Learning Emotion-Aware Contextual Representations for Emotion-Cause Pair Extraction

Emotion Cause Pair Extraction (ECPE), the task expanded from the previous emotion cause extraction (ECE) task, focuses on extracting emotion-cause pairs in text. Two reasons have made ECPE a more challenging, but more applicable task in real world scenarios: 1) an ECPE model needs to identify both emotions and their corresponding causes without the annotation of emotions. 2) the ECPE task involves finding causes for multiple emotions in the document context, while ECE is for one emotion. However, most existing methods for ECPE adopt an unified approach that models emotion extraction and cause extraction jointly through shared contextual representations, which is suboptimal in extracting multiple emotion-cause pairs. In addition, previous ECPE works are evaluated on one ECE dataset, which exhibits a bias that majority of documents have only one emotion-cause pair. In this work, we propose a simple pipelined approach that builds on two independent encoders, in which the emotion extraction model only provide input features for the cause extraction model. We reconstruct the benchmark dataset to better meet ECPE settings. Based on experiments conducted on the original and reconstructed dataset, we validate that our model can learn distinct contextual representations specific to each emotion, and thus achieves state-of-the-art performance on both datasets, while showing robustness in analyzing more complex document context.