Learning Grammars for Noun Phrase Extraction by Partition Search

This paper describes an application of Grammar Learning by Partition Search to noun phrase extraction, an essential task in information extraction and many other NLP applications. Grammar Learning by Partition Search is a general method for automatically constructing grammars for a range of parsing tasks; it constructs an optimised probabilistic context-free grammar by searching a space of nonterminal set partitions, looking for a partition that maximises parsing performance and minimises grammar size. The idea is that the considerable time and cost involved in building new grammars can be avoided if instead existing grammars can be automatically adapted to new parsing tasks and new domains. This paper presents results for applying Partition Search to the tasks of (i) identifying flat NP chunks, and (ii) identifying all NPs in a text. For NP chunking, Partition Search improves a general baseline result by 12.7%, and a methodspecific baseline by 2.2%. For NP identification, Partition Search improves the general baseline by 21.45%, and the method-specific one by 3.48%. Even though the grammars are nonlexicalised, results for NP identification closely match the best existing results for lexicalised approaches.