SSRGAT: integrating syntactic, semantic, and contextual knowledge in a knowledge-infused graph model for aspect-based sentiment analysis

Aspect-based sentiment analysis is a subfield of sentiment processing that enables the determination of a sentence’s polarity based on its aspects. A fundamental challenge in aspect-based sentiment analysis is the presence of multiple aspects with varying polarities within a single sentence. Recent studies have attempted to extract the relationships between each pair of aspects and sentences using graph convolutional methods and dependency trees. However, given the complexities of natural language and the number of aspects in a sentence, these models have struggled to construct accurate sentence vectors. Furthermore, the performance of knowledge-based models that rely on syntactic structures, such as dependency trees, is limited, as many texts do not adhere to grammatical rules, thereby decreasing the model’s efficacy. In this study, we propose a novel model called SSRGAT that leverages syntactic, contextual, and semantic information by constructing a semantic tree from an external knowledge base and generating a pseudo-semantic representation for each word using graph node embeddings. Also, we have examined a gating mechanism as the strategy of integrating multiple knowledge sources. Besides, a cross-attention mechanism is used to determine the impact of diverse inter-word relationships in determining the polarity. Experiments conducted on several benchmark datasets demonstrate that the proposed model enhanced the performance of other approaches in terms of in both accuracy and robustness.