Script Event Prediction Based on Large Model Reflection Mechanism

Script event prediction (SEP) involves predicting the most plausible next event given a sequence of historical events. This paper introduces ReGen-SEPM (Reflection and Generation for Script Event Prediction Model), a framework combining LLaMA-based generation and a reflection mechanism to improve prediction quality. The generative module produces initial predictions based on context and prompts, while the reflection module evaluates and refines them for semantic consistency and causal plausibility. A self-supervised learning mechanism further enhances the reflection module's effectiveness. Experiments on the New York Times Corpus show that ReGen-SEPM outperforms state-of-the-art baselines in accuracy, semantic similarity, and causal coherence, with ablation studies highlighting the contributions of reflection and self-supervised learning.