IGen: Redefining long-term event prediction with iterative generation and dynamic balancing

Abstract

Long-term event prediction is essential for anticipating future developments, enabling better decision-making and effective risk mitigation. However, long-term event prediction poses dual challenges: the accumulation of errors over time and the trade-off between factuality and diversity in predictions, which can significantly impact the reliability of predictions. We draw inspiration from the Chain-of-Thought (CoT) mechanism to tackle these challenges and propose a novel Iterative Generation framework (namely IGen) for long-term event prediction. We introduce an innovative temporal-coherence dual verification mechanism that evaluates the consistency between candidates and preceding events to enhance prediction accuracy. The temporal dimension of the dual verification emphasizes the chronological sequence of events, ensuring that predicted events follow a logical timeline. Meanwhile, another one focuses on the intrinsic connections between events, maintaining semantic alignment with preceding events for semantic and logical consistency. Additionally, we introduce an event-level dynamic nucleus sampling strategy that adjusts decoding probabilities based on the quality of preceding events, balancing diversity and factuality within and between events. Extensive experiments demonstrate that our framework outperforms SIF by approximately 22% in temporal score and 13% in coherence score, with both frameworks utilizing T5-large as the backbone. This highlights that our approach significantly outperforms traditional baselines in balancing diversity and factuality, effectively mitigating the adverse effects of error accumulation and thereby enhancing the reliability of long-term event prediction.