AWA-Net: An Adaptive Window Attention Network for Aspect-Based Sentiment Analysis

Aspect-based sentiment analysis is a subtask of sentiment classification that identifies the fine-grained sentiment of a sentence. As a cognitive-inspired computational task, ABSA reflects the human cognitive process of understanding and evaluating opinions by associating emotions with specific targets. In recent years, graph neural networks that incorporating the syntactic structure of sentences for modeling have achieved good results. However, these methods rely excessively on syntactic dependency trees with sparse relations for feature learning, which leads to insufficient learning ability of the model’s local information. To address the above problems, we propose an Adaptive Window Attention Network (AWA-Net) model for Aspect-based sentiment analysis. First, the components of a sentence are analyzed using a syntactic parse tree to obtain the most informative phrase structure. We construct an adaptive window to retain information within the phrase structure while masking information outside this range. Then, the model learns the local information within the adaptive window through the attention mechanism. Finally, the proposed interaction module is used to interactively learn and fuse the local and syntactic features. This module mines deep and shallow feature information, thereby improving the accuracy of aspect word sentiment prediction. Experimental results on four public datasets indicate that the accuracy and F1 score of the AWA-Net model show noticeable improvements over the baseline models.