Pattern Mining for Information Extraction Using Lexical, Syntactic and Semantic Information: Preliminary Results

Abstract

A method is being developed to mine a text corpus for candidate linguistic patterns for information extraction. The candidate patterns can be used to improve the quality of extraction patterns constructed by a pseudosupervised learning method--an automated method in which the system is provided with a high quality seed pattern or clue, which is used to generate a training set automatically. The study is carried out in the context of developing a system to extract disease-treatment information from medical abstracts retrieved from the Medline database. In an earlier study, the Apriori algorithm had been used to mine a sample of sentences containing a disease concept and a drug concept, to identify frequently occurring word patterns to see if these patterns could be used to identify treatment relations in text. Word patterns and statistical association measures alone were found to be insufficient for generating good extraction patterns, and need to be combined with syntactic and semantic constraints. In this study, we explore the use of syntactic, semantic and lexical constraints to improve the quality of extraction patterns.