Learning to segment speech using multiple cues: A connectionist model

Considerable research in language acquisition has addressed the extent to which basic aspects of linguistic structure might be identified on the basis of probabilistic cues in caregiver speech to children. This type of learning mechanism presents classic learnability issues: there are aspects of language for which the input is thought to provide no evidence, and the evidence that does exist tends to be unreliable. We address these issues in the context of the specific problem of learning to identify lexical units in speech. A simple recurrent network was trained on a phoneme prediction task. The model was explicitly provided with information about phonemes, relative lexical stress, and boundaries between utterances. Individually these sources of information provide relatively unreliable cues to word boundaries and no direct evidence about actual word boundaries. After training on a large corpus of child-directed speech, the model was able to use these cues to reliably identify word boundaries. The model shows that aspects of linguistic structure that are not overtly marked in the input can be derived by efficiently combining multiple probabilistic cues. Connectionist networks provide a plausible mechanism for acquiring, representing, and combining such probabilistic information.