A reinforcement learning paradigm for Vietnamese aspect-based sentiment analysis

Abstract

This paper presents the task of aspect-based sentiment analysis (ABSA) that recognizes the sentiment polarity associated with each aspect of entities discussed in customers’ reviews, focusing on a low-resourced language, Vietnamese. Unlike conventional classification approaches, we leverage reinforcement learning (RL) techniques by formulating the task as a Markov decision process. This approach allows an RL agent to handle the hierarchical nature of ABSA, sequentially predicting entities, aspects, and sentiments by exploiting review features and previously predicted labels. The agent seeks to discover optimal policies by maximizing cumulative long-term rewards through accurate entity, aspect, and sentiment predictions. The experimental results on public Vietnamese datasets showed that the proposed approach yielded new state of the art (SOTA) results in both hotel and restaurant domains. Using the best model, we achieved an improvement of 1% to 3% in the F1 scores for detecting aspects and the corresponding sentiment polarity.