Chain of History: Learning and Forecasting with LLMs for Temporal Knowledge Graph Completion

Ruilin Luo,Tianle Gu,Haoling Li,Junzhe Li,Zicheng Lin,Jiayi Li,Yujiu Yang

2024-02-14

Abstract:Temporal Knowledge Graph Completion (TKGC) is a complex task involving the prediction of missing event links at future timestamps by leveraging established temporal structural knowledge. This paper aims to provide a comprehensive perspective on harnessing the advantages of Large Language Models (LLMs) for reasoning in temporal knowledge graphs, presenting an easily transferable pipeline. In terms of graph modality, we underscore the LLMs' prowess in discerning the structural information of pivotal nodes within the historical chain. As for the generation mode of the LLMs utilized for inference, we conduct an exhaustive exploration into the variances induced by a range of inherent factors in LLMs, with particular attention to the challenges in comprehending reverse logic. We adopt a parameter-efficient fine-tuning strategy to harmonize the LLMs with the task requirements, facilitating the learning of the key knowledge highlighted earlier. Comprehensive experiments are undertaken on several widely recognized datasets, revealing that our framework exceeds or parallels existing methods across numerous popular metrics. Additionally, we execute a substantial range of ablation experiments and draw comparisons with several advanced commercial LLMs, to investigate the crucial factors influencing LLMs' performance in structured temporal knowledge inference tasks.

Artificial Intelligence,Computation and Language

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to complete the link prediction task in the Temporal Knowledge Graph (TKG), that is, to predict the missing event links at future time points after a series of observed facts are given. This task is known as Temporal Knowledge Graph Completion (TKGC). Specifically, the goals of the paper are:

1. **Explore the application of large - language models (LLMs) in temporal knowledge graph reasoning**: Research how to use the advantages of LLMs to understand and reason about the structural information in the temporal knowledge graph, especially when dealing with key nodes in the historical chain.

2. **Propose an efficient and transferable framework**: Develop a general framework that can use LLMs for temporal knowledge graph reasoning. This framework can not only effectively predict future event links but also adapt to different data sets and application scenarios.

3. **Solve the challenge of reverse logical reasoning**: Explore how to overcome the "reverse curse" problem of LLMs in dealing with structured expressions by introducing reverse logical reasoning, that is, the model performs well in forward reasoning but poorly in reverse reasoning.

4. **Optimize the model training strategy**: Use Parameter - Efficient Fine - Tuning (PEFT) technology to fine - tune LLMs to enhance the model's understanding of the historical context and improve its performance in the temporal knowledge graph completion task.

Through these goals, the paper aims to provide a comprehensive perspective on how to fully utilize the advantages of LLMs to solve the complex task of temporal knowledge graph completion.