Sentence Dependent-Aware Network for Aspect-Category Sentiment Analysis

The purpose of Aspect-Category Sentiment Analysis is to predict sentiment polarities of given aspect categories in sentences. Most previous methods used attention-based neural network models to Establish connections between aspect categories and sentiment words and generate aspect-specific sentence representations. However, these models may mismatch sentiment words with aspect categories due to the complexity of sentence structures. To solve this problem, we reconstruct the dependency tree into an ACSA-oriented dependency tree, which builds a direct or indirect semantic connection between sentiment words and corresponding aspect categories, and avoid introducing redundant information from the original dependency tree. On this basis, we propose a Sentence Dependent-Aware Network (SDAN) to encode the tree effectively. The experimental results of applying SDAN to three public datasets demonstrate its effectiveness.