故事启发大语言模型的时序知识图谱预测

Summary

Temporal knowledge graphs (TKGs) record facts as (head, relation, tail, timestamp) quadruples. They are very sparse: many entities sit in a long tail, and many events have little interaction history, so forecasting missing entities at future timestamps is unreliable. The authors propose NarrativeGraph, which converts graph history into text so that a large language model can apply its world knowledge. First, a rule-based module builds a "key event tree" for each query. It scores past quadruples by how often the head entity and the head–relation pair appear, how many distinct related entities exist, and how recent the latest event is, and it keeps only a compact set of relevant events. Second, a fine-tuned LLaMA generator rewrites these events as a chronological, causally coherent story. Third, a separately fine-tuned LLaMA reasoner reads the story together with the query and predicts the missing entity. On ICEWS14, YAGO and ICEWS18, the method reports strong results, especially on Hit@3 and Hit@10. Further analyses indicate gains for long-tail entities and sparse events. The work shows that narrative serialization is a workable bridge between structured graphs and LLMs.

Key points

• Framework name: NarrativeGraph. It has two modules: key event tree extraction and temporal story reasoning, which is split into a LLaMA generator and a LLaMA reasoner, each fine-tuned separately.
• The event-selection weight is W = f·k / (t_q − t_latest + 1). Here f is quadruple frequency, k is the number of distinct related entities, and the +1 prevents division by zero. This filtering also addresses the limits of LLM context windows (the paper cites GPT-3's 2,048 tokens as an example).
• The story generator is prompted with an instruction to link the quadruples into a narrative, adding historical background and supplementary knowledge.
• Evaluation uses three public datasets: ICEWS14, YAGO and ICEWS18. The clearest improvements are on the Hit@3 and Hit@10 metrics.
• Related work spans graph-based TKG models (TTransE, TComplEx, RE-NET, RE-GCN, CyGNet, TeMP) and LLM-based approaches (KG-BERT, PPT, GenTKG, Chain of History, GPT-NeoX in-context forecasting).