Part-of-speech- and syntactic-aware graph convolutional network for aspect-level sentiment classification

Aspect-level sentiment classification is a task within the realm of fine-grained sentiment analysis that focuses on identifying the sentiment polarity specific to a particular aspect of review data. However, most existing methods fail to account for the unique expression and language style used in review data, which limits their sentiment classification performance. To address this issue, we propose a novel method called Part-of-Speech- and Syntactic-Aware Graph Convolutional Network (PSA-GCN) that aims to integrate part-of-speech (POS) information and syntactic knowledge into word embeddings. Specifically, we simplify the complex POS tags into six basic categories. At the same time, we only consider the connection relationship between words in the syntax tree that does not involve dependency types, in order to avoid the inaccuracies of overly-subdivided POS tags and the adverse effects caused by syntax dependencies of erroneous types on sentiment analysis. By incorporating these components into our model, the PSA-GCN is able to enhance the representation power of word embeddings and thus improve the performance in aspect-level sentiment classification. PSA-GCN first extracts part-of-speech tags and the syntactic parse tree to model the linguistic information present in the review data. It then considers the sentiment priors of different part-of-speech pairs holistically to construct a part-of-speech dependency graph, and a syntactic dependency graph utilizing the syntactic information from the parse tree. These graphs are initialized with Bert embeddings, and graph reasoning is performed to obtain the final part-of-speech and syntactic-aware language representation. Finally, aspect-level sentiment polarity is obtained through the classification of the final language representations. Our experiments on Restaurant, Laptop, and Twitter datasets reveal that PSA-GCN outperforms baseline models significantly in all three datasets.