Word segmentation: The role of contrast

Abstract

Phonotactics are know to be a potential cue for word boundaries (Mattys & Jusczyk, 2001; Mattys et al., 1999); however, to learn and make use of phonotactics in this way, an infant must both perceive native language phones correctly most of the time, and also categorize them in an adult-like way. The role of perception in phonology has been a focus of study in both diachronic and synchronic phonology (Ohala, 1981; Steriade, 2001). That ease of perception should be a driving force for the organization of phonological grammar is nothing new. However, its role in word segmentation has not been examined in detail. Given that phonotactic cues may be important for word segmentation, it is possible that failure to perceive certain contrasts could be devastating to the learner, in which case phonotactic cues would hardly be practical as reliable cues for word boundaries. Alternatively, it could be that those contrasts which are least perceptable cause less damage to the learner’s ability to reliably locate word boundaries. I test this hypothesis on a supervised computational learning model, a modified version of DiBS Daland (2009), altering the input to take away specific contrasts. As a case study, I then ran input neutralized for place of articulation and input neutralized for voicing through the unsupervised learner StaGe Adriaans & Kager (2010). From these modeling tests, I conclude that a) a rich inventory of phones (ie, types) is not as crucial to segmentation as one would think; b) of contrasts tested, place of articulation is the least useful for phonotactic constraints as cues for word boundaries in Dutch, and c) there is a correlation between perceptual salience of a contrast and the usefulness of that contrast in cueing word boundaries.