Learning feature-value grammars from plain text
NeMLaP3/CoNLL '98 Proceedings of the Joint Conferences on New Methods in Language Processing and Computational Natural Language Learning
| Hi-index | 0.00 |
This paper describes the theory and implementation of Babel, a system which explores the hypothesis that much of the differences in the world's languages may be characterized by the inventory and properties of the lexical items and functional categories of those languages. The structure of Babel assumes that functional categories are originally lacking in a child's syntax, and are acquired through a statistical induction process of lexical acquisition. Babel then uses information induced from the structure of the lexicon to create a model of syntax via a deductive, rule-based process. This model makes a number of predictions about the time course of language acquisition. These predictions are tested by running Babel as a simulation of child language acquisition, using large samples of adult speech to children as input. The simulation results are shown to highly correlate to longitudinal studies of child language acquisition in English and Polish. Finally, the approach to handling noisy data with Babel is detailed.