Graph attention network based detection of causality for textual emotion-cause pair

To solve the problem that the existing model cannot adequately express inter-sentence structural information, this paper proposes a textual Emotion-Cause Pair (ECP) causal relationship detection method (GAT-ECP-CD) fused with graph attention network (GAT). A structural relationship graph directly propagates causal features from the context to integrate syntactic dependency information between different sentences in a document. First, using a word-level Bidirectional Long Short-Term Memory (BiLSTM) network to obtain intraclause semantic representations respectively. Then, the independent sentence vector is sent to the GAT as a graph node to capture the local and global dependency information between clauses to obtain richer features. Finally, a multi-task learning module bridges the first and second stages for dynamic prediction. On the benchmark dataset, compared with the existing method, the F1 score is improved by 4.38%, which verifies the effectiveness of the proposed model.