An Approach Based on Semantic Relationship Embeddings for Text Classification

Semantic relationships between words provide relevant information about the whole idea in texts. Existing embedding representation models characterize each word as a vector of numbers with fixed length. These have been used tasks involving text classification, such recommendation and question–answer systems. However, embedded provided by semantic has neglected. Therefore, this paper proposes an approach that involves for which is evaluated. Three based on relations extracted from Wikipedia are presented compared existing word-based models. Our considers following relationships: synonymy, hyponymy, hyperonymy. They were considered since previous experiments shown they knowledge. The using lexical-syntactic patterns identified literature. vector: hyponymy–hyperonymy, combination all relationships. A Convolutional Neural Network relationship embeddings was trained classification. An evaluation carried out proposed configurations to compare them two corpora. results obtained metrics precision, accuracy, recall, F1-measure. best 20-Newsgroup corpus hyponymy–hyperonymy embeddings, achieving accuracy 0.79. For Reuters corpus, F1-measure recall 0.87 synonymy–hyponymy–hyperonymy.