Erickson, L. C. & Thiessen, E. D. .
Both Carnegie Mellon University
Statistical learning (SL) is proposed to play an important role in language acquisition (e.g., statistical segmentation; Thiessen & Erickson, 2013). However, the extent to which natural language acquisition involves SL is unclear (e.g., Johnson & Tyler, 2010). If word forms in real language are discovered via SL, statistically segmented sequences should share properties with real lexical items. One feature exhibited by real words is the Ganong effect, in which lexical items shape the perception of ambiguous phonemes (Ganong, 1980; a sound ambiguous between /g/ and /k/ is perceived as /g/ in the context of [?]ift and /k/ in the context of [?]iss). We tested whether items segmented via statistical learning can also shape the perception of ambiguous phonemes. Participants segmented an artificial language on the basis of statistical structure. Subsequently, we tested whether items consistent with the input structure of the language biased their perception of ambiguous sounds (e.g., is the initial sound in [?]awroetee heard as /g/ if gawroetee but not kawroetee was a word in the language?). A word-consistent shift in the predicted direction was observed. These results are consistent with the possibility that statistical learning produces representations that share features with real words.