ONSEP: A Novel Online Neural-Symbolic Framework for Event Prediction Based on Large Language Model

TL;DR

ONSEP tackles temporal knowledge graph forecasting by enabling test-time adaptation through dynamic causal rule mining (DCRM) and dual history augmented generation (DHAG). DCRM continuously updates a causal rule base while DHAG fuses short-term patterns with long-term causal trajectories via a long short-term bi-branch retriever and a hybrid inference strategy, enabling self-improvement without fine-tuning. Empirical results on ICEWS datasets show consistent Hit@1 gains over In-Context Learning across model scales (e.g., InternLM2-7B) and demonstrate inductive generalization of mined rules, indicating strong robustness to data drift. The approach highlights the potential of neural-symbolic methods to augment black-box LLMs for dynamic event prediction, with broad applicability to domains requiring rapid adaptation and interpretable causal reasoning.

Abstract

In the realm of event prediction, temporal knowledge graph forecasting (TKGF) stands as a pivotal technique. Previous approaches face the challenges of not utilizing experience during testing and relying on a single short-term history, which limits adaptation to evolving data. In this paper, we introduce the Online Neural-Symbolic Event Prediction (ONSEP) framework, which innovates by integrating dynamic causal rule mining (DCRM) and dual history augmented generation (DHAG). DCRM dynamically constructs causal rules from real-time data, allowing for swift adaptation to new causal relationships. In parallel, DHAG merges short-term and long-term historical contexts, leveraging a bi-branch approach to enrich event prediction. Our framework demonstrates notable performance enhancements across diverse datasets, with significant Hit@k (k=1,3,10) improvements, showcasing its ability to augment large language models (LLMs) for event prediction without necessitating extensive retraining. The ONSEP framework not only advances the field of TKGF but also underscores the potential of neural-symbolic approaches in adapting to dynamic data environments.

Figures (6)

• Figure 1: Comparison of ONSEP and ICL Frameworks for Event Prediction with Schematic Overview of ONSEP's Core Components and Operational Processes.
• Figure 2: Detailed structure of ONSEP framework with two phases: (1) Dynamic causal rule mining(§ \ref{['DCRM']}) and (2) dual history augmented generation(§ \ref{['DHAG']}). Specially, the DCRM phase employs a semantic-driven algorithm(§ \ref{['csd-rl']}) to identify causal rules and dynamically updates the rule base(§ \ref{['DU']}), incorporating a filtering and sorting mechanism(§ \ref{['Filter&Sort']}). (2) The DHAG phase utilizes a long short-term bi-branch retriever(§ \ref{['LSTBBR']}) alongside a hybrid model inference (§ \ref{['HMI']}) strategy to improve prediction accuracy.
• Figure 3: Performance of ONSEP in terms of Hit@1 across various DHAG ensemble weights $\lambda$ of DHAG. The underlying LLM is InternLM2-7B, processing input histories of length 100. $\lambda$ represents the weight given to long-term causal event chains. This illustrates how varying $\lambda$ influences the integration of short-term and long-term reasoning contexts within ONSEP.
• Figure 4: Example of short-term and long-term historical event chains used by ONSEP.
• Figure 5: Performance comparison across various model series under ONSEP and ICL methods with the percentage improvement indicated in red above the green bars.