Effective segmentation relies on the flexible in- tegration of multiple types of segmentation cues, among them statistical regularities in phonemes and prosody, coarticulation, and allophonic variation. In- fants begin using multiple segmentation cues within their first year of life (Johnson & Jusczyk, 2001). Despite this, many state-of-the-art models look at only one type of information: phonemes. In this study, we expand an existing model to incorporate multiple cues, leading to an improve- ment in segmentation performance and opening new ways of investigating human segmentation acquisi- tion. On the latter point, we show that rational learn- ers can learn to segment without encountering words in isolation, and that human learners deviate from ra- tionality in certain segmentation tasks. 2 Previous work The prevailing unsupervised word segmentation sys- tems (e.g., Brent, 1999; Goldwater, Griffiths, & Johnson, 2006; Blanchard & Heinz, 2008) use only phonemic information to segment speech. However, human segmenters use additional information types, notably stress information, in their segmentation. We present an overview of these phonemic mod- els here before discussing the prosodic model ex- pansion. A more complete review is available in Goldwater (2007). 2.1 Goldwater et al (2006) The Goldwater et al model is related to Brent (1999)’s model, both of which use strictly phone- mic information to segment. The model assumes that the corpus is generated by a …

Content truncated to 500 words