EventRAG: Supportive Event Retrieval on Hypergraph for Future Forecasting

ACL ARR 2025 February Submission4113 Authors

Abstract: Recent advancements in Retrieval-Augmented Generation (RAG) have significantly enhanced the ability of large language models to ground their responses in external knowledge. However, solving complex future forecasting problems remains a challenge due to the need for retrieving supportive events. Current methods focusing on textual-similarity or entity-relevance are not able to capture supportive events due to incompleteness of the knowledge base and the inherent nuanced nature of events. This paper introduces EventRAG, an event-oriented RAG framework specifically designed for future forecasting tasks. Specifically, we first propose the supportive event retrieval where we construct the event hypergraph index on the knowledge base. Based on that, we denote the event supportiveness as random variables and maximize the expectation. We establish the maximum expected event cover program to solve this maximization. Finally, EventRAG integrates the retrieval and reasoning into the event-oriented agentic reasoning process. It enables the framework to retrieve the needed information to perform complicated forecasting. We conducted experiments and in-depth analysis to evaluate the effectiveness of EventRAG. The results demonstrate that EventRAG significantly outperforms competitive RAG baselines in future forecasting. The code and dataset are available on the ARR system.

Paper Type: Long

Research Area: Information Retrieval and Text Mining

Research Area Keywords: RAG;reasoning;future forecasting

Contribution Types: Model analysis & interpretability

Languages Studied: en

Submission Number: 4113