Contributions to Hypernym Patterns Representation and Learning based on Dependency Parsing and Sequential Pattern Mining

Hypernym relation is a semantic relationship between a specific term and a generic term of it. Approaches to extract such relations from texts have gained a large interest due to the huge availability of textual resources and their key role in ontology building. These approaches are commonly divided into two types: pattern-based and distributional. Patterns seem quite more interesting than distributional for building ontology due to their ability in extracting explicit relations from texts and their good precision. Thus, we focus on patterns and we describe an approach for systematically improving pattern performances. The approach is based on the coupled usage of sequential pattern mining and a specific pattern representation using grammatical dependencies to learn hypernym and anti-hypernym patterns. The results confirm that our approach can learn patterns that outperform unsupervised approaches and a supervised pattern-based approach. However, while the best performances are achieved by some supervised distributional approaches using word embedding, patterns can extract distinct hypernym relationships. This confirms that both types of approaches are complementary. Additionally, further experiments also confirm that the proposed approach tends to learn generically valid patterns across various corpora.