Aspect-Level Semantic and Syntactic Reinforcement for Aspect-Based Sentiment Analysis
Shiqi Wang, Hongliang Dong, Zhiyi Fang, Chongkuan Chen · 2022
Aspect-based sentiment analysis (ABSA) is a fine-grained task whose main target is to identify the sentiment polarity associated with a given aspect in a sentence. Its main challenge is aspect feature extraction of sentences. For the problem of insufficient aspect-level information extraction, such as long dependencies problems, we propose a novel model that enhances aspect information from semantic and syntactic perspectives to address this issue. In terms of semantics, we provide a mechanism to actively mask aspect words and enable the model to capture specific properties in context. In terms of syntax, we advance a new aspect weight graph that can directly connect aspect words with syntactically related words. Furthermore, we integrate semantic extraction and syntactic analysis methods to enhance the informative representation of aspect-level words. We evaluate the effectiveness of our method on publicly available datasets. Experimental results show that our model achieves excellent results on the ABSA tasks.