Order-guided deep neural network for emotion-cause pair prediction

Abstract

Emotion-Cause Pair Extraction (ECPE) is a prediction task aiming to extract the emotions and their corresponding causes in a target document. The existing methods for this problem mainly focus on modeling the dependence between emotion clauses and related cause clauses and the interaction among emotion-cause pairs. However, these methods ignore the order information between emotion clauses and their cause clauses, which can be proved useful for the ECPE task. In this paper, we propose an order-guided deep predictive model, which integrates different orders between emotion clauses and their cause clauses into an end-to-end framework to tackle this task. Specifically, we build an order-guided clause encoder with a three-level long-short term memory (LSTM) network to learn the different orders from forward LSTM, backward LSTM and Bi-LSTM, respectively. In this way, the deep networks with different directions can effectively capture different orders, and therefore improve the performance of our model in this prediction task. Additionally, the previous methods use only a shared word encoder to capture word-level emotion and cause information, resulting in paying more attention to emotion information and lacking the ability to capture cause information. In order to overcome this deficiency, we design both an emotion-aware word encoder and a cause-aware word encoder to enhance the ability to capture the emotion and cause information. The experiment results illustrate that our method outperforms the other baselines on two real-world datasets, and demonstrate the effectiveness of the proposed method.