Exploring the Role of Stress in Bayesian Word Segmentation using Adaptor Grammars

Exploring the Role of Stress in Bayesian Word Segmentation using Adaptor Grammars Benjamin B¨orschinger1,2 Mark Johnson1,3 1Department of Computing, Macquarie University, Sydney, Australia 2Department of Computational Linguistics, Heidelberg University, Heidelberg, Germany 3Santa Fe Institute, Santa Fe, USA {benjamin.borschinger|mark.johnson}@mq.edu.au Abstract Stress has long been established as a major cue in word segmentation for English infants. We show that enabling a current state-of-the-art Bayesian word segmentation model to take ad- vantage of stress cues noticeably improves its performance. We find that the improvements range from 10 to 4%, depending on both the use of phonotactic cues and, to a lesser ex- tent, the amount of evidence available to the learner. We also find that in particular early on, stress cues are much more useful for our model than phonotactic cues by themselves, consistent with the finding that children do seem to use stress cues before they use phono- tactic cues. Finally, we study how the model’s knowledge about stress patterns evolves over time. We not only find that our model cor- rectly acquires the most frequent patterns rel- atively quickly but also that the Unique Stress Constraint that is at the heart of a previously proposed model does not need to be built in but can be acquired jointly with word segmen- tation. 1 Introduction Among the first tasks a child language learner has to solve is picking out words from the fluent speech that constitutes its linguistic input.1 For English, stress has long been claimed to be a useful cue in infant word segmentation (Jusczyk et al., 1993; Jusczyk et al., 1999b), following the demonstra- 1The datasets and software to replicate our experiments are available from http://web.science.mq.edu.au/ ˜bborschi/ tion of its effectiveness in adult speech process- ing (Cutler et al., 1986). Several studies have investigated the role of stress in word segmenta- tion using computational models, using both neu- ral network and “algebraic” (as opposed to “statis- tical”) approaches (Christiansen et al., 1998; Yang, 2004; Lignos and Yang, 2010; Lignos, 2011; Lig- nos, 2012). Bayesian models of word segmenta- tion (Brent, 1999; Goldwater, 2007), however, have until recently completely ignored stress. The sole exception in this respect is Doyle and Levy (2013) who added stress cues to the Bigram model (Gold- water et al., 2009), demonstrating that this leads to an improvement in segmentation performance. In this paper, we extend their work and show how to integrate stress cues into the flexible Adaptor Gram- mar framework (Johnson et al., 2007). This allows us to both start from a stronger baseline model and to investigate how the role of stress cues interacts with other aspects of the model. In particular, we find that phonotactic cues to word-boundaries inter- act with stress cues, indicating synergistic effects for small inputs and partial redundancy for larger in- puts. Overall, we find that stress cues add roughly 6% token f-score to a model that does not account for phonotactics and 4% to a model that already in- corporates phonotactics. Relatedly and in line with the finding …

Content truncated to 500 words