Leveraging Evolution Patterns to Enhance Script Event Prediction by Large Language Models

As large language models (LLMs) continue to advance, there is a growing need to explore their script event prediction capability. Fine-tuning LLMs is considered impractical when computational resources are limited. Furthermore, in script-based prediction, when events are expressed as a quadruple and do not follow conventional sentence grammar, it increases the difficulty of reasoning for LLMs. In this study, we propose a Filter-then-Rerank and analogical reasoning method that leverages event evolution patterns to unlock the intrinsic knowledge of LLMs and enhance their in-context learning (ICL) capability. Our experimental investigations on the Multiple Choice Narrative Cloze (MCNC) task demonstrate that our approach improves script event prediction performance.