Understanding the causes of emotions has long been considered significant due to their great potential in many applications, such as assisting merchants in mining the reasons behind users' opinions and adjusting marketing strategies correspondingly. Indeed, extracting causal relationships between emotions and causes is crucial in that recognizing the causes of emotions can help inform why a specific type of emotion is evoked. However, the current performance of extracting such emotion-related causal relationships is hindered by two major problems.

The first problem is the limited extraction accuracy. Despite the rapid development of machine learning-based approaches, they encounter different challenges in extracting causal relationships between emotions and causes. Existing approaches can be mainly divided into end-to-end and pipeline approaches. For end-to-end approaches, the main challenge is that multiple pairs of emotions and causes can coexist in a given piece of text data. As for pipeline approaches, the main issue is the cascading errors propagating between multiple stages.

The second problem is the incomplete extraction results. Existing approaches commonly neglect that a specific condition may be required for some causal relationships to be valid. Without considering these causal conditions, the extracted causal relationships between emotions and causes are incomplete.

Corresponding to the two problems above, our focus in this thesis is on the following objectives. Firstly, we aim to design a novel end-to-end approach to distinguish and extract multiple pairs of emotions and causes from a paragraph. Secondly, we aim to propose an innovative manner to handle the cascading errors in a pipeline model and improve the accuracy of causal relationship extraction. Thirdly, we investigate the existence of causal conditions in the extracted causal relationships and how to recognize them correspondingly.

To achieve the first objective, we propose a sequence labeling algorithm to extract multiple pairs of emotions and causes simultaneously. We design a particular set of labels composed of emotion types and causal identities (e.g., whether a clause carries the cause information or the emotion information) to distinguish the boundaries of different pairs. By assigning each clause with a specially designed label, matching over the output label sequence can fulfill the task of causal relationship extraction. We find that such an approach can outperform multiple existing baselines, even without the help of a strong pre-trained language model.

To attain the second objective, we incorporate a Reinforcement Learning (RL) framework with a redesigned two-stage model to handle the cascading error challenge. By giving each decision an explicit reward representing its effect on the final performance, an RL agent clearly knows which decision in the first stage downgrades the overall performance and learns how to avoid it. Meanwhile, the intrinsic manner of sequential prediction within the RL framework enables the model to make use of the pairs extracted earlier as auxiliary information for the subsequent extraction process. We investigate the specific effect of RL and the sequential prediction on the extraction performance, and find that the model with these two components outperforms all end-to-end approaches proposed so far. Such an observation corroborates the great benefit of RL in reducing cascading errors between stages in a pipeline model.

For the third objective, we collect and construct a dataset containing these special conditional causal relationships and propose two general modules to address the task of extracting emotions, causes, and their conditions. We demonstrate that the proposed masking module and the prediction aggregation module are beneficial in recognizing causal conditions and further improving the overall extraction performance.

To summarize, in this thesis, we propose a sequence labeling approach and an RL-based two-stage approach to promote the extraction accuracy of end-to-end approaches and pipeline approaches, respectively. With higher accuracy, the extracted causal relationships can provide more reliable knowledge concerning why a specific type of emotion is evoked and further benefit the understanding of the emotion. Besides, we propose a new task with its corresponding dataset and solutions to articulate the importance of considering causal conditions for a complete causal relationship. By recognizing the participating causal conditions, more complete causal relationships can be extracted so that the comprehension concerning the occurrence of different emotions under different causal conditions can be promoted.