Exploring The Potentiality Of Semantic Features For Paraphrase Detection

Paraphrase is defined as the repetition of something written or spoken using different words. In this paper, we adopt a feature engineering strategy to perform paraphrase detection at the sentence level. In particular, we explore the potentiality of semantic features, as the similarity between two semantic graphs, a distance function between sentences and the cosine similarity between embedded sentences, using them within several machine learning-based classifiers. We evaluate our approach on the ASSIN benchmark corpus and achieve 80.5% of F-score, outperforming some other detection methods for Portuguese.