Semantic Relation Extraction by Analysis of Terms Correlation in Documents

Ontologies are important to organize and describe information, but are hard to create and maintain, which motivates the development of tools to help in this task. This article presents a strategy to extract, from a corpora of documents in a given domain, semantic elements expressing proximity relations between terms and concepts to help the construction of domain ontologies. The technique presented here, ACT, is based on linguistic processing, machine learning, and biclustering. Results show that concepts obtained by ACT are at least as good as those from similar techniques, such as LSI and NMF. In relation to those techniques, it additionally has the advantage of allowing the supervision by a domain expert.