Unifying emotion-oriented and cause-oriented predictions for emotion-cause pair extraction.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Artificial Intelligence

Background:

• Emotion-cause pair extraction (ECPE) identifies emotions and their causes in text without explicit emotion annotations.
• Existing methods often overlook the bidirectional relationship between emotion and cause clauses.
• This limitation hinders the comprehensive understanding and extraction of emotion-cause relationships.

Purpose of the Study:

• To propose a novel unified architecture for ECPE that capitalizes on the bidirectional nature of emotion-cause relationships.
• To redefine ECPE as a closed-loop structure detection problem to address clause mismatch.
• To enhance model training with a confidence estimation procedure for extracted pairs.

Main Methods:

• Introduced two new tasks: emotion-oriented cause prediction (EoCP) and cause-oriented emotion prediction (CoEP).
• Developed a unified single-chain architecture integrating EoCP and CoEP cells.
• Formulated ECPE as a closed-loop structure detection problem.
• Implemented a confidence estimation procedure for training.

Main Results:

• The proposed unified architecture achieved state-of-the-art performance on two public datasets.
• Demonstrated significant improvements in extracting multiple emotion-cause pairs.
• The bidirectional approach and closed-loop formulation proved effective.

Conclusions:

• The novel unified architecture significantly advances the state-of-the-art in emotion-cause pair extraction.
• The bidirectional prediction tasks (EoCP, CoEP) and closed-loop approach effectively address limitations of previous methods.
• The model shows particular strength in complex scenarios involving multiple emotion-cause pairs.