Paper deep dive
Beyond Similarity: Grounded Agentic Extraction and Expert-Adjudicated Evaluation of Intertextuality in Classical Chinese Histories
Zhaoji Wang, Wanyu Si, Jun Wang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 8/1/2026, 1:49:29 AM
Summary
This paper introduces a grounded agentic extraction framework for fine-grained intertextuality in Classical Chinese histories, moving beyond simple similarity scores. The method uses Large Language Models (LLMs) to identify exact character spans and classify reuse across five dimensions (form, aspect, source-marking, function, stance). The approach is validated against an expert-adjudicated benchmark of 2,533 pairs between the Analects and the Book of Han, and scaled to the full Twenty-Four Histories, revealing corpus-level structural patterns and a reliability gradient in expert agreement.
Entities (8)
Relation Signals (5)
Book of Han → cites → Analects
confidence 98% · In the Book of Han (first century CE), the abdication charge from the Analects... is transcribed almost verbatim.
LLM Agent → performs → Intertextuality Extraction
confidence 95% · We recast fine-grained intertextuality extraction as an agentic task in which a large language model (LLM) reads two text units in full
LLM Agent → uses → Five-Dimension Typology
confidence 93% · label it under a five-dimension typology of reuse (form, aspect, source-marking, function, stance)
Twenty-Four Histories → contains → Intertextual Pairs
confidence 90% · Scaling the validated extractor to the full Twenty-Four Histories (65,380 comparisons, 5,766 pairs) recovers corpus-level structure
Gu Jiegang → launched → Doubting Antiquity Movement
confidence 85% · the thesis that launched the Doubting Antiquity movement (Gu 1923).
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Computational approaches to intertextuality have advanced from string matching to neural retrieval, yet their outputs, similarity scores and parallel-passage lists, identify where texts reuse one another without characterizing how or why. We recast fine-grained intertextuality extraction as an agentic task in which a large language model (LLM) reads two text units in full and, through a constrained tool interface, must ground each proposed reuse in exact character spans on both sides and label it under a five-dimension typology of reuse (form, aspect, source-marking, function, stance). We validate the approach on an exhaustive comparison of the Analects with the Book of Han, where three domain experts adjudicate a pooled multi-model candidate set into a benchmark of 2,533 intertextual pairs. Against this standard we study twelve LLMs, reporting precision (56%-93%), a 51$\times$ cost spread at comparable quality, and how well their confidence is calibrated. Expert agreement traces a reliability gradient: dimensions legible on the textual surface are annotated consistently, while those requiring inference of intent are contested, delimiting the claims such annotation supports. Scaling the validated extractor to the full Twenty-Four Histories (65,380 comparisons, 5,766 pairs) recovers corpus-level structure a similarity score cannot express. The interpretive composition of citation shows no systematic change across eighteen centuries, yet the same passage is quoted less and less literally. Stability in the aggregate with drift in the individual case is what a cultural-attraction account expects. We release the extraction protocol and the expert-adjudicated benchmark.
Tags
Links
- Source: https://arxiv.org/abs/2607.27595v1
- Canonical: https://arxiv.org/abs/2607.27595v1
Trouble viewing inline? Open PDF directly →
Full Text
51,019 characters extracted from source content.
Expand or collapse full text
Beyond Similarity: Grounded Agentic Extraction and Expert-Adjudicated Evaluation of Intertextuality in Classical Chinese Histories Zhaoji Wang1,2, Wanyu Si1,2, Jun Wang1,2,* Abstract Computational approaches to intertextuality have advanced from string matching to neural retrieval, yet their outputs, similarity scores and parallel-passage lists, identify where texts reuse one another without characterizing how or why. We recast fine-grained intertextuality extraction as an agentic task in which a large language model (LLM) reads two text units in full and, through a constrained tool interface, must ground each proposed reuse in exact character spans on both sides and label it under a five-dimension typology of reuse (form, aspect, source-marking, function, stance). We validate the approach on an exhaustive comparison of the Analects with the Book of Han, where three domain experts adjudicate a pooled multi-model candidate set into a benchmark of 2,533 intertextual pairs. Against this standard we study twelve LLMs, reporting precision (56%–93%), a 51× cost spread at comparable quality, and how well their confidence is calibrated. Expert agreement traces a reliability gradient: dimensions legible on the textual surface are annotated consistently, while those requiring inference of intent are contested, delimiting the claims such annotation supports. Scaling the validated extractor to the full Twenty-Four Histories (65,380 comparisons, 5,766 pairs) recovers corpus-level structure a similarity score cannot express. The interpretive composition of citation shows no systematic change across eighteen centuries, yet the same passage is quoted less and less literally. Stability in the aggregate with drift in the individual case is what a cultural-attraction account expects. We release the extraction protocol and the expert-adjudicated benchmark. Introduction In the Book of Han (first century CE), the abdication charge from the Analects, “” (the heaven-ordained succession rests upon your person), is transcribed almost verbatim. A few chapters later, in the biography of the usurper Wang Mang, the same line reappears as “”, the second-person pronoun quietly rewritten into the first. By the 260s, in the abdication documents of the Records of the Three Kingdoms, only a formulaic skeleton survives, “”, detached from the Analects and fused into the standard script of dynastic transfer. One canonical sentence, three degrees of fidelity, spread across three centuries of official historiography. chunk A (Analects book)… “” … chunk B (history scroll)… “” … LLM agentreads both chunks in full,proposes candidate reuses find_positionexact search →[,)[start,end) + context(a) one extraction taskover one chunk pair add_pair: 3 checks schema: 5-dim labels ⋅· spans ⋅· confidence verbatim re-slice: chunk[,)[start,end) == fragment, char-for-char six-tuple dedup: the same intertextual pair, keyed by (Atext_A, Astart_A, Aend_A, Btext_B, Bstart_B, Bend_B), cannot enter twice any check fails ⇒ pair rejected, error returnedpasspass list / removeinspect & retract committedintertextual pairsboth chunks enter the model’s context whole:one Analects book, one full history scroll.No sentence splitting, no embedding retrieval—no similarity filter decides what the model sees. committed intertextual pairAnalects Book 13 → Book of Han scroll 6A[876,883)A\,[876,883): “”B[3156,3164)B\,[3156,3164): “” form:paraphrase aspect:content marking:marked function:application stance:positive evidence:medium confidence:0.72(b) verifiable outputsThe two fragments share zero characters: string similarity scoresthis real intertextual pair 0.00, and dense retrieval ranks itssource #232 of 1,470—only reading recovers how and why.(c) task termination123×n×\,n(n≥0)(n\!≥\!0)nothing leftto add?a run ends: n committed pairscall submitfirststop guard: no exit but submitstatuscheckmatch(n≥1)(n\!≥\!1)no_match(n=0)(n\!=\!0) any other end: failedsubmitdecides the outcometooltooltoollocateoffsetsrevisestateproposerejectaccept(d) validate-then-scale validate: 2,400 chunk-pairtasks × 12 LLMs scale: 24 histories, 65,380 tasks→ 5,766 intertextual pairs → clusters6,018 proposed intertextual pairs→ 3,489 pooled candidate pairs select extractor:precision ⋅· cost ⋅· calibration 2 experts + arbiter→ 2,533 gold intertextual pairs Figure 1: Framework overview. (a) The extraction agent for one chunk pair and its task tools (names abbreviated): find_position grounds candidate wordings in exact offsets; list / remove support revision; add_pair commits a pair only past the three checks shown. (b) The committed intertextual pair shown is real and shares zero characters across its fragments: an edict of Emperor Wu invoking Analects 13 by the sage’s name alone (cf. Beyond overlap). (c) A task can end only through submit (a stop guard blocks any earlier exit), and submit itself checks the declared status against the committed set; any other end is recorded as failed. (d) The validate-then-scale pipeline. A text-reuse detector can find that these passages resemble the Analects; it cannot say that the first is a faithful citation, the second an appropriation of authority, the third a dead formula. This is the gap we address, and it is consequential: a similarity score answers none of the questions historians ask of this corpus. Over roughly two millennia, China’s official histories were compiled by authors steeped in the classics, who quoted, paraphrased, and silently absorbed the Analects into judgments, memorials, and narrative alike. Tracing that reuse by hand is the labor of concordance scholarship, and it has carried arguments of the first rank. Gu Jiegang dated the strata of the legendary past by setting text against text. The Odes know Yu but not Yao and Shun. The Analects, later, praises both. The chapter that first chains them into a succession, the source of this paper’s opening abdication charge, he read as a late imitation of archaic style. Each age, he argued, extended the history it received, the thesis that launched the Doubting Antiquity movement (Gu 1923). Comparison of this kind is bounded by sustained expert attention, and in classical studies that attention is scarce. Gu himself planned to proceed one book at a time. It does not scale. Nor do the standard computational substitutes. On our data, retaining every human-confirmed reuse lets dense and sparse embedding prefilters discard only 2.5% and 1.1% of chunk pairs (0.3% combined), because literal and distributional similarity systematically miss paraphrase, precisely the cases that matter. The limit is representational: routing every comparison through one similarity value keeps only the degree of reuse and discards its kind. We therefore formalize fine-grained intertextuality extraction as span-grounded relation extraction between two texts. Given a pair of text units, a model must locate every reused fragment on both sides as an exact character span, and label each resulting intertextual pair under a five-dimension typology of reuse. This is a localization-and-classification task with verifiable outputs, not a similarity score. A prediction is correct only if the fragment texts and their in-unit character offsets match. We instantiate it as an agentic task in which an LLM, constrained by a tool interface, commits only span-verified, schema-valid pairs. Because exhaustive expert judgment is affordable only at small scale, we adopt a validate-then-scale design. We validate on the Analects against the Book of Han, where three domain experts adjudicate a pooled candidate set to a gold standard and twelve LLMs are measured against it. We then apply the validated extractor to the full Twenty-Four Histories and read its output as a distribution of calibrated model judgments. Our contributions are: (1) a verifiable extraction protocol for fine-grained intertextuality, executed by a tool-constrained LLM agent (Fig. 1), where annotations exist only as commitments that are span-verified, schema-checked, and de-duplicated at write time, abstention is explicit, and execution is reproducible and hermetic, so corpus-scale output is auditable rather than a parse of model prose; (2) an expert-adjudicated benchmark and a twelve-model study, with 2,533 gold pairs adjudicated from 3,489 triple-annotated candidates over the Analects–Book of Han, used to compare twelve LLMs on precision, cost, and calibration, and to surface a reliability gradient in expert agreement across the five dimensions; and (3) a validated large-scale application to all Twenty-Four Histories (65,380 comparisons, 5,766 pairs) that recovers structure a similarity score cannot express, with the interpretive composition of Analects citation showing no systematic change across eighteen centuries while literal fidelity to a fixed passage declines. Related Work Text reuse across traditions. Computational study of intertextuality began with surface matching and sequence alignment, pursued largely within separate philological traditions (for surveys, see Duan 2025; Sommerschield et al. 2023). An early computational model characterized reuse in the Greek New Testament (Lee 2007), and for Latin poetry the Tesserae project matches shared words and lemmata to surface parallels (Coffee et al. 2013). Alignment-based systems detect reprinted passages in nineteenth-century newspapers (Smith, Cordell, and Dillon 2013), parallel passages in Hebrew–Aramaic and Buddhist Chinese corpora (Shmidman, Koppel, and Porat 2018; Nehrdich 2020), related verses in the Qur’an (Sharaf and Atwell 2012), and reuse in Chinese corpora with a language-agnostic aligner (Vierthaler and Gelein 2019). For premodern Chinese, Sturgeon (2018) detects text reuse across the transmitted early corpus, the Evol line scales sentence-embedding retrieval to millions of intertextual pairs and mines the resulting networks for cultural-evolution questions (Duan et al. 2023; Wang et al. 2024a), and, closest to us in corpus, Deng, Yang, and Wang (2022) compare the Shiji and Book of Han through intertextual pairs. These methods scale, but operate on literal or distributional overlap, return passage-level candidates, and assign no typology of how a source is reused. Adding distributional semantics widens coverage, monolingually and across languages (Scheirer, Forstall, and Coffee 2016; Manjavacas, Long, and Kestemont 2019; Burns et al. 2021; Riemenschneider and Frank 2023), without changing this picture. From detection to a typology of reuse. That reuse comes in kinds is recognized (Forstall and Scheirer 2019), and a digital-humanities strand argues for moving “from quantitative to qualitative analysis” and toward a computational hermeneutics of how, not how much, texts are reused (Roe 2024; Moritz et al. 2016). LLMs have very recently been turned on intertextuality. At NLP venues, Yang et al. (2025) use LLMs to build a dataset of intertextual relations among philosophical texts, and Periti et al. (2024) benchmark how already-known reused passages are recontextualized. Both operate on relations between texts or given reuse pairs, not on extracting and localizing reuse. Extraction itself remains in preprints and digital-humanities venues, and none of these efforts combine the pieces we do. Loci Similes (Schelb et al. 2026) pairs a directed corpus with a two-way typology and expert-checked gold, but operates at the segment level and localizes no character spans. Lau and McManus (2024) mine asymmetric intertextuality from large Chinese corpora with an LLM-assisted pipeline framed as retrieval, with no reuse typology. A retrieval framing is moreover bounded by its segmentation. Where several passages on one side answer to one on the other, segment-to-segment matching cannot represent the relation, a structure our corpus-scale run surfaces below. Umphrey, Roberts, and Roberts (2024) prompt an LLM to emit intertextual pairs for expert validation, without span localization or adjudicated gold. Cameron (2026) traces biblical allusion in a modern novel. No prior system, in any tradition, combines generative-LLM extraction, two-sided character-span localization, a multi-dimension reuse typology, and an expert-adjudicated gold standard with reported agreement, applied at corpus scale. We target that combination, and validate it before applying it at scale. A further difference is architectural: these systems read conclusions out of model prose, whereas our annotations exist only as tool commitments checked at write time, which is what makes corpus-scale output auditable rather than parsed. Task and Agentic Extraction Protocol Task. We segment the two sources into text units (chunks): the Analects into its 20 books, each history into its scrolls. Given a chunk pair (A,B)(A,B), the task is to output every intertextual pair: a fragment of A and a fragment of B, each as an exact character span [,)[start,end), together with a label along five interpretive dimensions: form (direct quotation vs. paraphrase), aspect (content vs. structure), source-marking (explicit, marked, unmarked), function (background, support, application, critique), and stance (positive, neutral, negative, complex). Intertextual reuse is defined strictly as traceable textual dependence: shared topic, agreement, common vocabulary, or diffuse influence are excluded. A prediction is correct only if both fragment texts occur verbatim at their stated offsets. Identity is likewise strict: two candidates are the same only if they agree on the chunk pair, both fragment texts, and both offset ranges. Endpoint punctuation differences and shifted span boundaries are distinct candidates. Precision is the share of committed pairs judged valid against the adjudicated gold standard. We cast this as an agentic task in the tool-using paradigm (Yao et al. 2023; Schick et al. 2023; Wang et al. 2024b). For each chunk pair, an LLM agent reads both texts in full and commits candidate pairs through a constrained tool interface. The contribution is the verifiability contract this interface enforces and what that contract makes measurable. Both chunks enter the context whole, a complete book against a complete scroll, rather than cut into sentences and paired by n-gram or embedding similarity (Sturgeon 2018; Duan et al. 2023; Wang et al. 2024a). An index surfaces only what it already scores as close (the bottleneck quantified above), and marking and rhetorical function are legible only in discourse that sentence-level pairing discards. Three design principles address the failure modes of using an LLM for this task. The concrete tool schemas, isolation settings, and control parameters are given in the supplementary material. P1: Verifiable grounding. An LLM asked for the exact reused spans on both sides produces fluent but sometimes unanchored or fabricated parallels. The agent must therefore ground every proposed pair in exact character offsets, obtained from a positioning tool, before it can be recorded. The positioning tool is exact substring search over a chunk, returning [,)[start,end) offsets with ± 30 characters of surrounding context, queried singly or in batch. At commit time, the protocol re-slices each chunk at the stated offsets and rejects any pair whose cited fragment does not reproduce verbatim. This turns plausible-looking output into checkable textual evidence, echoing the grounded-parallel emphasis of human-in-the-loop systems for ancient texts (Assael et al. 2025). In the twelve-model run below, agents issued 319,465 positioning calls (about 11 per task attempt), and the gates rejected 39% of add_pair calls, a quarter at the verbatim re-slice, another 15% at the schema check. Rejection tracks grounding, not quality. The lowest-precision model alone drew two-thirds of the re-slice rejections, yet 81% of attempts that hit any rejection still ended with an accepted pair. Ablating the positioning tool multiplies the gate-rejection rate ninefold and output tokens fivefold (supplementary ablation). P2: Tool-mediated structured commitment. Rather than parse conclusions from the model’s prose, the agent commits each annotation through a submission interface that enforces the schema and rejects duplicate pairs at commit time, duplicate identity being the six-tuple of both fragment texts and their offsets. Six task tools implement the protocol: single and batch position lookup, pair commitment, listing and removal of committed pairs (the only route to revising an accepted pair), and a final submission call. A task’s output is exactly the set of committed pairs. The model’s final free text is never parsed. Execution follows the same discipline: each task runs hermetically, with a fresh temporary home and no inherited configuration, and is keyed by a content hash of every input that determines its result, including provider identity and endpoint, so the same model name served by two providers cannot silently reuse cached results. Configuration drift between runs is an error. P3: Enforced completion and explicit abstention. At corpus scale every task must end in a well-defined outcome so that aggregate statistics are trustworthy. Every task is classified into exactly one of three outcomes: the agent commits at least one grounded pair (a match), explicitly declares that the two chunks share no reuse (abstention), or is recorded as a failure. Concretely, a stop guard blocks the agent from terminating before it has submitted. A repetition guard detects identical tool calls returning identical results and injects corrective feedback from the third repetition onward. The final submission carries an explicit status, checked against the committed set in both directions: declaring a match requires at least one committed pair, and declaring no_match requires none, so the agent cannot hedge by abstaining while holding pairs. A task that ends any other way (an API or transport error, exhausted retries, a timeout, or a run that stops before submitting) is recorded as failed, never coerced into an empty result. Transient rate-limit errors are retried with deterministic backoff. Because most chunk pairs share no reuse, explicit abstention is a first-class outcome. Expert-Adjudicated Evaluation Data. Source A is the Analects (20 books, 22,919 characters), and source B is the Book of Han (120 scrolls, 942,693 characters), both from a publicly available punctuated digital edition (Shidian Guji). Their exhaustive pairing yields 2,400 chunk-pair tasks. Annotation and adjudication. Twelve LLMs (below) each processed all 2,400 tasks, proposing 6,018 candidate pairs. De-duplicated under a strict fragment-pair criterion (same source pair, same chunk pair, same fragment texts and same in-chunk offsets), these yield 3,489 distinct candidates. Pooling across all twelve models keeps the benchmark from favoring any single model’s proposal distribution. The pool is genuinely a union, with 68% of candidates proposed by exactly one model. Manual enumeration offers no substitute reference. Two philological surveys count Analects quotation in the Book of Han; one reports 237, the other 447 (Zhang 2008; Wang 2014). The totals differ by nearly a factor of two, and both authors expect to have missed cases. Even the larger falls well below the 954 marked pairs the experts confirm on this source pair alone. The limit is not effort but sustained attention over a long text, which is what an exhaustive machine pass supplies. Two experts independently annotated all 3,489 on a purpose-built web interface that shows each candidate as its two fragments highlighted within fixed context windows: first a validity judgment (yes/no/uncertain), then, for valid pairs, the five interpretive dimensions, all forced single-choice. A third expert adjudicated every item on which the two disagreed, seeing both independent annotations and the fields in conflict before issuing the final judgment. This yields the adjudicated gold standard: 2,533 intertextual pairs, 72.6% of the pool. Inter-annotator agreement is computed only from the two independent annotations, before adjudication (Cohen 1960; Artstein and Poesio 2008). Retrieval baselines: an oracle upper bound. The adjudicated set also bounds what any retrieve-then-judge baseline could recover, since a judge only classifies what retrieval surfaces: 25.8% of the 2,533 pairs share no character trigram between their fragments (40.8% share no 4-gram). An edict of Emperor Wu justifies shifting policy priorities by how Confucius answered each duke differently, “”, naming the sage yet sharing not one character with the teaching it invokes (“”, Book 13). Ban Gu’s rhapsody in the autobiographical postface compresses the ford encounter of Book 18 into “”, recluse and disciple each a single name-character, unmarked. We give dense retrieval an oracle advantage no deployed pipeline has: BGE-M3 (Chen et al. 2024) is queried with the expert-identified history-side fragment itself. Even so, it ranks an overlapping source sentence top-10 for only 61.0% of trigram-free pairs (top-1: 33.0%), against 91.7% for verbatim-anchored ones. The edict and the rhapsody rank #232 and #470 among 1,470 Analects sentence units. The stratum’s median rank is 4, against 1 for verbatim-anchored pairs. These oracle hit rates cap the recall of any retrieve-then-judge pipeline exactly where interpretation begins (the fragment-level counterpart of the chunk-pair bound above). Moreover, retrieval output carries neither spans nor labels. Validity: a threshold gap, closed by adjudication. On the validity judgment the two experts differ in strictness, not in criterion. One accepted 55.3% of candidates, the other 99.4%. Their decisions nest perfectly: no candidate was accepted by the stricter expert and rejected by the lenient one. The disagreement is a one-sided threshold gap of 1,539 candidates, adjudicated item by item, the arbiter siding with the stricter expert on 59.9% and accepting 617. Label conflicts also sent 1,214 of the both-accepted candidates to the arbiter, who overturned 14 of them to invalid. The arithmetic closes: 1,930−14+617=2,5331,930-14+617=2,533. Marginals this skewed make chance-corrected agreement coefficients uninformative (Feinstein and Cicchetti 1990). The check that matters is robustness. Rescored against the stricter expert alone, all twelve models’ precisions drop by 9–23 points. However, the ranking (Spearman ρ=0.94ρ=0.94, no move over two places) and the deployment choice are unchanged. Absolute precision inherits the adjudicated threshold, and the comparison does not depend on it. A reliability gradient. Agreement on the five interpretive dimensions, computed over the 1,930 candidates both experts accepted, is not uniform (Table 1). Dimensions whose evidence sits on the surface of the text (is the wording identical, is the source named) are settled by inspection, while function and stance, which require reconstructing intent and attitude, admit principled disagreement between competent experts. Raw agreement is read against chance: aspect’s 98.9% rides a 97%-content base rate (chance 96.1%), so its label is reliable but nearly constant. The gradient bounds what the annotations support. The distributions of form and source-marking may be discussed with confidence, while function and stance serve only as exploratory signals. The disagreement is a property of the task, not a failure of annotation. Moreover, the models reproduce the gradient where it is informative. Relative to each model’s majority-class baseline, label accuracy lifts by 18–37 points on form and 25–43 on source-marking, by 12–25 on function, and by at most 2 on aspect and 9 on stance (per-model table in the supplementary material). One contested item makes this concrete: both experts accepted the pair linking the Analects phrase “” (diligent and fond of learning) to the Book of Han scene where a courtier defends the crown prince: “what is called talent is: diligent and fond of learning, reviewing the old to know the new, and such is the crown prince” (the second phrase itself quotes Analects 2, a two-source composite). Both agreed it is an unmarked verbatim quotation with positive stance. They split only on why: canonical support for the courtier’s claim, or application of the canon’s standard of talent to a living person (the arbiter ruled application). Nothing on the textual surface settles this, and 345 items in the pool share this shape: pair accepted, form and source-marking agreed, function split between support and application. Dimension Obs. agr. Chance Cohen’s κ Form (quotation/paraphrase) 87.3% 56.1% 0.71 Aspect (content/structure) 98.9% 96.1% 0.70 Source-marking (explicit/marked/unmarked) 85.9% 39.8% 0.77 Function (background/support/…) 58.9% 39.5% 0.32 Stance (positive/neutral/…) 77.0% 68.4% 0.27 Table 1: Inter-annotator agreement on the five interpretive dimensions, over the 1,930 candidates both experts accepted (pre-adjudication). Chance: expected agreement from the annotators’ marginals, the baseline κ corrects for. Form and source-marking are reliable; function and stance are contested; aspect is reliable but near-constant. Twelve-model study. Against the gold standard we score twelve recent LLMs from seven providers (among them DeepSeek-AI 2026; GLM-5 Team 2026; OpenAI 2026; Anthropic 2025) (Fig. 2; per-model counts in Table 2). Precision ranges from 55.9% to 92.9%. Eight of twelve reach at least 82%, so the strict protocol is executable for most of the models tested. Quality and efficiency vary independently. At comparable precision, per-valid-pair cost spans a 51× range, and the chosen deployment model attains the highest precision-per-dollar. The top-precision model is also the slowest (median 521s) and the most expensive per valid pair. One of its valid pairs buys fifty-one from the deployment model at 4.9 points lower precision. Meanwhile, the fastest models (37–39s) span a wide precision range, so precision, cost, and latency are three separate axes. Verbalized model confidence (Guo et al. 2017; Tian et al. 2023), pooled over the twelve models, discriminates valid from invalid candidates (AUC 0.85). Candidates at confidence ≥ 0.9 are 96.6% valid (the deployment model alone: AUC 0.84, 96.1%). However, models are systematically over-confident at lower scores (Fig. 3). Finally, inter-model agreement is a strong validity signal. Candidates proposed by ≥ 2 models are 88.5% valid, against 65.1% for single-model ones. Models find different slices of the intertextual space, which the pooled evaluation aggregates. A single pass already rivals hand enumeration. Philological surveys of this source pair counted 237 and 447 quotations by hand (Zhang 2008; Wang 2014), while each model alone commits 323 to 572 adjudicated-valid pairs. Figure 2: Twelve LLMs on the 2,533-pair gold standard: precision against cost per valid pair (log scale). Quality and efficiency vary independently; the selected deployment model (deepseek-v4-flash) attains the best precision-per-cost (counts in Table 2). Model Cand. Valid Prec. Cost ($) $/valid Lat. (s) doubao-2.1-turbo 367 341 92.9% 853 2.501 521 deepseek-v4-pro 426 394 92.5% 42 0.107 101 glm-5-turbo 460 410 89.1% 215 0.525 126 deepseek-v4-flash 443 390 88.0% 19 0.049 131 doubao-2.1-pro 495 426 86.1% 241 0.566 183 qwen3.7-plus 492 416 84.6% 68 0.163 178 gemini-3.1-flash-lite 384 324 84.4% 115 0.355 37 glm-5.2 583 478 82.0% 98 0.205 162 qwen3.7-max 520 414 79.6% 122 0.294 194 gpt-5.4-mini 455 350 76.9% 227 0.648 39 gpt-5.5 815 572 70.2% 1,029 1.799 38 claude-haiku-4.5 578 323 55.9% 647 2.003 163 Total 6,018 4,838 3,676 Table 2: Per-model results on the 2,400 evaluation tasks: candidates, valid pairs, precision (validity), total cost, cost per valid pair, and median latency. Precision is the share of committed pairs judged valid in expert adjudication. Figure 3: Calibration of verbalized confidence, pooled over the twelve models: share of candidates judged valid per confidence bucket. Scaling to the Twenty-Four Histories We run the selected extractor (deepseek-v4-flash; DeepSeek-AI 2026) over the full Twenty-Four Histories: the Analects paired exhaustively with every history scroll, 65,380 chunk-pair tasks over ≈ 28.5M characters whose composition dates span eighteen centuries (91 BCE–1739 CE), with no failed tasks, yielding 5,766 intertextual pairs. Because precision was measured only on the Book of Han, we read the output as a distribution of calibrated model judgments and state uncertainties explicitly. At the deployment model’s validated precision, roughly one pair in eight is expected to be spurious. Claims therefore rest on distributional structure rather than raw totals, and each pair’s confidence lets a reader threshold harder. A worked example. The introduction’s abdication charge forms a single merged-fragment cluster of 85 extracted pairs across 22 histories, spanning composition dates from 91 BCE to 1739 CE. Three members show what the labels add. In the Book of Han, the charge is quoted near-verbatim under the source’s name, “”, labeled quotation / explicit / support (similarity 0.87): the canon is named to lend authority. In the same history, an imperial patent appointing a commander absorbs the charge’s wording, “” (go, devote your whole heart … hold faithfully to the mean), labeled quotation / unmarked / application (similarity 0.37): investiture language repurposed as working government prose. In the Records of the Three Kingdoms, the abdication edict transfers the mandate with “” (Heaven’s favor ends forever; the succession lies with Jin), labeled paraphrase / unmarked / application (similarity 0.25): the dead formula of the introduction, now datable and countable. A similarity ranking orders these three by overlap alone. The labels separate public appeal to canonical authority, silent administrative repurposing, and formulaic script, and the cluster assembles the passage’s reception history. Within this one cluster, fidelity already trends downward with composition year, the pattern quantified across all clusters below (Fig. 4). The mix of citation shows no detectable drift; volume and fidelity do. Across the twenty-four histories we detect no systematic diachronic change in the interpretive composition of Analects citation (Table 3). Comparing histories composed by 659 CE against those from 945 CE onward, the source-marking mix is unchanged (unmarked 77.5% vs. 77.6%), as are the function and stance mixes. In addition, per-history usage profiles do not diverge with composition date (Jensen–Shannon divergence uncorrelated with the year gap, permutation p = .12–.28). The table tracks each dimension’s dominant category, the divergence test the full mix. Dimension Dominant category Early Late ρ p Form quotation 40.9% 42.4% −0.07-0.07 .73 Aspect content 98.1% 98.2% −0.28-0.28 .18 Source-marking unmarked 77.5% 77.6% 0.110.11 .60 Function support 67.8% 70.8% −0.11-0.11 .61 Stance positive 73.8% 79.8% 0.160.16 .44 Table 3: Diachronic stability of the interpretive composition. Early/Late: pooled share over histories composed by 659 CE vs. from 945 CE onward. ρ, p: Spearman correlation of the dominant category’s per-history share with composition year over the 24 histories, permutation p (2,000 permutations); no trend reaches significance. Raw citation density does fall about threefold, but this is a count-level trend obtainable without our framework, consistent with the received periodization (Pi 1959), so we rest no claim on it. What the framework adds is that the way the canon is cited shows no systematic change: its rhetorical form was fixed early and persisted, a stability only annotation can establish. This stability does not conflict with the constant rewriting of individual passages below. A cultural-attraction account expects a stable group-level distribution produced by convergent transformation, not faithful copying (Buskell 2017). Unmarked absorption as the durable norm is, in native terms, licensed practice: the “compositional” register, which Zhang (1832, Shuolin) distinguishes from the “evidential”: in composition one draws on old meanings and takes over existing text without marking the source, and this is no plagiarism, a license Zhang grants only where the author’s own argument stands without leaning on the borrowed words. The norm is older than Zhang’s formulation: Liu Xie’s sixth-century poetics already holds that old material, aptly used, reads “as if it came from one’s own mouth” (Liu 502, Shilei). The corpus shows that this ideal was, measurably and durably, the practice. Marking and fidelity are distinct axes. Aligning each history fragment to its Analects source (character-level longest common subsequence) shows that how literally a citation is quoted does not fix how it is marked. Named citations are more faithful than unattributed ones (median 0.78 vs. 0.38; Kruskal–Wallis p<10−3p<10^-3), and the gap survives controlling for form (0.85 vs. 0.50 among direct quotations), so a similarity score cannot stand in for the marking label, the point of a reuse typology. Naming a source (“the Master said”) is a public appeal to canonical authority that holds the wording to a higher standard; the relation is a regularity, not a law (unmarked citation can be verbatim when concealment is the motive; Edelstein, Morrissey, and Roe 2013). Function tracks fidelity; application is unattributed. Invoked as authoritative support, the canon is quoted more literally (mean similarity 0.50); turned to application, repurposed into governance and policy, it is looser (0.42) and 93% unattributed: the pragmatic use native scholarship names tong-jing zhi-yong, applying the classics to affairs, where a citation’s force is that the sage “would have approved” the decision, not its exact words (Yü 2021). Because function is a low-agreement dimension (Table 1), we report this as exploratory. Fidelity declines in later-compiled histories. Holding a passage fixed (a merged-fragment cluster), later-compiled histories show lower literal fidelity (Fig. 4): across 283 clusters spanning ≥ 5 distinct composition years, the mean Spearman correlation is −0.15-0.15 (69.3% negative; cluster-bootstrap 95% CI [−0.19,−0.12][-0.19,-0.12]; permutation p<0.001p<0.001), and the decline is steeper for direct quotation (−0.18-0.18) than paraphrase (−0.08-0.08). The pattern strengthens in the confidence ≥ 0.9 subset (172 clusters, ρ=−0.19ρ=-0.19, p<0.001p<0.001). That transmission necessarily alters wording is, again, a native observation (Zhang 1832, Shuolin): successive historians recast their sources’ wording not from a taste for difference but by the necessity of circumstance, which is the transformation-as-limiting-case view of cultural transmission (Sperber 1985). Figure 4: Literal fidelity declines with the citing history’s composition date, held constant per passage: distribution of per-cluster Spearman correlations of fidelity vs. year (boxes: interquartile range; diamonds: stratum mean; ∗ permutation p<0.05p<0.05; k qualifying clusters). Asymmetric reuse fixed segmentation cannot align. Reading each chunk in full, rather than matching pre-segmented sentences, also recovers reuse whose two sides differ in extent, the asymmetric case a fixed segmentation cannot represent (Lau and McManus 2024). In the Book of Han, a single clause praising Emperor Cheng’s bearing, “” (on mounting the carriage he stood straight; he did not look round, spoke not hastily, pointed at nothing), draws on two separate passages of Analects “Xiangdang”: “” (mounting the carriage, he stood straight and grasped the cord) and “” (in the carriage he did not look round, spoke not hastily, pointed at nothing). Both were accepted against the one history clause. A sentence-to-sentence ranking must pick one Analects sentence or split its score across the two; the many-to-one relation is expressible only when fragments are localized freely. It surfaces here as two accepted pairs sharing one history-side span, one of 117 groups in the corpus run where a single history span answers to two or more disjoint Analects passages. Discussion and Limitations Span-grounded records placed under a reuse typology turn detection into description: a corpus-scale run can now say what kind of reuse occurs, under what marking, to what rhetorical end. The object of interpretation here is the textual relation, not the model’s reasoning. Expert agreement bounds the account: form and source-marking are annotated consistently and carry the load-bearing findings, while function and stance are contested and reported as exploratory. First, corpus-scale extraction uses a single model calibrated on the Book of Han. The load-bearing findings persist and strengthen in the confidence ≥ 0.9 subset (55.3% of pairs; the deployment model’s ≥ 0.9 candidates are 96.1% valid on the benchmark), so they are not carried by lower-confidence pairs. Second, the pool is the union of twelve models’ candidates, as in pooled evaluation (Buckley and Voorhees 2004), so corpus totals are lower bounds. Third, the model may have absorbed prior scholarly identifications, so we frame the system as assistive. Fourth, each history enters the diachronic analyses with a single composition date, though several took decades, and the early/late split follows received periodization. Contamination also cuts unevenly: memorized parallels concentrate in famous, marked cases, while the corpus-scale value lies in unmarked paraphrase, which concordances do not list. An anonymized platform (URL withheld) shows every pair in context, with spans, in clusters tracing reception; the full pair set ships in the code-and-data supplement as JSONL, SQLite, TEI stand-off XML, and RDF/Turtle. The benchmark is a controlled evaluation setting, not a universal intertextuality benchmark; what travels is the protocol. Little is specific to the Analects or to Chinese: the task requires only a segmentable source pair and an annotation schema, and the three principles apply wherever reuse must be tied to exact spans. Next: ensemble extraction and a second adjudicated evaluation on a later history. Conclusion We recast fine-grained intertextuality extraction as a grounded, typology-based agentic task, built an expert-adjudicated benchmark whose reliability gradient marks which dimensions annotation can support, and applied the validated extractor at corpus scale. Across eighteen centuries the interpretive composition of citation shows no systematic change while the same passage is quoted ever less literally. Stability in the aggregate produced by transformation in the individual case, as a cultural-attraction account expects, and structure no similarity score can state. The finding is legible only because every pair is a write-time-checked commitment, so the run is audited rather than trusted. Comparison once proceeded one book at a time. It now keeps pace with the corpus. Acknowledgments This work was supported by the National Key Research and Development Program of China (“Research on Cross-Context Barrier-Free Interactive Technology and Equipment for Ethnic Regions,” No. 2025YFC3309300) and the National Natural Science Foundation of China (“The Construction of the Knowledge Graph for the History of Chinese Confucianism,” No. 72010107003). References Anthropic (2025) Anthropic. 2025. Claude Haiku 4.5 System Card. https://w.anthropic.com/claude-haiku-4-5-system-card. Artstein and Poesio (2008) Artstein, R.; and Poesio, M. 2008. Inter-Coder Agreement for Computational Linguistics. Computational Linguistics, 34(4): 555–596. Assael et al. (2025) Assael, Y. M.; Sommerschield, T.; Cooley, A.; Shillingford, B.; Pavlopoulos, J.; Suresh, P.; Herms, B.; Grayston, J.; Maynard, B.; Dietrich, N.; Wulgaert, R.; Prag, J.; Mullen, A.; and Mohamed, S. 2025. Contextualizing ancient texts with generative neural networks. Nature, 645(8079): 141–147. Buckley and Voorhees (2004) Buckley, C.; and Voorhees, E. M. 2004. Retrieval Evaluation with Incomplete Information. In Proceedings of the 27th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 25–32. ACM. Burns et al. (2021) Burns, P. J.; Brofos, J. A.; Li, K.; Chaudhuri, P.; and Dexter, J. P. 2021. Profiling of Intertextuality in Latin Literature Using Word Embeddings. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 4900–4907. Association for Computational Linguistics. Buskell (2017) Buskell, A. 2017. What are cultural attractors? Biology & Philosophy, 32(3): 377–394. Cameron (2026) Cameron, E. 2026. ‘The Order in the Horse’s Heart’: A Case Study in LLM-Assisted Stylometry for the Discovery of Biblical Allusion in Modern Literary Fiction. arXiv:2604.19447. Chen et al. (2024) Chen, J.; Xiao, S.; Zhang, P.; Luo, K.; Lian, D.; and Liu, Z. 2024. M3-Embedding: Multi-Linguality, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation. In Findings of the Association for Computational Linguistics: ACL 2024, 2318–2335. Association for Computational Linguistics. Coffee et al. (2013) Coffee, N.; Koenig, J.-P.; Poornima, S.; Forstall, C. W.; Ossewaarde, R.; and Jacobson, S. L. 2013. The Tesserae Project: intertextual analysis of Latin poetry. Literary and Linguistic Computing, 28(2): 221–228. Cohen (1960) Cohen, J. 1960. A Coefficient of Agreement for Nominal Scales. Educational and Psychological Measurement, 20(1): 37–46. DeepSeek-AI (2026) DeepSeek-AI. 2026. DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence. arXiv:2606.19348. Deng, Yang, and Wang (2022) Deng, Z.; Yang, H.; and Wang, J. 2022. A comparative study of Shiji and Hanshu from the perspective of digital humanities. In Proceedings of the 21st Chinese National Conference on Computational Linguistics (CCL), 656–670. Chinese Information Processing Society of China. Duan (2025) Duan, S. 2025. Quantitative Intertextuality from the Digital Humanities Perspective: A Survey. arXiv:2510.27045. Duan et al. (2023) Duan, S.; Wang, J.; Yang, H.; and Su, Q. 2023. Disentangling the cultural evolution of ancient China: a digital humanities perspective. Humanities and Social Sciences Communications, 10: 310. Edelstein, Morrissey, and Roe (2013) Edelstein, D.; Morrissey, R.; and Roe, G. 2013. To Quote or not to Quote: Citation Strategies in the Encyclopédie. Journal of the History of Ideas, 74(2): 213–236. Feinstein and Cicchetti (1990) Feinstein, A. R.; and Cicchetti, D. V. 1990. High agreement but low kappa: I. The problems of two paradoxes. Journal of Clinical Epidemiology, 43(6): 543–549. Forstall and Scheirer (2019) Forstall, C. W.; and Scheirer, W. J. 2019. Quantitative Intertextuality: Analyzing the Markers of Information Reuse. Springer. GLM-5 Team (2026) GLM-5 Team. 2026. GLM-5: From Vibe Coding to Agentic Engineering. arXiv:2602.15763. Gu (1923) Gu, J. 1923. Yu Qian Xuantong xiansheng lun gushi shu [Letter to Qian Xuantong on Ancient History]. First published in Dushu zazhi (1923); collected in Gushi bian [Debates on Ancient History], Vol. 1 (Pushe, Beijing, 1926). Guo et al. (2017) Guo, C.; Pleiss, G.; Sun, Y.; and Weinberger, K. Q. 2017. On Calibration of Modern Neural Networks. In Proceedings of the 34th International Conference on Machine Learning, 1321–1330. PMLR. Lau and McManus (2024) Lau, P. K.; and McManus, S. M. 2024. Mining Asymmetric Intertextuality. arXiv:2410.15145. Lee (2007) Lee, J. 2007. A Computational Model of Text Reuse in Ancient Literary Texts. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, 472–479. Association for Computational Linguistics. Liu (502) Liu, X. 502. Wenxin diaolong [The Literary Mind and the Carving of Dragons]. Ca. 501–502; chapter “Shilei” (on allusion) cited; verified against the Guangming Daily Press edition (Beijing, 2014), p. 142. Manjavacas, Long, and Kestemont (2019) Manjavacas, E.; Long, B.; and Kestemont, M. 2019. On the Feasibility of Automated Detection of Allusive Text Reuse. In Proceedings of the 3rd Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature (LaTeCH-CLfL), 104–114. Association for Computational Linguistics. Moritz et al. (2016) Moritz, M.; Wiederhold, A.; Pavlek, B.; Bizzoni, Y.; and Büchler, M. 2016. Non-literal text reuse in historical texts: An approach to identify reuse transformations and its application to Bible reuse. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, 1849–1859. Association for Computational Linguistics. Nehrdich (2020) Nehrdich, S. 2020. A Method for the Calculation of Parallel Passages for Buddhist Chinese Sources Based on Million-scale Nearest Neighbor Search. Journal of the Japanese Association for Digital Humanities, 5(2): 132–153. OpenAI (2026) OpenAI. 2026. GPT-5.5 System Card. https://openai.com/index/gpt-5-5-system-card. Periti et al. (2024) Periti, F.; Cassotti, P.; Montanelli, S.; Tahmasebi, N.; and Schlechtweg, D. 2024. TRoTR: A Framework for Evaluating the Re-contextualization of Text Reuse. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 13972–13990. Association for Computational Linguistics. Pi (1959) Pi, X. 1959. Jingxue lishi [History of Classical Studies]. Beijing: Zhonghua Shuju. Annotated by Zhou Yutong. Riemenschneider and Frank (2023) Riemenschneider, F.; and Frank, A. 2023. Graecia capta ferum victorem cepit. Detecting Latin Allusions to Ancient Greek Literature. In Proceedings of the Ancient Language Processing Workshop, 30–38. INCOMA Ltd. Roe (2024) Roe, G. 2024. Text reuse as cultural practice: intertextuality in the 18th-century digital archive. Digital Enlightenment Studies, 2: 1–30. Scheirer, Forstall, and Coffee (2016) Scheirer, W. J.; Forstall, C. W.; and Coffee, N. 2016. The sense of a connection: Automatic tracing of intertextuality by meaning. Digital Scholarship in the Humanities, 31(1): 204–217. Schelb et al. (2026) Schelb, J.; Wittweiler, M.; Revellio, M.; Feichtinger, B.; and Spitz, A. 2026. Loci Similes: A Benchmark for Extracting Intertextualities in Latin Literature. arXiv:2601.07533. Schick et al. (2023) Schick, T.; Dwivedi-Yu, J.; Dessì, R.; Raileanu, R.; Lomeli, M.; Hambro, E.; Zettlemoyer, L.; Cancedda, N.; and Scialom, T. 2023. Toolformer: Language Models Can Teach Themselves to Use Tools. In Advances in Neural Information Processing Systems 36 (NeurIPS). Sharaf and Atwell (2012) Sharaf, A.-B.; and Atwell, E. 2012. QurSim: A corpus for evaluation of relatedness in short texts. In Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC’12), 2295–2302. European Language Resources Association (ELRA). Shmidman, Koppel, and Porat (2018) Shmidman, A.; Koppel, M.; and Porat, E. 2018. Identification of Parallel Passages Across a Large Hebrew/Aramaic Corpus. Journal of Data Mining and Digital Humanities. ArXiv:1602.08715. Smith, Cordell, and Dillon (2013) Smith, D. A.; Cordell, R.; and Dillon, E. M. 2013. Infectious Texts: Modeling Text Reuse in Nineteenth-Century Newspapers. In 2013 IEEE International Conference on Big Data. IEEE. Sommerschield et al. (2023) Sommerschield, T.; Assael, Y.; Pavlopoulos, J.; Stefanak, V.; Senior, A.; Dyer, C.; Bodel, J.; Prag, J.; Androutsopoulos, I.; and de Freitas, N. 2023. Machine Learning for Ancient Languages: A Survey. Computational Linguistics, 49(3): 703–747. Sperber (1985) Sperber, D. 1985. Anthropology and Psychology: Towards an Epidemiology of Representations. Man (New Series), 20(1): 73–89. Sturgeon (2018) Sturgeon, D. 2018. Unsupervised identification of text reuse in early Chinese literature. Digital Scholarship in the Humanities, 33(3): 670–684. Tian et al. (2023) Tian, K.; Mitchell, E.; Zhou, A.; Sharma, A.; Rafailov, R.; Yao, H.; Finn, C.; and Manning, C. 2023. Just Ask for Calibration: Strategies for Eliciting Calibrated Confidence Scores from Language Models Fine-Tuned with Human Feedback. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 5433–5442. Association for Computational Linguistics. Umphrey, Roberts, and Roberts (2024) Umphrey, R.; Roberts, J.; and Roberts, L. 2024. Investigating Expert-in-the-Loop LLM Discourse Patterns for Ancient Intertextual Analysis. In Proceedings of the 4th International Conference on Natural Language Processing for Digital Humanities, 31–40. Association for Computational Linguistics. Vierthaler and Gelein (2019) Vierthaler, P.; and Gelein, M. 2019. A BLAST-based, Language-agnostic Text Reuse Algorithm with a MARKUS Implementation and Sequence Alignment Optimized for Large Chinese Corpora. Journal of Cultural Analytics, 4(2). Wang (2014) Wang, J. 2014. Hanshu yin Lunyu kao [A Study of the Book of Han’s Quotations of the Analects]. Master’s thesis, Qufu Normal University. Wang et al. (2024a) Wang, J.; Duan, S.; Fu, B.; Gao, L.; and Su, Q. 2024a. Evol project: a comprehensive online platform for quantitative analysis of ancient literature. Humanities and Social Sciences Communications, 11: 291. Wang et al. (2024b) Wang, L.; Ma, C.; Feng, X.; Zhang, Z.; Yang, H.; Zhang, J.; Chen, Z.; Tang, J.; Chen, X.; Lin, Y.; Zhao, W. X.; Wei, Z.; and Wen, J.-R. 2024b. A survey on large language model based autonomous agents. Frontiers of Computer Science, 18(6): 186345. Yang et al. (2025) Yang, Y.; Xu, Y.; Huang, C.; Jurgensen, J.; Hu, H.; and Wang, H. 2025. InterIDEAS: Philosophical Intertextuality via LLMs. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, 23151–23172. Association for Computational Linguistics. Yao et al. (2023) Yao, S.; Zhao, J.; Yu, D.; Du, N.; Shafran, I.; Narasimhan, K.; and Cao, Y. 2023. ReAct: Synergizing Reasoning and Acting in Language Models. In The Eleventh International Conference on Learning Representations (ICLR). Yü (2021) Yü, Y.-s. 2021. Confucian Culture vs. Dynastic Power in Chinese History. Asia Major, 3rd series, 34(1): 1–10. Zhang (1832) Zhang, X. 1832. Wenshi tongyi [General Principles of Literature and History]. Qing dynasty; chapter “Shuolin” cited; verified against the Guji Chubanshe edition (Beijing, 1956, punctuated by Liu Gongchun), p. 116–118. Zhang (2008) Zhang, Z. 2008. Cong Shiji, Hanshu, Hou Hanshu kan Lunyu zai Liang Han de Liuchuan [The Transmission of the Analects in the Two Han Dynasties as Seen from the First Three Histories]. Master’s thesis, Shandong University.