Aspect-level Sentiment Analysis Based on Semantic and Affective Dependency Enhancement

Graph neural networks based on the dependency tree can use rich syntactic information widely used in aspect-based sentiment analysis (ABSA). However, most of these models focus on the syntactic dependencies of sentences and lack the modeling of affective dependencies between aspects and context, which can clearly express affective expressions related to specific aspects in a sentence. Additionally, although graph neural networks based on the dependency tree can aggregate more attribute dependency information through the graph structure, they need to pay attention to the semantic relationship between ordered words, resulting in insufficient mining of semantic information. To address those issues, we propose a semantic and affective dependency enhancement model (SADE) in this paper, which can better handle emotional context knowledge for specific aspects and context semantic information. Specifically, the SADE model proposed an aspect-aware attention mechanism combined with self-attention to learn the aspect-related semantic correlations and the global semantics of the sentence. Additionally, SADE incorporates the senticWordNet lexical resource and dynamically considers affective dependencies between specific aspect terms and contexts in a particular domain, thereby obtaining affective knowledge-enhanced dependency information. The SADE model's experimental results demonstrate that it performs best on two public benchmark datasets.