CHisAgent: A Multi-Agent Framework for Event Taxonomy Construction in Ancient Chinese Cultural Systems

TL;DR

The paper tackles the challenge that large language models struggle with historical-cultural reasoning in non-English contexts. It introduces CHisAgent, a three-stage, multi-agent framework (Inducer, Expander, Enricher) to induce, expand, and enrich a large-scale domain-aware event taxonomy derived from the Twenty-Four Histories, covering politics, military, diplomacy, and society in ancient China. Through extensive reference-free and reference-based evaluations, CHisAgent achieves strong structural coherence, broader coverage, and cross-cultural alignment across East Asian corpora, outperforming baselines and ablations in key metrics. The work offers a scalable approach to structuring historical knowledge that can improve cross-cultural analysis and downstream reasoning in AI systems.

Abstract

Despite strong performance on many tasks, large language models (LLMs) show limited ability in historical and cultural reasoning, particularly in non-English contexts such as Chinese history. Taxonomic structures offer an effective mechanism to organize historical knowledge and improve understanding. However, manual taxonomy construction is costly and difficult to scale. Therefore, we propose \textbf{CHisAgent}, a multi-agent LLM framework for historical taxonomy construction in ancient Chinese contexts. CHisAgent decomposes taxonomy construction into three role-specialized stages: a bottom-up \textit{Inducer} that derives an initial hierarchy from raw historical corpora, a top-down \textit{Expander} that introduces missing intermediate concepts using LLM world knowledge, and an evidence-guided \textit{Enricher} that integrates external structured historical resources to ensure faithfulness. Using the \textit{Twenty-Four Histories}, we construct a large-scale, domain-aware event taxonomy covering politics, military, diplomacy, and social life in ancient China. Extensive reference-free and reference-based evaluations demonstrate improved structural coherence and coverage, while further analysis shows that the resulting taxonomy supports cross-cultural alignment.

Figures (13)

• Figure 1: Illustration of the multi-agent CHisAgent framework, consisting of three key stages: the Inducer, the Expander, and the Enricher.
• Figure 2: Radar chart of evaluation metrics of different methods across various domains. GPT4-4o (C + D) denotes GPT4-4o (Corpus + Domain), and "Economy" represents "Economy-Livelihood".
• Figure 3: Average human evaluation scores of taxonomies generated by different methods across five dimensions.
• Figure 4: Guidelines for Human Evaluation.
• Figure 5: Taxonomy (partial) Example for Diplomacy and Society Domain.