Emotion-cause pair extraction (ECPE), which aims to extract emotions and the corresponding causes in documents, has a wide range of applications in network public opinion analysis. Current two-stage methods first extract emotion and cause clauses, and then pair them. However, there are two problems in these methods: 1) the unidirectional enhancement between emotion and cause extraction fails to make full use of the correlation between them; 2) the errors from the first stage directly degrade the performance of the second stage. To address these problems, we firstly propose a mutually auxiliary multitask model to promote the extraction of emotion and cause clauses by adding two auxiliary tasks which are identical to the original tasks. The proposed model uses the predicted results generated by the two auxiliary tasks as extra features of each other’s main tasks, so as to establish the bidirectional correlation between emotion and cause extraction. Secondly, to reduce the influence of error propagation on the second stage, we design a self-distillation method for pairwise tasks to train the proposed model, which further improve the accuracy of emotion and cause extraction. Experimental results on the ECPE benchmark dataset show that the proposed model has achieved good performance on emotion-cause pair extraction, outperforming the baseline models by 1.92% in F1 score.

Recently, graph neural networks (GNN), due to their compelling representation learning ability, have been exploited to deal with emotion-cause pair extraction (ECPE). However, current GNN-based ECPE methods mostly concentrate on modeling the local dependency relation between homogeneous nodes at the semantic granularity of clauses or clause pairs, while they fail to take full advantage of the rich semantic information in the document. To solve this problem, we propose a novel hierarchical heterogeneous graph attention network to model global semantic relations among nodes. Especially, our method introduces all types of semantic elements involved in the ECPE, not just clauses or clause pairs. Specifically, we first model the dependency between clauses and words, in which word nodes are also exploited as an intermediary for the association between clause nodes. Secondly, a pair-level subgraph is constructed to explore the correlation between the pair nodes and their different neighboring nodes. Representation learning of clauses and clause pairs is achieved by two-level heterogeneous graph attention networks. Experiments on the benchmark datasets show that our proposed model achieves a significant improvement over 13 compared methods.