Text Corpora, Local Grammars and Prediction

Proceedings of the Fourth International Conference on Language Resources and Evaluation (LREC 2004)

Abstract

A corpus-based method for identifying and learning patterns describing events in a specific domain by examining the manner in which: (a) a small number of keywords in the domain are distributed throughout the corpus; and, (b) a local grammar that is idiosyncratic of a class of events in the domain, governs the usage of the keywords. We used a 3.63 million words corpus, and the results are encouraging. More importantly, the method can be applied to any arbitrary domains.