Word segmentation: TP or OCP? A re-analysis of Batterink & Paller (2017)

Abstract

Research on statistical learning suggests that to segment speech into words, infants keep track of transitional probabilities (TPs) between syllables: the likelihood that syllable X occurs given syllable Y. TPs between neighboring syllables within words are higher than TPs at word boundaries. Batterink and Paller (2017) measured neural oscillations with EEG during statistical learning, which are known to phase-lock to the rhythm of an auditory stimulus. In the study of Batterink and Paller (2017), participants listened to a structured stream, consisting of four tri-syllabic words (TPs within words: 1.0, between: 0.33), and a random stream (TPs 0.09). Exposure to the structured stream but not the random stream led to an increase of phase-locking to the word frequency (1.1 Hz), compared to the syllable frequency (3.3 Hz).

However, some participants unexpectedly segmented the random stream into tri-syllabic units as well. The current study provides an alternative explanation for the findings of Batterink and Paller (2017) through the Obligatory Contour Principle (OCP) with a constraint on place of articulation (OCP-PLACE). Boll-Avetisyan and Kager (2014) showed that OCP-PLACE can influence word segmentation in Dutch. We performed a data re-analysis of Batterink and Paller (2017), replicating their analysis with Linear Mixed Modelling (LMM) and investigating the OCP-PLACE constraint as a possible alternative explanation of the data, including participants’ triplet segmentation in the random stream.

We confirmed the statistical robustness of the results found by B&P2017, reporting the same results with our LMM approach as their ANOVA. Furthermore, we found a significant effect of OCP that is parallel to the effect of condition in the data of B&P2017. Further research should investigate the independent effects of OCP-PLACE on word segmentation in English and consider OCP-PLACE as a possible confounder that should be controlled for in further statistical language learning experiments.