Abstract

The pattern based methods are widely used in Information Extraction (IE). But the methods suffered from the limited coverage and difficulty in using various information. This paper presents a novel approach using multi-level patterns to identify Chinese event roles. The characters of the approach are: (1) Four levels of patterns are introduced and patterns of each level involve different source of information. (2) Patterns of the 3rd and 4th levels are soft patterns, which encompass flexible information, support fuzzy match. Experiments on ACE2005 corpus show that patterns of each level are benefit to the system, and the performances are improved continuously when patterns are added level by level. And Soft patterns are very useful for improving recall because of their flexible structures, strong expressive power and fuzzy match.

Paper Details