Reasoning over multiplex heterogeneous graph for Target-oriented Opinion Words Extraction

Abstract

Target-oriented Opinion Word Extraction (TOWE) is a new emerging subtask of Aspect Based Sentiment Analysis (ABSA), which aims to extract fine-grained opinion terms for a given aspect term from a sentence. In this task, the key point is how to find the correct opinion that is far away from its corresponding aspect. Ideally, reinforcement learning (RL) seems to be a promising approach due to its delayed reward mechanism. However, as aspect-opinion interaction data is likely to be complicated, it is not easy to directly apply RL techniques to improve the performance. In this paper, we propose a novel Padding-Enhanced Reinforcement learning model (PER) to address this issue. Specifically, PER first designs a multiplex heterogeneous graph to cover both sequential structure and syntactic structure in order to enrich their interactions and alleviate the long distance issue. By formulating the extraction task as a Markov Decision Process (MDP), PER then walks on the designed graph to infer corresponding opinions for each aspect. In addition, a padding module is further designed to aggregate rich information from distant nodes to guide the exploration process. Extensive experimental results on four widely used datasets illustrate that our proposed model consistently outperforms the state-of-the-art methods.