cs.CLAug 29, 2026
SaveAI Historian: Helping historians organize and verify person-centred temporal clues from dispersed historical narratives
Organizations: Westlake University · Peking University
Abstract
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.
Figures & tables
Figure 1: AIH organizes dispersed source sentences into person-centred timelines and aligns them on a shared temporal axis. a, From example sentences in the Basic Annals of Gaozu , the Basic Annals of Xiang Yu and the Hereditary House of Chancellor Xiao , AIH identifies people and temporal cues, groups adjacent sentences sharing the same temporal context into TimeBlocks, and records inferred ranges using start and end dates; the bilingual table on the right retains the person, source chapter, sentence number and TimeBlock number. b, Six relevant sentences concerning Liu Bang are assembled from the three source texts while retaining their source identities and original numbering. Coloured dots on the right of each sentence card mark associated people, so a single sentence can be linked to Liu Bang, Xiang Yu and/or Xiao He. c, The six pieces of evidence are reordered according to their inferred temporal ranges to produce a single-person timeline for Liu Bang; shapes continue to denote their respective sources. d, Candidate ranges generated separately for Liu Bang, Xiang Yu and Xiao He are aligned on a shared temporal axis, making temporal overlaps directly visible; the diagram on the left shows their convergence from person-specific ranges into a common temporal coordinate system. Researchers assess the historical significance of these overlaps against the source texts. Colours denote people, shapes denote source texts and bars denote inferred temporal ranges.
Figure 2: AIH Agent architecture and TimeBlock-level temporal reasoning. a, AIH Agent takes biographies from histories in the annals-biographies format as input. Sentence-level agents (A1–A4) identify people, extract temporal expressions from the original text, determine narrative functions and mark temporal shifts caused by retrospective passages. TimeBlock-level agents (A5–A9) then perform within-document temporal reasoning, cross-biography evidence verification and normalized temporal-coordinate generation, producing traceable person–time evidence; the biographical chronology and temporal-range views are system outputs. b, A TimeBlock groups sentences governed by the same temporal marker or temporal context into a sortable unit, supporting completion of incomplete temporal markers, determination of temporal granularity and anchors, and reordering from narrative into chronological order. c, Before temporal information is transferred, A8 assesses whether the selected source and target TimeBlocks provide sufficient evidence to support the same event or historical phase. Relations that meet the propagation criteria enter A9; other candidates are retained for human review. A9 converts verified temporal evidence from the source TimeBlock into normalized boundary constraints that narrow the temporal range of the target TimeBlock. These internal coordinates provide the system’s standardized representation for ordering and scoring; their serialization is defined in Supplementary Section 2.2.1 . Detailed architecture and methods are provided in Supplementary Fig. 4 – 6 .
Figure 3: Evaluation design and end-to-end reconstruction of person-centred temporal ranges. a, Evaluation design: source materials were expert-annotated and adjudicated to produce sentence-level month ranges and diagnostic labels, and then entered two complementary branches. Experiment 1 compares human-only annotation, direct large-language-model prompting and AIH Agent in sentence-level temporal-range reconstruction; it outputs sentence-level month ranges and evaluates them using MicroIoU and cumulative elapsed time. Experiment 2 compares human responses, direct large-language-model prompting and structured large-language-model prompting in diagnostic reasoning; it outputs row-level answers and evaluates them using strict accuracy and cumulative elapsed time. b, End-to-end temporal-reconstruction performance across six evaluation cases. Bars show month-level, case-size-weighted MicroIoU for human-only annotation, direct large-language-model prompting and AIH Agent over 244 eligible sentence-level temporal ranges. c, Cumulative elapsed time for the same end-to-end task, shown in seconds on a logarithmic axis. Accuracy and elapsed time by case group are provided in Supplementary Fig. 7 ; accuracy and elapsed time for the diagnostic experiment are provided in Supplementary Fig. 9 and Supplementary Fig. 10 .
Figure 4: Westlake Historian incorporates AIH-processed person–time evidence into a historian-led research process. a, Historians begin with person records, temporal ranges and source links organized by AIH, and delimit the people, period and sources under examination. b, Within the same scope, the platform juxtaposes relevant original passages from different texts, allowing researchers to compare people’s activities, formulate questions for investigation and use the original text to assess further the historical relations among the materials. c, For a selected lead, the platform presents its surrounding context, original temporal information, temporal marker completed from context, inferred temporal range and related sources, enabling the temporal judgement to be verified against the original text. d, Historians inspect the evidence and propose revisions to entries or temporal ranges; AI review provides suggestions, but the final decision is made and recorded by a human. Accepted versions may remain in a private workspace or be published to the Chronology Market. The numbers along the bottom summarize the processed corpus analysed in this study.
| Evaluation | Conditions | Evaluation set | Primary metric | Aggregation and timing |
|---|---|---|---|---|
| Experiment 1: temporal reconstruction | Direct large-language-model prompting; human-only; AIH Agent | Six case packets; 244 eligible sentence-level temporal ranges | Month-level MicroIoU between predicted and reference ranges | Case-size-weighted mean; six human participants; AIH Agent row-level consensus across three runs; cumulative elapsed time |
| Experiment 2: diagnostic reasoning | Human; direct large-language-model prompting; structured large-language-model prompting | 96 human responses across six forms; 48 unique questions per model condition | Strict row accuracy; all required fields must match | Majority vote across three model runs; total elapsed time by condition |
Table 1: Design and reporting criteria for the human and model evaluations.
Figure 1: Traditional and extended biographical chronologies. a, A traditional biographical chronology records the principal events of a person’s life in temporal order. b, An extended biographical chronology preserves selected source excerpts under dated entries; source selection and quotation scope can produce multiple chronological versions for the same person.
Figure 2: Biographical material dispersed across chapters. A historical process may be distributed across several biographies, each preserving passages from a participant’s perspective or a different narrative context. AIH compares these fragments across chapters and organizes them as source-linked chronological entries while preserving their original chapter and sentence locations.
| Participant | Stage 1 | Stage 2 |
|---|---|---|
| P1 | C1 | T1 |
| P2 | C2 | T2 |
| P3 | C3 | T3 |
| P4 | C4 | T1 |
| P5 | C5 | T2 |
| P6 | C6 | T3 |
Figure 3: Participant allocation across the two evaluation stages. Stage 1 was the end-to-end evaluation under the human-only condition, and Stage 2 was the component-level diagnostic evaluation. Blue boxes indicate cases C1–C6 and green boxes indicate diagnostic forms T1–T3. C1–C2 evaluate direct temporal localization, C3–C4 within-document ambiguity reasoning and C5–C6 cross-document temporal propagation. Each case was completed once by one participant, and each diagnostic form by two participants.
Figure 4: Complete AIH Agent architecture. Sentence-level agents A1–A4 identify person entities, extract temporal cues, classify textual function and detect retrospective passages. TimeBlock-level agents A5–A7 transform temporal markers, reconstruct temporal order and determine temporal granularity. A8 coordinates cross-text temporal propagation, and A9 converts eligible anchors and constraints into normalized temporal ranges. Preprocessing, TimeBlock assembly and summary generation are supporting pipeline steps; A1–A9 are the nine semantic agents.
Figure 5: Expanded TimeBlock construction and temporal ordering. a, Sentence-level preprocessing labels source temporal expressions, sinking statements, retrospective passages and ordinary narrative sentences. b, The labelled sequence is assembled into regular TimeBlocks and TimeBlocks containing retrospective passages. c, The Temporal Marker Buffer completes missing calendar elements from preceding TimeBlocks and local context. d, TimeBlocks containing retrospective passages are reordered from narrative into chronological order; when several blocks share a temporal anchor, the original hierarchical sentence number provides a secondary ordering key.
Figure 6: Expanded cross-text temporal propagation. a, A8 compares a source TimeBlock with target candidates and retains relations supported by evidence of the same event or historical phase. b, A9 converts the source evidence into normalized boundary constraints. c, A9 applies these boundaries in the target chronology, narrowing the candidate event range and updating adjacent open temporal ranges.
| Agent | Level | Principal function | Main output |
|---|---|---|---|
| A1 | Sentence | Person recognition and construction of cross-text traceable referent sets. | Person fields |
| A2 | Sentence | Extraction of source temporal expressions. | Source temporal expression |
| A3 | Sentence | Text-function classification and identification of descriptive or preparatory material for sinking. | Sinking flag |
| A4 | Sentence | Detection of retrospective passages and temporal displacement from the surrounding narrative sequence. | Retrospective-passage flag |
| A5 | TimeBlock | Conversion of abbreviated or relative expressions into comparable temporal markers. | Temporal marker |
| A6 | TimeBlock | Temporal ordering of TimeBlocks on a shared timeline. | Ordered sequence |
Table 1: Levels, functions and principal outputs of the nine AIH semantic agents.
| Field | Meaning |
|---|---|
| sentence_id | Stable sentence identifier in the form book_uuid.chapter_id.paragraph_id.sentence_id . |
| text | Sentence text. |
| character_flags | Boolean fields for candidate people such as liu_bang , xiang_yu and xiao_he . |
| has_time_expression | Whether the sentence contains an extractable source temporal expression. |
| time_expression | Extracted source wording, such as “秦二世元年七月” or “这时”. |
| is_sinking | Whether the sentence is assigned to the background layer of the principal temporal-reasoning chain. |
Table 2: Principal variables in sentence-level annotation.
| Field | Meaning |
|---|---|
| ID | Unique TimeBlock identifier, usually derived from the first sentence in the block. |
| timeblock_range | Sentence span covered by the TimeBlock. |
| Interlude | Whether the TimeBlock contains a retrospective passage. |
| Conversioninformation | Temporal-marker conversion record, including source wording, conversion requirement, evidence and rationale. |
| Granularity | Temporal granularity, with values 0 – 3 . |
| TM | Normalized temporal marker. |
Table 3: Principal variables in the TimeBlock representation.
Figure 7: Multi-model end-to-end results by case group in Experiment 1. a, Month-level MicroIoU for direct temporal localization (easy; C1–C2), within-document ambiguity reasoning (medium; C3–C4) and cross-document temporal propagation (hard; C5–C6). The two internal panels show AIH Agent and DirectLLM results for DeepSeek-V4-Flash (non-thinking mode), GPT-5.6 Sol, Gemini 3.1 Pro, Claude Opus 5 and Qwen 3.6; the human-only condition is the common reference line. b, Observed wall-clock time for the same groups and models on a logarithmic scale. AIH Agent values sum three independent runs per case; DirectLLM values are the observed times recorded in the run logs.
Figure 8: Case-level multi-model results in Experiment 1. a, Month-level MicroIoU for the human-only condition and each model–method combination in C1–C6, shown as percentages. b, Observed case-level wall-clock time. AIH Agent cells report the sum of three independent runs; direct-prompting cells report the observed times recorded in the run logs.
Figure 9: Diagnostic accuracy in Experiment 2. a, Overall strict accuracy for human responses, direct prompting and structured prompting across 48 unique diagnostic questions. b, Strict accuracy for temporal-information extraction and judgement, temporal-information reasoning, and cross-text event verification and temporal alignment. The figure label “cross-text temporal-information ordering” encompasses cross-text candidate-relation and temporal-alignment judgements. Human values aggregate 96 responses; model values are aggregated over three runs of the same 48 questions.
Figure 10: Diagnostic elapsed time in Experiment 2. a, Cumulative elapsed time for human responses, direct prompting and structured prompting. b, Elapsed time for temporal-information extraction and judgement, temporal-information reasoning, and cross-text event verification and temporal alignment. The figure label “cross-text temporal-information ordering” encompasses cross-text candidate-relation and temporal-alignment judgements. Time is shown in seconds on a logarithmic scale.
| Case | Sentence | Chapter | Passage | Internal reference range |
|---|---|---|---|---|
| C5 | 8.16.1 | Gaozu | In the third year of Qin Er Shi, King Huai of Chu saw that Xiang Liang’s army had been defeated, moved his capital to Pengcheng and combined the armies of Lü Chen and Xiang Yu. | -0207-10to-0206-11 |
| C5 | 7.10.1 | Xiang Yu | After the Chu army was defeated at Dingtao, King Huai moved to Pengcheng, combined the armies of Xiang Yu and Lü Chen and assumed command. | -0207-10to-0206-11 |
| C5 | 8.23.3 | Gaozu | In the eleventh month of the first Han year, Xiang Yu led the armies of the feudal lords westward to enter the passes, but the gate was closed. | -0206-11to-0206-12 |
| C5 | 7.17.3 | Xiang Yu | Xiang Yu heard that Pei Gong had taken Xianyang, became enraged and sent Lord Dangyang and others to attack the pass. | -0206-11to-0206-12 |
| C6 | 53.4.1 | Xiao He | After Liu Bang rose and became Pei Gong, Xiao He frequently assisted him in administering affairs. | -0209-10to-0208-10 |
| C6 | 53.4.2 | Xiao He | When Pei Gong entered Xianyang, Xiao He first secured the Qin laws, maps and household registers. | -0208-10to-0205-01 |
Table 4: Examples of internal reference temporal coordinates in the cross-text annotation.
| Sentence | Passage | Internal reference range |
|---|---|---|
| 53.4.1 | After Liu Bang rose and became Pei Gong, Xiao He repeatedly assisted him in administering affairs. | -0209-10to-0208-10 |
| 53.4.2 | When Pei Gong entered Xianyang, Xiao He first secured the laws, maps and household registers. | -0208-10to-0205-01 |
| 53.4.3 | When Pei Gong became King of Han, he appointed Xiao He as chancellor. | -0205-01to-0205-10 |
| 53.5.1 | When the King of Han led his army east and pacified the Three Qin, Xiao He remained as chancellor to govern Ba and Shu. | -0205-10to+infinity |
Table 5: Cross-text internal reference coordinates for Xiao He passages in C6.
Figure 11: Text upload and monitored AIH processing. Users upload biographical texts and inspect the process that generates source-traceable chronology and temporal-range views.
Figure 12: Source-grounded question answering. The interface supports questions over AIH-structured historical material while retaining access to the underlying evidence.
Figure 13: Bilingual translation. The interface provides AI-assisted Chinese–English translation to support reading of AIH-organized chronological material.
Figure 14: Biographical-draft generation. The interface generates a readable draft from AIH-structured chronological material for continued scholarly review and revision.
Figure 15: Time-point source inspection. A chronological point can be traced to its source sentence and temporal evidence for examination by historians.
Figure 16: AI-assisted revision review. The interface examines a submitted revision and its rationale, returns an AI recommendation and records the final human judgement.
Figure 17: Platform corpus coverage. The interface lists processed corpus collections and document tracks spanning East Asian historiographical materials and modern histories of science and technology.
| Corpus collection | Texts | Sentences | TimeBlocks | Temporal anchors | Normalized ranges | Cross-text relations |
|---|---|---|---|---|---|---|
| Samguk Sagi | 3 | 73 | 23 | 20 | 23 | 19 |
| Sanguozhi | 3 | 345 | 57 | 19 | 57 | 10 |
| History of Yuan | 3 | 7,187 | 3,405 | 950 | 3,405 | 36 |
| History of the Northern Dynasties | 3 | 930 | 174 | 66 | 174 | 26 |
| Book of Northern Qi | 3 | 639 | 216 | 122 | 216 | 57 |
| History of the Southern Dynasties | 3 | 433 | 30 | 1 | 30 | 14 |
Table 6: Westlake Historian corpus coverage and structured outputs as of 10 July 2026.
| Chapter | TimeBlocks | Non-empty ISO | Cross-text update records | Summaries |
|---|---|---|---|---|
| Basic Annals of Xiang Yu | 88 | 77 | 57 | 88 |
| Basic Annals of Gaozu | 166 | 139 | 85 | 166 |
| Hereditary House of Chancellor Xiao | 30 | 23 | 12 | 30 |
| Total | 284 | 239 | 154 | 284 |
Table 7: Final TimeBlock output statistics for the three Shiji chapters.
| Chapter | ID | Normalized time | summary |
|---|---|---|---|
| Xiang Yu | 7.3.1 | -0209-07-01 | In the seventh month of the first year of Qin Er Shi, Chen She and others launched the uprising at Daze Village, beginning the late-Qin rebellions. |
| Xiang Yu | 7.5.14 | -0209-09-01 | Pei Gong raised troops at Pei and led an attack on Xue. |
| Gaozu | 8.23.3 | -0206-11to-0206-12 | In the eleventh month of the first Han year, Xiang Yu led the armies of the feudal lords west and was blocked at Hangu Pass. |
| Gaozu | 8.26.2 | -0205-01to-0205-04 | Xiang Yu made Pei Gong King of Han, ruling Ba, Shu and Hanzhong from Nanzheng. |
| Xiao He | 53.4.1 | -0209-10to-0208-10 | After Liu Bang rose as Pei Gong, Xiao He repeatedly assisted him in administration and became a principal aide. |
| Xiao He | 53.4.2 | -0208-10to-0205-01 | When Pei Gong entered Xianyang, Xiao He first preserved the laws, maps, registers and archival materials. |
Table 8: Representative TimeBlock output entries.
| Chinese term | English usage | Definition in this study |
|---|---|---|
| 人物编年（年谱） | biographical chronology | Chronological organization of a person’s life, activities and relationships; nianpu is the traditional Chinese form discussed here. |
| 传统人物年谱 | traditional biographical chronology | Concise chronology centred on the principal events of a person’s life. |
| 年谱长编 | extended biographical chronology | Evidence-rich chronology incorporating source excerpts, quotations and editorial notes. |
| 多人物合谱 | multi-person chronology | Combined chronology representing interactions among several historical people. |
| 编年 | chronology | Events arranged by time; here, a method or temporal structure. |
| 编年体 | annalistic format | Historiographical format organized primarily by chronological sequence. |
Table 9: Chinese and English terminology used in the manuscript.
| Chinese title | English usage | Scope and note |
|---|---|---|
| 史记 | Shiji ( Records of the Grand Historian ) | Historical text used in all method demonstrations; Shiji is used after first definition. |
| 项羽本纪 | Basic Annals of Xiang Yu | Shiji , chapter 7; Xiang Yu’s biographical perspective. |
| 高祖本纪 | Basic Annals of Gaozu | Shiji , chapter 8; Liu Bang’s biographical perspective. |
| 萧相国世家 | Hereditary House of Chancellor Xiao | Shiji , chapter 53; Xiao He’s biographical perspective. |
| 二十四史 | Twenty-Four Histories | Canonical collection of Chinese official histories in the annals-biographies format. |
| 三国史记 | Samguk Sagi | Korean history written in the annals-biographies format. |
Table 10: Historical texts and chapter titles cited in the manuscript.
| Person | Chinese form | English usage | Relation and note |
|---|---|---|---|
| Liu Bang | 刘邦 | Liu Bang | Personal name and default manuscript usage. |
| 高祖 | Gaozu | Temple name; Emperor Gaozu of Han when imperial status is foregrounded. | |
| 汉王 | King of Han | Royal title before the establishment of the Han empire. | |
| 沛公 | Pei Gong | Early title in the uprising narrative; also translated as Duke or Lord of Pei. | |
| 刘季 | Liu Ji | Early name. | |
| Xiang Yu | 项羽 | Xiang Yu | Conventional name and default manuscript usage. |
Table 11: Historical people, aliases and titles used in the manuscript.
More in this topic
Historical Document Processing
Historical document processing uses AI to transcribe, structure, restore, and analyze manuscripts, books, newspapers, and archival records.
See all 76 papers