Enhancing Large Language Model with Decomposed Reasoning for Emotion Cause Pair Extraction

TL;DR

This paper tackles Emotion-Cause Pair Extraction (ECPE), where conventional models overfit position-based cues. It introduces the Decomposed Emotion-Cause Chain (DECC), a multi-step chain-of-thought framework that splits ECPE into emotion extraction and cause extraction, guided by iterative prompting and pruning; it is paired with in-context learning via diverse demonstrations. Across Chinese and English ECPE datasets, DECC achieves competitive results compared with state-of-the-art supervised fine-tuning, while remaining robust to dataset biases and varying LLM bases. The work demonstrates that structured, decomposed reasoning can unlock strong information-extraction performance from large language models without additional training, with practical benefits for multi-pair extraction and bias-resilient deployment.

Abstract

Emotion-Cause Pair Extraction (ECPE) involves extracting clause pairs representing emotions and their causes in a document. Existing methods tend to overfit spurious correlations, such as positional bias in existing benchmark datasets, rather than capturing semantic features. Inspired by recent work, we explore leveraging large language model (LLM) to address ECPE task without additional training. Despite strong capabilities, LLMs suffer from uncontrollable outputs, resulting in mediocre performance. To address this, we introduce chain-of-thought to mimic human cognitive process and propose the Decomposed Emotion-Cause Chain (DECC) framework. Combining inducing inference and logical pruning, DECC guides LLMs to tackle ECPE task. We further enhance the framework by incorporating in-context learning. Experiment results demonstrate the strength of DECC compared to state-of-the-art supervised fine-tuning methods. Finally, we analyze the effectiveness of each component and the robustness of the method in various scenarios, including different LLM bases, rebalanced datasets, and multi-pair extraction.

Figures (5)

• Figure 1: An Example of ECPE task
• Figure 2: Illustration of DECC framework. At locating step, emotional keyword angry is pruned. The analyzing chain corresponding to worshipful is pruned at step 3. The analyzing chain of surprise is pruned at the last step. e1-e4 denote the id of specific emotional keyword. ec and cc denote emotion clause and cause clause, respectively. ac stands for the corresponding analyzing chain.
• Figure 3: Results under different ICL shot numbers on Chinese dataset and English dataset
• Figure 4: The prompt settings for English dataset.
• Figure 5: The prompt settings for Chinese dataset.