A BERT-based Hierarchical Model for Vietnamese Aspect Based Sentiment Analysis

Abstract

Aspect based sentiment analysis (ABSA) is the task of identifying sentiment polarity towards specific entities and their aspects mentioned in customers’ reviews. This paper presents a new and effective hierarchical model using the pre-trained language model, Bidirectional Encoder Representations from Transformers (BERT). This model integrates the context information of the previous layer (i.e. entity type) into the prediction for the following layer (i.e. aspect type) and optimizes the global loss functions to capture the entire information from all layers. Experimental results on two public benchmark datasets in Vietnamese showed that the proposed model is superior to the existing ones. Specifically, the model achieved 84.23% and 82.06% in the F1_micro scores in detecting entities and their aspects on the domains of restaurants and hotels, respectively. In identifying aspect sentiment polarity, the model gained 71.3% and 74.69% in the F1_micro scores on the domains of restaurants and hotels, respectively. These results outperformed the best submission of the campaign by a large margin and gained a new state of the art.