Reconstructed semantic relative distance and global and local attention fusion network for aspect-based sentiment analysis

Aspect-based sentiment analysis aims to analyze the sentiment tendencies towards a specific aspect within a given sentence. As a fine-grained sentiment classification task, it plays an integral role in detecting users’ comments. Recent studies have used relational labels in dependency trees to focus on aspect items in local contexts. However, opinion words in context are affected by irrelevant dependency labels, which can interfere with their accurate evaluation. Moreover, the combination of feature sequences with long and short-distance dependencies has not been thoroughly explored. To this end, we propose a reconstructed semantic relative distance and global and local attention fusion network (RAGN), which can extract syntactic and semantic features and fully fusing feature vectors from multiple modules. Firstly, the dependency distance in the context dynamic weights layer is replaced with the reconstructed semantic relative distance, which is recalculated based on the relational labels in a syntactic dependency tree rooted in aspects. Secondly, a global and local attention fusion network captures long-distance dependencies and emphasizes parts of sentences with salient sequence features. Ultimately, combining the aspect sentiment classification task (ASC) and the aspect entity recognition task (AER) and utilizing AER as an auxiliary task facilitates the final classification of ASC. Experimental results on three publicly available datasets verify the superiority, effectiveness, and robustness of the proposed model.