Multi-step script event prediction based on reinforcement learning

Script Event Prediction aims to predict multiple subsequent events based on a given sequence of past events. While one-step script event prediction methods have achieved relatively high accuracy, their applicability is limited, as they cannot capture the evolving trends of event sequences. Moreover, most existing multi-step prediction approaches rely on one-step probability accumulation, which often leads to local optima while neglecting the overall quality of the predicted event chain. To address this issue, we propose a reinforcement learning framework to model the entire multi-step event prediction process. Specifically, we treat the current event chain as the state and employ a multi-layer perceptron as the agent to generate conditional probabilities for each candidate event. Each time a new event is predicted, the state of the event chain is dynamically updated before proceeding to the next prediction step. Additionally, we design a reward function based on the Bilingual Evaluation Understudy (BLEU) to ensure that the predicted event chain maintains high overall quality. Experiments on the MCNC task demonstrate that our framework outperforms existing methods to a certain extent, achieving significant improvements in multi-step script event prediction accuracy.