A Semantic-Proximity Term-Weighting Scheme for Aspect Category Detection

Resumen

Aspect category detection is a subtask of aspect-level sentiment analysis, which aims at identifying the aspect categories present in an opinion. It is a difficult task because the category must be inferred from the terms of the opinion, and also because each opinion may include judgments for more than one aspect category. In recent years, the use of attention mechanisms has improved performance in different tasks, allowing the identification and prioritization of terms that mostly contribute to the classification. However, in multi-label problems, such as aspect category detection,different terms must be selected based on each category, which is a drawback for these models. Motivated by the same idea of identifying and highlighting the importance of terms, this paper proposes a weighing scheme that emphasizes terms in an opinion based on their semantic proximity to each aspect category. The proposed scheme has been evaluated on different SemEval datasets, demonstrating its effectiveness in this multi-label scenario. Moreover, it can be applied in scenarios with limited training data and can be combined with different classification models, including deep neural networks.