Semantic Textual Similarity on Brazilian Portuguese: An approach based on language-mixture models

Abstract

The literature describes the Semantic Textual Similarity (STS) area as a fundamental part of many Natural Language Processing (NLP) tasks. The STS approaches are dependent on the availability of lexical-semantic resources. There are several efforts to improve the lexicalsemantics resources for the English language, and the state-of-art report a large amount of application for this language. Brazilian Portuguese linguistics resources, when compared with English ones, do not have the same availability regarding relation and contents, generation a loss of precision in STS tasks. Therefore, the current work presents an approach that combines Brazilian Portuguese and English lexical-semantics ontology resources to reach all potential of both language linguistic relations, to generate a language-mixture model to measure STS. We evaluated the proposed approach with a well-known and respected Brazilian Portuguese STS dataset, which brought to light some considerations about mixture models and their relations with ontology language semantics.