Segment predictability as a cue in word segmentation

Abstract

Several computational simulations of how children solve the word segmentation problem have been proposed, but most have been applied only to a limited number of languages. One model with some experimental support uses distributional statistics of sound sequence predictability (Saffran et al. 1996). However, the experimental design does not fully specify how predictability is best measured or modeled in a simulation. Saffran et al. (1996) assume transitional probability, but Brent (1999a) claims mutual information (MI) is more appropriate. Both assume predictability is measured locally, relative to neighboring segment-pairs. This paper replicates Brent's (1999a) mutualinformation model on a corpus of childdirected speech in Modern Greek, and introduces a variant model using a global threshold. Brent's finding regarding the superiority of MI is confirmed; the relative performance of local comparisons and global thresholds depends on the evaluation metric.

Adding generalization to statistical learning: The induction of phonotactics from continuous speech

“…Computational models have shown this interpretation to be effective in segmentation (e.g., Brent, 1999a). A trough-based approach, however, is not capable of extracting unigram words, since such words would require two adjacent local minima (Rytting, 2004;Yang, 2004). The implication is that the learner is unable to discover monosyllabic words in case of syllable-based statistical learning (Yang, 2004), or single-phoneme words in case of segment-based statistical learning (Rytting, 2004).…”

“…Optimal candidate mutual information (MI), which corresponds to the log 2 value of the observed/expected ratio, and which has been used in several computational segmentation studies (Brent, 1999a;Rytting, 2004;Swingley, 2005). The difference between transitional probability and observed/expected ratio (or MI) lies in the directionality of the dependency (Brent, 1999a).…”

“…Two different interpretations exist with respect to how co-occurrence probabilities affect speech segmentation (Rytting, 2004). suggest that word boundaries are hypothesized at troughs in transitional probability.…”

“…regardless of neighboring bigrams. The interpretation of probabilities in isolation has been modeled either by inserting boundaries at points of low probability (Cairns et al, 1997;Rytting, 2004), or by clustering at points of high-probability (Swingley, 2005). In both cases the learner needs to use a threshold on the probabilities in order to determine when a boundary should be inserted, or when a cluster should be formed.…”

“…For the statistical learning models, we use a threshold-based segmentation strategy (Cairns et al, 1997;Rytting, 2004;Swingley, 2005). A segmentation threshold can be derived from the properties of the statistical distribution.…”