Avaliando a utilização de Redes de Grafos com Atenção para a tarefa de Análise de Sentimentos em Nível de Aspecto em Línguas de Baixo Recurso

Abstract

Opinions about products and services hold immense importance for the industry, as they aid in the decision-making process. However, due to the continuous growth of user-generated content, particularly on the internet, the manual analysis of such data has become unfeasible. Sentiment Analysis techniques are essential for understanding and quantifying human sentiments expressed through different types of data, including audio, video, and images. Nonetheless, a substantial part of these analyses is conducted on textual data, which is the focus of this study. Among the various levels of analysis, aspect-level analysis is the most granular, as it examines all parts of an entity where a sentiment can be assigned. The use of graphs to represent textual data allows for the integration of the grammatical structural relationships of the text, enriching the representation and potentially benefiting the task f Aspect-Based Sentiment Analysis (ABSA). This approach not only incorporates information derived from syntactic analysis of the text but also enables the application of new techniques, expanding the possibilities of analysis. Although studies have demonstrated the effectiveness of this representation for ABSA using Graph Neural Networks in English, there is limited evidence of improvements with these techniques in low resource languages such as Portuguese. We developed a Graph Attention Network model for the ABSA task in Brazilian Portuguese. The proposed approach achieves a balanced accuracy of 0.74, presenting competitive results and achieving a performance equivalent to the third place in the ABSAPT competition.