Dependency graph enhanced interactive attention network for aspect sentiment triplet extraction

Abstract

Aspect sentiment triplet extraction is an extremely daunting task designed to identify the triplets from comments, where each triplet is composed of an aspect term, the related opinion term, and the sentiment between them. Existing research efforts majorly construct a novel tagging scheme to avoid the disadvantages of pipeline methods. However, the improvement is limited due to neglecting the implicit grammatical relationships among the three elements in a triplet. To cope with this limitation, we put forward an innovative Dependency Graph Enhanced Interactive Attention Network, which explicitly introduces the syntactic and semantic relationships between words. Specifically, an interactive attention mechanism is conceived to jointly consider both the contextual features learned from Bi-directional Long Short-Term Memory and the syntactic dependencies learned from the correspondent dependency graph in an iterative interaction manner. In addition, we notice that words with different Part-of-Speech categories have different contributions to the semantic expression of sentences. Accordingly, the information of different Part-of-Speech categories is recognized during the modeling process to properly capture the semantic relationships. Experiments on the benchmark datasets originally derived from SemEval Challenges illustrate that our presented approach has superiority over strong baselines.