ECPE-Qwen: Zero-shot Emotion-cause Pair Extraction with Fine-tuning Large Language Models

The goal of the Emotion-Cause Pair Extraction (ECPE) task is to extract emotion clauses and the corresponding cause clauses from a given text. To enhance the performance of LLMs on the ECPE task, we introduce a zero-shot instruction construction method using numerical identifiers. From a global text perspective, we employ the Low-Rank Adapter (LoRA) method for fine-tuning the Qwen1.5 base model across three stages: emotion clause extraction, cause clause extraction, and emotion-cause pair extraction. This approach guides the LLMs in modeling inter-sentence emotional relationships. Experimental results demonstrate that ECPE-Qwen model, with only 1.8 billion parameters, achieves performance on par with GPT-3.5.