Path-Enhanced Multi-hop Graph Attention Network for Aspect-based Sentiment Analysis

Aspect-based Sentiment Analysis is a task aims to identify the sentiment polarities of given aspects. The core of this task consists of distinguish and understand in complex sentences. Most recent work used attention-based neural network models to extract information of words and their connection. However, over-smoothing often occurs due to the complexity of the language. In this paper, we improve this problem by means of path enhancement. We reconstruct the dependency tree to fit for the model. Then, we propose a path-enhanced multi-hop graph attention network model. We conduct experiments on the SemEval 2014 dataset, and the experimental results show that our method improves the graph attention network significantly.