Maier, R. & Fausey, C. M.
University of Oregon
People talk about coherent episodes of their experience, leading to strong dependencies between words and the contexts in which they appear (Goldwater, Griffiths, & Johnson, 2009; Mintz, 2003; Roy, Frank, & Roy, 2012). Here, we ask whether this fact also yields context-specific distributions of transitional probabilities between syllables - a statistic hypothesized to help learners discover words in the first place (Saffran et al., 1996). This is important because the status of transitional probabilities as a cue to word segmentation in infant-directed speech is unclear (Swingley, 2005; Gambell & Yang, 2006). We consider the transitional probabilities that are available within typical contexts (e.g., mealtime, diaper change, bedtime) sampled from a large corpus of infant-directed speech (Korman, 1984). We find that the distribution of these probabilities within each context differs significantly from size-matched "nontexts" randomly sampled from the same corpus (Kolmogorov-Smirnov tests, p<.001). Word boundaries may be clearer within-context because frequent syllables (e.g., milk) co-occur with a range of subsequent syllables (milk hmm, milk to, milk in) in that context. The structure provided by coherent contexts may allow learners to use a cue that may not otherwise rise as signal out of noise in the full complexity of infant-directed speech.