DCN-ECPE: Dual-Channel Network for Emotion-Cause Pair Extraction

It is a challenge task to extract the potential pairs of emotion clause and corresponding cause clause from the documents. The existing state-of-the-art ECPE method formulates the task in an end-to-end model, which processes the interactions of emotion-cause pairs based on joint two-dimensional. The model has two shortcomings: 1) the potential semantic particularity of the causal relation between the emotion-cause pair is not fully considered; 2) it falls short of capturing various regional features of contextualized representation. In this work, we propose an end-to-end model named DCN-ECPE. The model generates the representation of emotion-cause pairs with dual-channel, which takes both potential causal features and contextualized interactions of the clause pairs into consideration. One channel extracts potential semantic feature of the causal relation from constructed statements, and the other channel processes the representation of clause pairs with CNN to capture various regional features. Our method outperforms existing state-of-the-art end-to-end ECPE method in all aspects.