Domain terms validation by means of a fuzzy lexical-semantic base

Abstract

Term extraction or recognition searches a given corpus to provide a list of domain specific terms for further use in more advanced tasks as in terminology and ontology building. Several statistical measures and Natural Language Processing techniques have been researched to improve precision of retrieved lists. However, to keep recall high, lists contain a number of false positives. To validate candidates as true positives in the domain, terms have to be manually evaluated or automatically checked against external resources such as specialized glossaries. Starting with a baseline of 50 candidate terms with 52% precision, we perform a series of experiments to show that a lexical knowledge base can significantly improve glossary performance. Furthermore, using a fuzzy lexical base, words clustered by a semantic association value, we research cutting points to reach 100% rates for either precision or recall for the baseline list, while keeping F-Measure > 80%, achieving 90% as best result. We conclude that, considering further research for limits and different case scenarios is also needed, a fuzzy lexical base can improve current state-of-the art approaches in automatic term extraction .