Syntax-enhanced aspect-based sentiment analysis with multi-layer attention

Abstract

As a key task of fine-grained sentiment analysis, aspect-based sentiment analysis aims to analyse people’s opinions at the aspect level from user-generated texts. Various sub-tasks have been defined according to different scenarios, extracting aspect terms, opinion terms, and the corresponding sentiment. However, most existing studies merely focus on a specific sub-task or a subset of sub-tasks, having many complicated models designed and developed. This hinders the practical applications of aspect-based sentiment analysis. Therefore, some unified frameworks are proposed to handle all the subtasks, but most of them suffer from two limitations. First, the syntactic features are neglected, but such features have been proven effective for aspect-based sentiment analysis. Second, very few efficient mechanisms are developed to leverage important syntactic features, e.g., dependency relations, dependency relation types, and part-of-speech tags. To address these challenges, in this paper, we propose a novel unified framework to handle all defined sub-tasks for aspect-based sentiment analysis. Specifically, based on the graph convolutional network, a multi-layer semantic model is designed to capture the semantic relations between aspect and opinion terms. Moreover, a multi-layer syntax model is proposed to learn explicit dependency relations from different layers. To facilitate the sub-tasks, the learned semantic features are propagated to the syntax model with better semantic guidance to learn the syntactic representations comprehensively. Different from the conventional syntactic model, the proposed framework introduces two attention mechanisms. One is to model dependency relation and type, and the other is to encode part-of-speech tags for detecting aspect and opinion term boundaries. Extensive experiments are conducted to evaluate the proposed novel unified framework, and the experimental results on four groups of real-world datasets explicitly demonstrate the superiority of the proposed framework over a range of baselines.