Emotion-Cause Pair Extraction: A New Task to Emotion Analysis in Texts

Rui Xia,Zixiang Ding

DOI: https://doi.org/10.48550/arXiv.1906.01267

2019-06-04

Abstract:Emotion cause extraction (ECE), the task aimed at extracting the potential causes behind certain emotions in text, has gained much attention in recent years due to its wide applications. However, it suffers from two shortcomings: 1) the emotion must be annotated before cause extraction in ECE, which greatly limits its applications in real-world scenarios; 2) the way to first annotate emotion and then extract the cause ignores the fact that they are mutually indicative. In this work, we propose a new task: emotion-cause pair extraction (ECPE), which aims to extract the potential pairs of emotions and corresponding causes in a document. We propose a 2-step approach to address this new ECPE task, which first performs individual emotion extraction and cause extraction via multi-task learning, and then conduct emotion-cause pairing and filtering. The experimental results on a benchmark emotion cause corpus prove the feasibility of the ECPE task as well as the effectiveness of our approach.

Computation and Language

What problem does this paper attempt to address?

The problems that this paper attempts to solve are two main flaws in extracting emotions and their underlying causes from text:

1. **Dependence on emotion annotation**: In traditional Emotion - Cause Extraction (ECE) tasks, it is necessary to annotate emotions first before extracting causes. This dependence restricts the application of ECE in real - world scenarios.
2. **Independent handling of emotions and causes**: Traditional methods first annotate emotions and then extract causes, ignoring the mutual indicative relationship between emotions and causes.

To solve these problems, the author proposes a new task: Emotion - Cause Pair Extraction (ECPE). The goal of ECPE is to directly extract all potential emotion - cause pairs from documents without prior emotion annotation. This new task not only solves the above two problems but also enables emotion - cause analysis to be better applied in practical scenarios.

### Specific problem description
- **Dependence on emotion annotation**: In traditional ECE tasks, emotions in the test set must be annotated first, which is impractical in practical applications.
- **Independent handling of emotions and causes**: Traditional methods ignore the mutual influence between emotions and causes, that is, emotions can help discover causes, and knowing the causes can also extract emotions more accurately.

### Solution
The author proposes a two - step framework to solve the ECPE task:
1. **Individual emotion and cause extraction**: Extract emotions and causes separately through a multi - task learning network.
2. **Emotion - cause pair matching and filtering**: Pair the extracted emotions and causes and train a filter to eliminate pairs without a causal relationship.

### Experimental results
- **Overall performance**: The experimental results show that the proposed Inter - EC method performs best on the ECPE task and its subtasks, especially with a significant improvement in recall in the cause extraction task.
- **Upper - bound experiment**: The upper - bound experiment using annotated data further verifies the mutual indicative relationship between emotions and causes.
- **Effect of emotion - cause pair filtering**: After introducing the pairing filter, the F1 score is significantly improved, the precision is greatly increased, and the recall is slightly decreased.

### Conclusion
The new task ECPE proposed by the author and its solution effectively overcome the limitations of traditional ECE tasks, prove the mutual indicative relationship between emotions and causes, and show potential in practical applications. Future work will attempt to build an end - to - end model to directly extract emotion - cause pairs.