Combining complex networks and word embeddings in text classification tasks

Abstract

Complex networks have been used to model a myriad of complex systems. Even though this model has already been employed in text classification tasks, most of the studies are based on the co-occurrence model, which has some drawbacks. We propose here an extension of the traditional co-occurrence model by using information obtained from word embeddings. The proposed model includes additional virtual edges reflecting the similarity between words (vertices). The enriched networks are expected to provide an improved characterization of texts. As a consequence, we expect a gain in performance and robustness in the considered classification tasks. Because the proposed method is generic, the same approach could be applied to analyze any networked systems by using information encoded in node embeddings. (AU)