Triplet extraction network with dual gating mechanism and dependency-oriented attention

Aspect Sentiment Triplet Extraction (ASTE) refers to the identification of aspect terms, opinion terms that imply related sentiments, and sentiment polarity. The latest tendency to address this task is to exploit syntactic knowledge of sentences using Graph Convolutional Networks (GCNs), however this approach introduces noise that is not relevant to the extraction target and generates numerous features that lack a suitable processing strategy. Therefore, this paper proposes a dual-gating and dependency-oriented attention network for aspect-sentiment triplet extraction. The model applies BERT and graph convolutional network (GCN) to encode semantic and syntactic features respectively. In order to preliminary filter the noise and to distinguish the importance of different connections in syntactic dependencies, the model adopts a dependency-directed attention mechanism to encode syntactic knowledge. The obtained feature representations are then input into the dual gating mechanism for further syntactic filtering and multi-feature fusion. Experimental results demonstrate that our model performs significantly better than current similar baseline models.