Script Event Prediction Based on Knowledge Distillation of Large Language Models

Script event prediction selects the most likely event from a set of candidates based on the sequence of events. Previous studies primarily relied on event pairs, chains, and evolution diagrams, using probabilistic theories to determine event transfer direction. However, as data volume increases, the transfer direction becomes unpredictable, leading to reduced model accuracy. Additionally, these studies often overlook the inclusion of background knowledge and conditional constraints, resulting in a broader solution space for event prediction. To address these issues, we propose a novel framework called Script Event Prediction Based on Knowledge Distillation of Large Language Models (SEKD). We establish a formal definition for script event data and utilize LLMs' reasoning capabilities to guide internal thinking and logical explanations through a thinking chain. These explanations serve as additional supervision for a smaller model trained through knowledge distillation, enabling it to inherit and learn reasoning abilities from LLMs. Experimental results on two datasets demonstrate the strong performance of our proposed model, showcasing robust understanding and learning of reasoning cognition and logic from LLMs, thus possessing fundamental reasoning capabilities.