Getting the Input Right: Refining our Understanding of What Children Hear

摘要

As models for language learning become increasingly sophisticated, it is essential to pay close attention to the purported input received by learners. This talk presents two examples of phonological input for which a failure to account for relevant factors has led to misleading conclusions. A fully annotated dataset of infant-directed speech is now allowing a more refined analysis of what children actually hear. The first example concerns vowel-duration contrasts (long vs. short) in Japanese. One previous study, working under the assumption that long and short vowels occurred with equal frequency, concluded that the distinction could be learned by a simple distributional model. Our dataset, however, reveals that (a) in reality over 90% of vowels in Japanese are short, and (b) the distribution of long vowel duration is entirely encompassed within that of short vowels. The second example concerns the widely accepted claim that when adults speak to infants (infant-directed speech, JDS), they speak with a slower speech rate than when speaking to adults (adult-directed speech, ADS). Studies supporting this conclusion, however, have consistently failed to account for the fact that IDS utterances are shorter than those of ADS. Our dataset differentiates between utterance-internal speech rate and utterance-final lengthening, and finds taken separately, these values almost identical between IDS and ADS. As it turns out, IDS appeared to have a slower overall rate only because of the greater frequency of utterance-final segments.