Collaborate SLM and LLM with latent answers for event detection
Youcheng Yan,Jinshuo Liu,Donghong Ji,Jinguang Gu,Ahmed Abubakar Aliyu,Xinyan Wang,Jeff Z. Pan
DOI: https://doi.org/10.1016/j.knosys.2024.112684
IF: 8.139
2024-11-10
Knowledge-Based Systems
Abstract:Event detection (ED) intends to identify events from text and classify them into predefined event types. One of the major issues for ED is the low-resource problem due to inadequate samples. Some studies address the low-resource issue with retrieving knowledge entries directly from knowledge bases while introducing a lot of irrelevant knowledge or failing the lookup. Moreover, recent work has attempted to employ large language models (LLMs, e.g., ChatGPT) that directly access event types in unstructured text under low-resource scenarios. Although LLM-based approaches have obtained promising results, we consider that the full potential of LLMs has not been activated due to insufficient prompt information. Our research proposes a two-stage event detection method that collaborates small language models (SLMs) and LLMs, namely LSLAED. Specifically, we first fine-tune the SLM to generate three types of latent answers: answer-aware examples, structure-aware examples, and corresponding answer candidates. Subsequently, all latent answers will form the prompt and enable the LLM to improve performance through in-context learning. We evaluate the proposed method using precision, recall, and F1-score as evaluation metrics. Experiments on the ACE2005 and ERE-EN datasets have demonstrated that LSLAED achieves significant improvement in both full-shot and few-shot scenarios.
computer science, artificial intelligence