How to combine text-mining methods to validate induced Verb-Object relations?

Abstract

This paper describes methods using Natural Language Processing approaches to extract and validate induced syntactic relations (here restricted to the Verb-Object relation). These methods use a syntactic parser and a semantic closeness measure to extract such relations. Then, their validation is based on two different techniques: A Web Validation system on one part, then a Semantic-Vectorbased approach, and finally different combinations of both techniques in order to rank induced Verb-Object relations. The Semantic Vector approach is a Roget-based method which computes a syntactic relation as a vector. Web Validation uses a search engine to determine the relevance of a syntactic relation according to its popularity. An experimental protocol is set up to judge automatically the relevance of the sorted induced relations. We finally apply our approach on a French corpus of news by using ROC Curves to evaluate the results.