REVOLUTIONIZING EVENT DETECTION: A NOVEL PROMPT-DRIVEN METHOD ENHANCED BY RETRIEVAL-AUGMENTED PARADIGM

Xi Zhang, Zeping Wang, Chenghua Guo, Pan Suo, Sanchuan Guo

Keywords: Event Detection, Large Language Models(LLMs), Lightweight LLMs, Cascading Prompt-Based Framework, Automated Prompt-Design

Abstract: Event Detection (ED) task involves extracting event triggers from sentences and classifying them into predefined event types. While large language models (LLMs) have become widely adopted across various NLP tasks, their application to ED remains relatively unexplored. All existing LLM-based approaches follow a traditional prompt-based paradigm, which requires designing distinct prompts for each event type. This strategy, however, suffers from a fundamental limitation: as the number of event types grows, the number of prompts needed increases linearly, resulting in significant manual effort and computational costs. To overcome this limitation, we propose a novel approach that integrates a retrieval-augmented mechanism with a redesigned cascading prompt-based framework. Specifically, the prompt-based component is employed to extract candidate triggers, while the retrieval-augmented module applies heuristic filtering strategies to coarsely eliminate irrelevant candidates. In addition, we put forward an innovative automated prompt-design method to accurately match valid triggers with their corresponding event types based on retrieved information. Experimental results on ACE-05 benchmark demonstrate the state-of-the-art performance under our scheme. Furthermore, the approach remains highly effective when using lightweight LLMs, indicating its strong potential for efficient large-scale data processing. This capability may have profound implications and become a fundamental work for future research.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 6441