PaperlayerView exact revision on arXiv

Source-linked AI summary

AI Historian: Helping historians organize and verify person-centred temporal clues from dispersed historical narratives

Yifeng Lu, Zijie Yang, Jie Li, Qingkai Min, Yue Zhang

arXiv:2608.29133v1cs.CL

TL;DR

Historians must reconnect dispersed, person-centred temporal evidence across chapters and sources while preserving provenance and interpretive context. AI Historian uses a structured, multi-agent workflow to organize and verify this evidence, achieving 86.2% MicroIoU in six Shiji cases while reducing annotation time.

  • Problem

    Historical evidence about people, actions and events is dispersed across documents, chapters and narrative perspectives, requiring retrieval, identification and comparison to reconstruct temporal sequences.

  • Method

    AI Historian combines large language models with multi-agent technology to organize source sentences into traceable person–time evidence, verify cross-text associations and support revisable temporal judgements.

  • Results

    86.2% MicroIoU across 244 sentence-level temporal ranges was achieved in 13 min 56 s, versus 1 h 32 min for human-only annotation.

  • Takeaways & Limitations

    AIH supports large-scale, cross-text historical research by preserving primary-source traceability while enabling historians to compare, revise and reject candidate associations.

  • Takeaways & Limitations

    Direct model performance remains affected by long-context evidence position, temporal reasoning limitations and inconsistently reliable citation support and self-evaluation.

Abstract

from arXiv · show

History is not preserved in complete, continuous form. Accounts of a person's activities, relationships and historical contexts are scattered across texts, chapters and narrative perspectives; historians must retrieve, identify and compare these materials to reconstruct temporal sequences and verify them against sources. Here we present AI Historian (AIH), an AI agent system that helps historians organize person-time evidence from dispersed biographical narratives. It takes source sentences as evidence units, identifies people and temporal cues, verifies candidate cross-text associations and infers comparable temporal ranges while preserving traceable source-text evidence. We evaluated AIH on six Shiji cases concerning Liu Bang, Xiang Yu and Xiao He. AIH Agent achieved a temporal-localization MicroIoU of 86.2%, compared with 81.3% for human-only annotation and 17.1% for direct large-language-model prompting; it required about 14 min, versus 1 h 32 min for human-only annotation. We further applied AIH to the Twenty-Four Histories and other ancient Chinese histories, ancient Japanese and Korean histories, and modern and contemporary historical materials, and released the results through Westlake Historian. These results indicate that AIH can reduce the cost of organizing historical materials at scale while turning connections obscured by chapter-based narration into traceable, revisable research questions for collaborative testing.

Introduction

Historical evidence about people and events is dispersed across chapters, texts and narrative perspectives, making temporal reconstruction and source verification difficult. AI Historian organizes these fragments into traceable person-centred timelines and shared temporal representations.

  • Research context: Biographical and annals-biographies formats scatter evidence about the same person or event across chapters and perspectives, requiring retrieval, comparison and source criticism.The Shiji exemplifies this challenge, with Liu Bang’s activities distributed across multiple biographies.
  • Research gap: Existing digital-history components address dates, event ordering, timelines and knowledge graphs, but do not integrate dispersed judgements into item-level traceable results.Large language models also have limitations in temporal reasoning, citation support and self-evaluation, motivating inspectable workflows.
  • AIH approach: AIH uses source sentences as evidence units, preserves provenance, inferential paths and uncertainty, and organizes person-centred narratives on a unified timeline.Sentences may be linked to multiple people, with candidate temporal ranges generated for each person and aligned across people.
  • Evaluation: AIH was evaluated against human-only annotation and direct large-language-model prompting across six Shiji cases concerning Liu Bang, Xiang Yu and Xiao He.The evaluation targets both accuracy and efficiency in organizing person–time evidence.
  • Results and deployment: 86.2% MicroIoU exceeded 81.3% for human-only annotation and 17.1% for direct prompting, while AIH required approximately 14 min versus 1 h 32 min for human-only annotation.The system was also implemented in Westlake Historian with materials from Chinese, Japanese, Korean, modern and contemporary histories.

Results

AIH reconstructs traceable person-centred temporal evidence from dispersed biographical narratives by decomposing extraction, ordering, verification and normalization into auditable stages. Across six Shiji cases, it improved temporal localization and reduced completion time while supporting historian-led inspection and revision.

  • Evidence organization: AIH organizes dispersed materials into temporal evidence that remains traceable to source sentences and can be questioned or revised.It groups narratives sharing temporal context, adjusts retrospective ordering, verifies cross-biography relations and standardizes only verified temporal ranges.
  • Evidence organization: AIH retrieves passages across Liu Bang’s biographies and places them on a shared temporal axis for comparison among Liu Bang, Xiang Yu and Xiao He.The resulting evidence chain links temporal ranges to people, chapters, source sentences and the basis of inference.
  • End-to-end evaluation: 86.2% MicroIoU exceeded human-only annotation (81.3%) and direct prompting (17.1%) across 244 sentence-level temporal ranges.MicroIoU measures month-resolution overlap between predicted and reference temporal ranges.
  • End-to-end evaluation: 13 min 56 s reduced AIH Agent’s elapsed time relative to 1 h 32 min for human-only annotation, while direct prompting took 2 min 28 s.AIH Agent used approximately 15% of the human time; direct prompting was faster but had low overlap with reference ranges.
  • Diagnostic evaluation: 100.0% MicroIoU was achieved for direct temporal localization and ambiguous within-document reasoning, while cross-document propagation reached 71.2%.The cross-document score remained above human-only annotation (67.1%) and direct prompting (11.1%).
  • Diagnostic evaluation: Structured prompting increased cross-text event-verification and temporal-alignment accuracy from 75.0% to 91.7%, but did not improve within-document temporal reasoning.Cross-text candidate relations and temporal-boundary judgements were identified as decisive links in the final result.
  • Historian-led research: Westlake Historian lets researchers compare original passages, inspect temporal inferences, revise entries and record human acceptance decisions.AI review supplies suggestions, while historians make and record final decisions.
  • Historian-led research: AIH’s reorganization connects fragmented accounts of troop movements and supplies across biographies into a view of the Han army’s recovery and return to battle.The example illustrates how juxtaposition can expose relations among activities distributed across chapters.

Discussion

AIH addresses dispersed, difficult-to-verify historical evidence by separating and preserving chronology construction, cross-text comparison and source inspection. Its evaluations and applications indicate more accurate, faster reconstruction that produces traceable leads without replacing historians’ source judgements.

  • Discussion: AIH separates temporal-cue identification, narrative-order adjustment, cross-text verification and temporal-range normalization in its chronology workflow.The paper presents this separation as an explanation for structured prompting’s largest gain in cross-text event verification and temporal alignment.
  • Discussion: AIH turns chapter-obscured connections into testable historical leads while preserving links to the original passages for source criticism.The Pengcheng case illustrates juxtaposition as an expanded field of inspection rather than a causal conclusion.
  • Discussion: AIH supports comparison across broader historical materials, including different regions, periods and languages, while historians retain responsibility for accepting, revising or rejecting associations.Westlake Historian presents examples across Chinese, Japanese, Korean, modern and contemporary sources.

Methods

The methods evaluate AIH’s staged, auditable reconstruction of person-centred temporal evidence using six Shiji case packets and complementary diagnostic questions. Inputs are modern Chinese translations, and outputs are assessed through temporal-range overlap, strict component accuracy and reported elapsed time.

  • Evaluation-case selection and input texts: Six Shiji case packets contain 244 eligible sentence-level temporal ranges drawn from three interrelated biographies and organized into direct, ambiguous and cross-text cases.The cases concern Liu Bang, Xiang Yu and Xiao He and use modern Chinese vernacular translations.
  • Diagnostic reasoning: Experiment 2 tests temporal-information extraction, temporal reasoning and cross-text verification with 48 unique questions and 96 scored human responses.Each of three diagnostic forms contains 16 questions, and model results use majority vote across three runs.
  • AIH workflow: AIH Agent uses nine semantic agents across sentence-level signals and TimeBlock-level chronological units to identify people, temporal cues, narrative functions and temporal order.Later stages propagate only evidence-verified cross-text relations into normalized anchors or boundary constraints.
  • Temporal reconstruction: Experiment 1 compares predicted and reference temporal ranges after projection to finite month sets using month-level MicroIoU.The reported overall value is a case-size-weighted mean across eligible sentence-level ranges.
  • Evaluation criteria: Strict row accuracy in Experiment 2 counts an item as correct only when every required field matches the reference answer.Human and model conditions use separate response sets, with model outputs determined by majority vote across three independent runs.
  • Auditability: AIH saves intermediate outputs at each rule-based and model-based stage, enabling final chronology entries to be traced through normalized ranges, TimeBlocks, sentence identifiers and source text.This workflow separates preprocessing, annotation, reasoning, propagation, normalization and summary generation.

Data availability

The derived Shiji materials and annotations are available through Westlake Historian for tracing chronology entries to their source evidence.

  • Data availability: Westlake Historian provides the three vernacular Shiji translations, derived annotations, ordering sequences and final TimeBlock outputs.The platform supports tracing final chronology entries to TimeBlocks, sentence positions and source-text evidence.

1 Biographical chronology and source organization

Biographical chronologies organize a person’s life in temporal order, while extended versions preserve selected source excerpts under dated entries. Source selection and quotation scope can yield multiple chronological versions.

  • Traditional chronology: Traditional biographical chronologies record a person’s principal life events in temporal order.
  • Extended chronology: Extended biographical chronologies preserve selected source excerpts under dated entries.Different source selections and quotation scopes can produce multiple chronological versions for the same person.
  • Source organization: AIH organizes dispersed biographical fragments as source-linked chronological entries while preserving their original chapter and sentence locations.The fragments may reflect different participants’ perspectives or narrative contexts.

2 Evaluation and methods

AIH was evaluated on six Shiji cases using human-only annotation, direct prompting, and structured agent workflows. Its pipeline decomposes sentence and TimeBlock processing into traceable temporal extraction, ordering, cross-text verification, and range normalization.

  • Evaluation design: Experiment 1 used six Shiji cases spanning direct localization, within-document ambiguity, and cross-document temporal propagation.Cases were drawn from biographies of Xiang Yu, Liu Bang, and Xiao He; each case was completed once by one participant.
  • Evaluation design: AIH organized full chapters into sentence-level annotations, TimeBlock representations, and ordered sequences.Modern-Chinese translations reduced confounding from differences among models in Classical Chinese processing.
  • AIH architecture: A1–A4 identify people, temporal expressions, textual functions, and retrospective passages, while A5–A9 perform temporal transformation, ordering, granularity assessment, propagation, and normalization.Preprocessing, TimeBlock assembly, and summary generation support the nine semantic agents.
  • Cross-text processing: A8 verifies candidate cross-document relations and citations before A9 converts eligible evidence into temporal constraints and normalized ranges.Unverified candidates enter human review rather than temporal propagation.
  • Temporal representation: AIH represents temporal precision at four levels and serializes interval endpoints as fixed-width internal date coordinates.Closed and open ranges use explicit interval notation, while original reign-title expressions are retained alongside internal coordinates.

3 Evaluation results

Across end-to-end and diagnostic evaluations, structured AIH workflows generally outperformed direct prompting in temporal accuracy, with especially strong gains in cross-text verification and alignment. Performance and elapsed time varied by case group, model, and evaluation condition.

  • End-to-end results: 90.7% was the highest descriptive macro-average MicroIoU for AIH Agent, achieved with Claude Opus 5 across six cases.The corresponding values were 90.2% for DeepSeek-V4-Flash, 88.6% for Gemini 3.1 Pro, 86.9% for GPT-5.6 Sol, and 78.6% for Qwen 3.6.
  • End-to-end results: 92.9%–100.0% was AIH Agent MicroIoU in easy cases versus 86.0% for human-only observation, while hard-case values ranged from 57.6% to 79.4% versus 67.1%.Direct prompting yielded 3.7%–12.0% in easy cases and 8.7%–11.2% in hard cases.
  • Diagnostic results: 75.0% was structured-prompting strict accuracy across 48 diagnostic questions, compared with 68.8% for direct prompting and 64.6% for human responses.Cumulative elapsed times were 14 min 2 s, 3 min 43 s, and 2 h 26 min, respectively.
  • Diagnostic results: 91.7% versus 75.0% was structured versus direct prompting accuracy for cross-text event verification and temporal alignment.For temporal-information reasoning, both prompting conditions reached 75.0%, whereas human responses reached 79.2%.
  • Diagnostic results: 52 min, 30 min 36 s, and 1 h 4 min were human elapsed times across extraction, reasoning, and cross-text modules.Direct prompting was fastest in each listed module, while structured prompting took 12 min 14 s, 50 s, and 58 s.

4 Cross-text cases

AIH aligns dispersed biographies by verifying shared events or phases and propagating only supported temporal constraints. The C5 and C6 cases illustrate shared-phase alignment and completion of a biographical chronology from another chapter’s timeline.

  • Case scope: C5 links Xiang Yu and Gaozu biographies, while C6 links Gaozu and Xiao He materials on a shared comparable timeline.The cross-text annotation covers passages describing the same historical process or adjacent historical phases.
  • C5: shared historical phase: In C5, A8 verifies that passages from the Xiang Yu and Gaozu biographies describe the same sequence of military reorganization and movement.The example passages concerning King Huai’s move to Pengcheng receive the reference range -0207-10to-0206-.
  • C6: biographical completion: In C6, Xiao He passages receive Chu–Han timeline ranges supplied by the Gaozu biography.The listed ranges progress from -0209-10to-0208-10 through -0205-01to-0205-10.
  • Temporal propagation: A9 combines target-sentence order, adjacent boundaries, and verified cross-text evidence to generate iso_range.Relation types determine whether propagation uses transferable anchors, source intervals, or preceding and following phase boundaries.

5 Platform and structured outputs

Westlake Historian exposes AIH’s structured evidence layer through tools for inspection, querying, revision, and publication. The platform processes broad historiographical collections while preserving traceable TimeBlock and cross-text outputs for human review.

  • Platform workflow: Westlake Historian supports evidence retrieval, source comparison, source-grounded questions, translation, draft generation, time-point tracing, and monitored AIH processing.These functions share the AIH structured-evidence layer.
  • Platform workflow: Users can submit revisions with reasons, receive AI-assisted review recommendations, and retain the final human decision before private storage or publication.Accepted versions may be published to the Chronology Market for discovery, inspection, and discussion.
  • Corpus coverage: 91 strong cross-text relations and 684 candidate relations were recorded in the platform corpus snapshot.Strong relations enter A9 propagation, while candidate relations enter subsequent human review.
  • Corpus coverage: 25 corpus collections covered 26 historical works and one modern-history collection, yielding 75 person timelines.The number of historical works exceeds the number of collections because two comparative collections combine old and new histories.
  • Structured outputs: 154 TimeBlocks contained cross-text update records in the structured outputs for the three Shiji chapters.TimeBlocks preserve source spans, temporal markers, normalized ranges, cross-text updates, and summaries for querying and tracing.

6 Terminology and proper names

This section standardizes Chinese and English terminology for historiographical forms, AIH components, temporal reasoning units, and source-linked chronological judgements.

  • The terminology preserves Chinese source terms initially, then uses corresponding English or romanized forms; names use pinyin in Chinese order and group aliases under one identity.
  • An extended biographical chronology is an evidence-rich chronology incorporating source excerpts, quotations and editorial notes.
  • A multi-person chronology combines several historical people’s timelines to represent their interactions.
  • A TimeBlock is contiguous text governed by a common temporal marker and used as the basic temporal-reasoning unit.
  • A Temporal Marker Buffer is document-level memory that normalizes elliptical or relative temporal expressions.
  • An evidence chain links a final chronological judgement to TimeBlocks, sentences and source texts.
Loading 2608.29133v1…