SPOKEN WORD ACCESS PROCESSES (SWAP) May 29-31, 2000 Nijmegen, The Netherlands

Lexical segmentation and ambiguity: Investigating the recognition of onset-embedded words

Matt H. Davis (1), M. Gareth Gaskell (2), William D. Marslen-Wilson (1)

(1) MRC Cognition & Brain Sciences Unit, Cambridge
(2) Department of Psychology, University of York

Connected speech contains few reliable acoustic markers of word boundaries. Temporary ambiguities may therefore exist between short and long words in connected speech. For instance, a word like "cap" may be ambiguous with the start of longer words in which it is embedded like "captain" and "captive" (McQueen, Norris and Cutler, 1994). The prevalence of onset-embedded words has been used to argue for accounts of spoken word recognition in which following context can be used to resolve this ambiguity. Computational models of spoken word recognition such as Trace (McClelland and Elman, 1986) therefore incorporate direct, inhibitory connections between lexical items. Lexical competition allows speech that rules out longer competitors to increase the activation of embedded words - even where mismatching input arrives after the acoustic offset of an embedded word.

However, data from cross-modal repetition-priming experiments (Davis, Marslen-Wilson and Gaskell, 1997) suggest that acoustic differences between syllables in short and long words help distinguish embedded words from longer competitors. In these experiments, greater priming of embedded words (such as "cap") was observed from the syllable of a short word ("cap") compared to the equivalent syllable from a longer word (e.g. /k{p/ from "captain"). Conversely greater priming of the target "captain" was observed from the initial syllable of the word "captain" than from the short embedded word "cap". These results show that listeners favour the correct interpretation of short and long words before the offset of the embedded syllable and suggest that following context is not the only cue involved in the identification of onset-embedded words. Consequently, models in which lexical competition and following context are necessary for the identification of onset-embedded words provide an incomplete account of the identification of embedded words in connected speech.

In this paper we present a recurrent network model of spoken word recognition in which following context can assist the identification of onset-embedded words without direct inhibitory connections between units at the lexical level. Simulations that incorporate input cues analogous to the acoustic differences between syllables in short and long words show differences between the activation of embedded words and longer competitors consistent with the priming data of Davis et al. (1997). These results indicate that recurrent neural networks can be used to model the integration of lexical and acoustic cues in the recognition of embedded words without direct competition at the lexical level. 

References

McClelland, J. L., & Elman, J. L. (1986). The TRACE model of speech perception. Cognitive Psychology, 18, 1-86. 

McQueen, J. M., Norris, D., & Cutler, A. (1994). Competition in spoken word recognition: spotting words in other words. Journal of Experimental Psychology: Learning, Memory and Cognition, 20, 621-638.

Full Paper

Bibliographic reference.  Davis, Matt H. / Gaskell, M. Gareth / Marslen-Wilson, William D. (2000): "Lexical segmentation and ambiguity: Investigating the recognition of onset-embedded words", In SWAP-2000, 71-74.