Aspect-Level Sentiment Analysis Using a Syntax-Enhanced and Semantically Informed Graph Convolutional Network

Aspect-level sentiment analysis seeks to determine the sentiment orientation towards aspects within a sentence. This task's primary difficulty lies in accurately defining the connection between aspects and sentiments, thus eliminating opinion words that do not pertain to the aspects. Over recent years, sentiment analysis based on aspects models using Graph Convolutional Networks (GCNs) have advanced considerably. Nevertheless, current models still have limitations, such as relying solely on a single dependency relation and ignoring semantic similarities between words and other syntactic structures. The lack of prior knowledge can lead to misinterpretation of relevant sentiment words, making it necessary to incorporate external knowledge to enrich the textual information. To resolve the issues mentioned above, this paper presents a Syntax and Semantics-Enhanced Graph Convolutional Network (SES-GCN) model for sentiment analysis. The model first utilizes dependency parsing to reveal the dependency relationships between words, especially long-distance dependencies, capturing the relationships among specific aspects and sentiment terms. It then uses a syntactic structure tree to provide hierarchical information of phrases and clauses within the sentence. Ultimately, a hybrid attention mechanism is employed to extract the semantic information of the sentence while factoring in diverse syntactic structures and semantic attributes. Experimental results on four public datasets show that the SES-GCN model surpasses the performance of baseline methods.