Levels of Representation in Language Development

In this paper we provide a computational exploration of changes in levels of representation in children’s language development. Children demonstrate a progression from awareness of larger units in their language, such as words or syllables, and only later do they develop the ability to manipulate phoneme level representations. We employed a minimum description length approach to encode a corpus of child-directed speech. The analysis indicated that, for early stages of language learning, a word level representation was most efficient, whereas after extended exposure, a phoneme level representation was more efficient. The analysis also accurately predicted which phonemes children would be able to distinguish from syllables at different stages of development.