Span-level emotion-cause-category triplet extraction via table-filling

Emotion cause analysis has been a prominent focus in natural language processing. The existing studies mainly focus on the clause-level emotion cause extraction (ECE) or emotion-cause pair extraction (ECPE). However, such clause-level emotion cause analysis is ambiguous and imprecise. Moreover, only identifying the emotion expression is not able to accurately convey the emotion information in many scenarios. In this paper, we aim at span-level emotion-cause-category triplet extraction (SECCE) task. The purpose of SECCE task is to identify the specific emotion spans, corresponding cause spans and the emotion categories from the given document. In contrast to the widely studied ECE and ECPE tasks, SECCE is a finer-grained task. We formulate SECCE task as a table-filling problem and design a table-filling-based label-aware convolutional neural network (TF-LaC) to solve it. Specifically, our model represents emotion-cause-category triplet(s) as word pair relations within a two-dimensional (2D) table. Then it utilizes the convolution block to exploit the regional information within the representation of the 2D table. Furthermore, a label-aware mechanism is designed to integrate the semantics of the candidate labels into the 2D table to enhance the quality of table representation. To evaluate the performance of the proposed model, we construct a dataset based on the previous benchmark dataset of emotion cause analysis. The results on this dataset demonstrate that TF-LaC outperforms the baselines in terms of the triplet extraction, achieving an improvement of span-level F1 score by at least 3.55%. The dataset and source code are available at https://github.com/ZZY-GraphMiningLab/TF-LaC.