Third International Conference on Spoken Language Processing (ICSLP 94)
We report in this paper our investigations aimed at evolving an approach to develop a grammar model for the speech of a child acquiring English as the first language. We use the occurrence of words in similar syntac-to-semantic environment as the criterion for classifying words into equivalence classes. Consequently, words in the same class are interchangeable in a phrase or larger context. We identify phrases on the basis of n-gram frequencies of 'class exemplars' in the corpus and their meaningfulness. We group them into a small number of interchangeable phrase types. It turns out that the corpus of 723 sentences consists of 161 distinct phrase type sequence patterns and that of these, just 70 patterns have a frequency more than one and account for as many as 632 sentences in the original corpus. We use an algorithmic procedure to build a 'State Transition Network' which accounts for the sentence patterns. The STN consists of 11 states and generates most of the sentences in the corpus. This indicates that the approach is effective for modeling the grammar of the child.
Bibliographic reference. Rao, P. V. S. / Bondale, Nandini (1994): "BSLP based language grammars for child speech", In ICSLP-1994, 1711-1714.