From Inline Notes to Collected Commentaries: Toward Context-Preserving Organization of Exegetical Knowledge in Classical Chinese Texts

# From Inline Notes to Collected Commentaries: Toward Context-Preserving Organization of Exegetical Knowledge in Classical Chinese Texts
Source: [https://arxiv.org/html/2607.29044](https://arxiv.org/html/2607.29044)
\\setCJKmainfont

FandolSong\-Regular\.otf

Ke Liang1, Qi Su2,3,4, Churen Huang1 1Department of Language Science and Technology, The Hong Kong Polytechnic University 2,3,4School of Foreign Languages; Institute for Artificial Intelligence; Research Center for Digital Humanities, Peking University \{leo\-ke\.liang,churen\.huang\}@connect\.polyu\.hk sukia@pku\.edu\.cn

###### Abstract

Inline notes and collected commentaries are important forms of scholarly communication that evolved within the Confucian exegetical tradition, yet have received little computational attention\. Drawing on traditional Chinese exegetics and philology, this paper formulates collected commentary compilation as an NLP task and proposes a computational framework that preserves the contextual dependency of inline notes while enabling their automatic compilation and exegetical knowledge organization\. It combines two\-step prompt chaining for identifying the associated main\-text segments and exegetical functions of annotations with cross\-source mention clustering for integrating commentary across editions, achieving a CoNLL F1score above 97% in a case study on theClassic of Mountains\. Our framework lays the foundation for the large\-scale organization of historical exegetical knowledge, thereby supporting a broad range of downstream philological and NLP tasks\.

## 1Introduction

Commentaries on classical texts are explanatory texts that developed alongside the transmission of these works, aiming to interpret their linguistic\(Ling,[2021](https://arxiv.org/html/2607.29044#bib.bib49)\)and historical meanings\(Dong,[1993](https://arxiv.org/html/2607.29044#bib.bib20)\)\. They not only facilitate the understand of ancient culture, but also play an important role in natural language processing \(NLP\) for ancient text, yet this potential remains largely underexplored\. In this paper, we explore the computational organization of ancient commentaries based on their intrinsic characteristics\. The resulting resources could be used for both philological downstream tasks and NLP applications\.

The tradition of annotating classical texts can be traced back to the Spring and Autumn period \(770\-476 BCE\)\. Owing to the Confucian emphasis on reverence for antiquity, it became the dominant mode of scholarly writing from the Western Han dynasty \(206 BCE\-9 CE\) throughout the premodern period\(Gardner,[1998](https://arxiv.org/html/2607.29044#bib.bib26)\)\. Unlike the Western tradition of marginalia or footnotes, Chinese classical commentaries have, since the Han Dynasty, been embedded within the main text as inline notes \(隨文注釋\) and differentiated from it by smaller double\-column script\(Wilkinson,[2018](https://arxiv.org/html/2607.29044#bib.bib94)\)\. As commentaries accumulated, scholars often needed to consult multiple commentarial editions \(注本\), each containing a distinct body of commentary, to obtain a comprehensive understanding of a classical text, a process that was both complicated and time\-consuming\. To address this challenge, scholars in the Eastern Han \(25\-220 CE\) began integrating commentaries across different editions into a single base edition, sometimes supplemented by the compiler’s own interpretations\(Zhang,[2011](https://arxiv.org/html/2607.29044#bib.bib108)\), giving rise to the tradition of collected commentaries, orjijie\(集解, alsojizhu集注, orjishi集釋\), which was subsequently adopted to numerous Confucian and non\-Confucian classics\(Ying,[2005](https://arxiv.org/html/2607.29044#bib.bib104)\)\. Having last for nearly 2000 years, this tradition remains valuable for organizing the vast body of historical commentaries in today’s digital age\.

However, automatically compiling collected commentaries is far from straightforward due to the contextual exegesis \(隨文釋義\) practiced in inline notes, which means that inline notes can only be interpreted within their original context\. Successful compilation needs to preserve their contextual dependency after the detachment and reattachment, requiring consideration of the main\-text variants, commentary placement, and potential repetition or contradiction across editions\. Failure to do so may lead to inconsistency and imprecision, as observed in theShiji\(Ma,[2022](https://arxiv.org/html/2607.29044#bib.bib57)\)\.

In this paper, we formulate the compilation of collected commentary as an NLP task and propose a context\-preserving computational framework for its automatic compilation and exegetical knowledge organization\. Drawing on Chinese exegetics and philology, we derive a set of criteria for collected commentary compilation, among which identifying the main\-text segment associated with each commentary is particularly challenging\. Accordingly, we decompose the framework into two subtasks: \(1\) large language model \(LLM\)\-based identification of associated main\-text segments and exegetical functions; and \(2\) semantic similarity\-based cross\-source mention clustering for commentaries from different editions\. To implement this framework, we take theClassic of Mountains\(山經,Shan Jing, hereafterSJ\), one section of theClassic of Mountains and Seas\(山海經,Shanhai Jing, hereafterSHJ\)111We focus on theSJbecause of its intrinsic differences from the other parts of theSHJ\(Chen,[2012](https://arxiv.org/html/2607.29044#bib.bib11)\)and downstream\-task considerations\., as a case study\. Our framework achieves a CoNLL F1score above 97%\.

Our emphasis on contextual dependency reflects not only the intrinsic nature of traditional commentaries but also the context\-sensitive nature of language, aligning with the development of contextual representation in NLP\(Vaswaniet al\.,[2017](https://arxiv.org/html/2607.29044#bib.bib81); Peterset al\.,[2018](https://arxiv.org/html/2607.29044#bib.bib66); Devlinet al\.,[2019](https://arxiv.org/html/2607.29044#bib.bib16); Choiet al\.,[2021](https://arxiv.org/html/2607.29044#bib.bib13)\)\. The main contributions of this paper are twofold\. First, grounded in Chinese exegetics and philology, we formulate collected commentary compilation as an NLP task and propose a computational framework for its compilation and exegetical knowledge organization\. Second, our framework provides a general approach to organizing large\-scale historical exegetical knowledge from collected commentaries\. The resulting structured knowledge can serve as a foundation for a broad range of downstream philological and NLP tasks, facilitating both humanities research on classical Chinese and the adaptation of modern NLP methods to classical texts\.

## 2Related Work

Many NLP tasks have been applied to ancient texts, whereas ancient commentaries have received comparatively little attention\. In this section, we review previous work in two areas—NLP research on ancient classics and the computational processing of ancient commentaries, with a particular focus on Chinese—and position our work within this literature\.

### 2\.1NLP Research on Ancient Classics

Regarding NLP research on ancient Chinese classics, early studies mainly focused on tasks such as part\-of\-speech tagging \(POS\), Chinese word segmentation \(CWS\), named entity recognition \(NER\)\(Huanget al\.,[2010](https://arxiv.org/html/2607.29044#bib.bib33); Tang,[2013](https://arxiv.org/html/2607.29044#bib.bib76); Huanget al\.,[2015](https://arxiv.org/html/2607.29044#bib.bib36); Wanget al\.,[2016](https://arxiv.org/html/2607.29044#bib.bib88); Huang and Wu,[2018](https://arxiv.org/html/2607.29044#bib.bib35); Liet al\.,[2018](https://arxiv.org/html/2607.29044#bib.bib46)\), mainly based on rule\-based approaches or traditional machine learning methods\. More recently, Transformer architecture and subsequent LMs have substantially advanced research on these tasks\(Huet al\.,[2021](https://arxiv.org/html/2607.29044#bib.bib38); Tang and Su,[2022](https://arxiv.org/html/2607.29044#bib.bib77); Wanget al\.,[2022](https://arxiv.org/html/2607.29044#bib.bib84); Wang and Ren,[2022](https://arxiv.org/html/2607.29044#bib.bib90); Yasuoka,[2023](https://arxiv.org/html/2607.29044#bib.bib103); Zhuet al\.,[2024](https://arxiv.org/html/2607.29044#bib.bib114); Tanget al\.,[2024](https://arxiv.org/html/2607.29044#bib.bib73); Xuet al\.,[2024](https://arxiv.org/html/2607.29044#bib.bib98); Denget al\.,[2024](https://arxiv.org/html/2607.29044#bib.bib15); Kanget al\.,[2025](https://arxiv.org/html/2607.29044#bib.bib42); Tanget al\.,[2026](https://arxiv.org/html/2607.29044#bib.bib75)\)\. For example,Huet al\.\([2021](https://arxiv.org/html/2607.29044#bib.bib38)\)employed BERT\-CRF and BERT\-CNN models for CWS on ancient texts\.Tanget al\.\([2024](https://arxiv.org/html/2607.29044#bib.bib73)\)constructed a cross\-era dataset for developing and evaluating NER and relation extraction\.Denget al\.\([2024](https://arxiv.org/html/2607.29044#bib.bib15)\)integrated temporal features into PLMs for named entity disambiguation in historical documents\.

Besides these task\-oriented studies, considerable effort has also been devoted to constructing specialized linguistic resources, including a word sense annotation corpus\(Shuet al\.,[2022](https://arxiv.org/html/2607.29044#bib.bib70)\), ancient Chinese allusion resources\(Moet al\.,[2024](https://arxiv.org/html/2607.29044#bib.bib60)\), and a Tongjiazi \(通假字\) dataset\(Wanget al\.,[2024](https://arxiv.org/html/2607.29044#bib.bib83)\)\. These resources provide valuable training datasets for downstream NLP models\. Additionally, recent studies have further explored knowledge\-oriented applications, such as knowledge graph\(Lianget al\.,[2024](https://arxiv.org/html/2607.29044#bib.bib44)\), GraphRAG\(Yanget al\.,[2026](https://arxiv.org/html/2607.29044#bib.bib100)\), multimodal fusion classification model\(Shenet al\.,[2026](https://arxiv.org/html/2607.29044#bib.bib69)\), aiming to facilitate knowledge retrieval, reasoning, and intelligent question answering over classical texts\.

### 2\.2Commentary Processing

Regarding ancient commentaries processing, to our knowledge, computational studies of Chinese commentaries began withMaet al\.\([2012b](https://arxiv.org/html/2607.29044#bib.bib54),[a](https://arxiv.org/html/2607.29044#bib.bib55)\), who performed automatic analysis of commentaries on theAnalects\. Their workflow consisted of sentence alignment, which aligns the main text across different editions at sentence level, and annotation alignment, which identifies the main\-text segment associated with each annotation\(Chenet al\.,[2013](https://arxiv.org/html/2607.29044#bib.bib12)\)\. Similar approaches were subsequently applied to other Confucian classics, including theYijing\(Jia,[2015](https://arxiv.org/html/2607.29044#bib.bib41)\),Shijing\(Wang,[2018](https://arxiv.org/html/2607.29044#bib.bib89)\),Zuozhuan\(Xu and Liang,[2019](https://arxiv.org/html/2607.29044#bib.bib99)\), andMengzi\(Liang,[2021](https://arxiv.org/html/2607.29044#bib.bib45)\)\. More recently,Liet al\.\([2025](https://arxiv.org/html/2607.29044#bib.bib51)\)formulated annotation alignment as a semantic graph construction problem, using modularity\-maximization clustering to align annotations with their corresponding main\-text sentences within each individual edition\.

Comparable concerns also appear in Western digital humanities research on medieval fragments\(Bertiet al\.,[2014](https://arxiv.org/html/2607.29044#bib.bib3),[2016](https://arxiv.org/html/2607.29044#bib.bib2); Dobcheva and Mackert,[2018](https://arxiv.org/html/2607.29044#bib.bib18)\), biblical catenae\(Paparnakis and Domouchtsis,[2017](https://arxiv.org/html/2607.29044#bib.bib65)\), the Commentaria inAristotelem Graeca et Byzantina\([CAGB,](https://arxiv.org/html/2607.29044#bib.bib7)\), and the scholia of theVenetus Amanuscript\([HMT,](https://arxiv.org/html/2607.29044#bib.bib32)\)\. Although these studies primarily focus digitization, they likewise highlight the central role of annotations in preserving and interpreting ancient knowledge\.

### 2\.3Research Positioning

Existing studies on commentary processing have established several core research tasks, including automatic commentary analysis and annotation alignment, and have identified challenges such as main\-text variants\(Xu and Liang,[2019](https://arxiv.org/html/2607.29044#bib.bib99)\)and weak semantic similarity between targets and explanations\(Liet al\.,[2025](https://arxiv.org/html/2607.29044#bib.bib51)\)\. However, little attention has been paid to collected commentaries, the contextual dependency of inline notes, and the context\-preserving organization of exegetical knowledge\.

More importantly, most NLP studies reviewed above, from NER to knowledge graph construction, rely on manually curated datasets\(Huanget al\.,[2010](https://arxiv.org/html/2607.29044#bib.bib33),[2015](https://arxiv.org/html/2607.29044#bib.bib36); Liet al\.,[2018](https://arxiv.org/html/2607.29044#bib.bib46); Tang and Su,[2022](https://arxiv.org/html/2607.29044#bib.bib77); Zhuet al\.,[2024](https://arxiv.org/html/2607.29044#bib.bib114); Denget al\.,[2024](https://arxiv.org/html/2607.29044#bib.bib15); Kanget al\.,[2025](https://arxiv.org/html/2607.29044#bib.bib42); Lianget al\.,[2024](https://arxiv.org/html/2607.29044#bib.bib44)\), whose construction often requires expert knowledge derived directly or indirectly from commentaries\. Research has also demonstrated that external knowledge\(Nieet al\.,[2022](https://arxiv.org/html/2607.29044#bib.bib61)\), including commentarial information\(Xu and Chen,[2012](https://arxiv.org/html/2607.29044#bib.bib97)\)can facilitate NER and CWS\. Therefore, our work on collected commentary compilation and exegetical knowledge organization may provide a foundation for future development in these tasks\.

## 3Theoretical Foundations

This section outlines the theoretical foundations for the computational organization of historical exegetical knowledge, covering commentary hierarchy and placement, the contextual dependency of inline notes, exegetical functions, and integration criteria\. Because Chinese commentary systems vary across texts and require individual analysis\(Ni,[2004](https://arxiv.org/html/2607.29044#bib.bib62)\), we focuse on theSJwhile drawing on other classical texts to illustrate the broader applicability\.

Table[1](https://arxiv.org/html/2607.29044#S3.T1)lists the fiveSHJeditions examined in this study \(Editions A\-E\): Guo Pu’s, Wu Renchen’s, Wang Fu’s, Bi Yuan’s and Hao Yixing’s commentaries \(hereafter Guo, Wu, etc\.\)\. In all five editions, the commentaries are embedded within the main text in smaller double\-column script\. Figure[1](https://arxiv.org/html/2607.29044#S3.F1)shows a page excerpt on*xuncao*\(薰草\) from Edition D and part of its corresponding translation\. Table[2](https://arxiv.org/html/2607.29044#S3.T2)presents the digitized main text and selected commentaries from the “Qiuru Mountain” paragraph across the five editions\. Character variants in the main text are marked with red circles\. Each commentary is assigned a unique ID\. These examples are used throughout to illustrate the theoretical foundations\.

Table 1:Selected inline\-commentary versions on theSHJTable 2:Main text and commentaries for the “Qiuru Mountain” paragraph\.![Refer to caption](https://arxiv.org/html/2607.29044v1/Fig1-xuncao.png)Figure 1:The*Xuncao*passage of Edition D\.### 3\.1Commentary Hierarchy and Placement

The main genres of ancient Chinese commentaries includezhuanzhu\(傳注\),yishu\(義疏\), andzhangju\(章句\)\. Since the inline notes examined in this study belong tozhuanzhuandyishu, we briefly introduce these two genres below\.Zhuanzhuprimarily focuses on lexical items\.Yishuis a form of sub\-commentary whose explanatory target may lie in either the main text or an earlier commentary\. They differ in commentary hierarchies\.Zhuanzhucontains a single commentary layer, whileyishucontains two commentary layers, whose second layer explains either the main text or the first layer\. For clarity, we distinguish betweenbase\-commentary\(b\-commentary\), inherited from earlier commentators \(if present\), andcore\-commentary\(c\-commentary\), newly contributed layer in a given edition\. Additionally, although all five editions comment on the same classic, their commentary placement differs owing to differences in exegetical conventions among commentators\.

Among the five editions, Guo’s Edition A iszhuanzhuand the earliest extant commentary on theSHJ, whereas the other four areyishuthat all draw on Guo’s interpretations\. The first layer of Edition B, D and E consists of Guo’s annotations, separated from the second layer by formulaic expressions such as “任臣案”,“沅曰” and “懿行案” \(“Renchen/Yuan/Yixing comments”\)\. By contrast, Edition C integrates Guo’s annotations with Wang’s own interpretation into a single layer, due to the underdevelopment of evidential scholarship in Wang’s time\(Chen,[2012](https://arxiv.org/html/2607.29044#bib.bib11)\)\. Accordingly, Editions A and C contain only c\-commentary, while the other three contain both\.

The commentary placement in Edition C differs from the other four editions, placing commentaries at the end of complete sentences rather than after the associated main\-text segments\. This is evident from the placement of the commentary IDs in Table[2](https://arxiv.org/html/2607.29044#S3.T2)\. Wang’s placement strategy often cause one commentary to incorporate material from multiple entries in Guo’s commentary\.

### 3\.2Contextual Exegesis

The primary exegetical units of inline notes range from words and phrases to sentences and larger textual units\(Wang,[1996](https://arxiv.org/html/2607.29044#bib.bib92)\)\. In our study, most commentaries target words, although their interpretations often extend beyond them because of contextual exegesis\. The contextual dependency gives rise to two additional features that make identifying exegetical targets non\-trivial and motivate the design of our framework\.

First, different spans within a commentary may have different exegetical targets, both within and across commentary layers\. For example, AN2\.5 in Table[2](https://arxiv.org/html/2607.29044#S3.T2)first explains*bi*\(臂\) and then shifts to the referent of*shuima*\(水馬\)\. In EN2\.7, the c\-commentary explains*lou*\(膢\), a term introduced in its b\-commentary \(derived from AN2\.5\) rather than the main text\. This shows that interpreting c\-commentaries may require reference to the hierarchical context formed by the main text and the preceding b\-commentary\. Similar phenomena also occur inLunyu Jijie\([He,](https://arxiv.org/html/2607.29044#bib.bib31)\)and the Baina edition of theShiji\([Sima,](https://arxiv.org/html/2607.29044#bib.bib71)\)\.

Second, the text segment to which a commentary is attached is not always its exegetical target\. In AN2\.5, the commentary is attached to “文臂牛尾”, yet its latter part interprets*shuima*\. Similarly, in*xuncao*passage \(Figure[1](https://arxiv.org/html/2607.29044#S3.F1)\), commentary DW9\.7 is attached to “麻葉而方莖赤華而黑實” but ultimately targets*xuncao*\. Comparable examples also occur in theZhuangzi\([Zhuanget al\.,](https://arxiv.org/html/2607.29044#bib.bib113)\)\. We therefore define the attached text segment as theanchor\_text, distinguishing it from the actual exegetical target,

### 3\.3Exegetical Functions

Although inline notes must be interpreted in context, traditional Chinese philology also produced works such as theEryaandJingji Zuangu, which reorganize exegetical knowledge extracted from diverse classical sources\(Wang,[1996](https://arxiv.org/html/2607.29044#bib.bib92)\)\. Inspired by this tradition, we classify and, when necessary, decompose commentaries by their primary exegetical functions to facilitate downstream applications\.

We propose a threefold classification: textual criticism \(TC\), lexical exegesis \(LE\), and referential exegesis \(RE\)\.TCaddresses textual variants and textual emendation,LEexplains the pronunciation, orthography, and meanings of words; andREinvestigates named entities, including their attributes, identities, and historical transformations\. Compared with earlier schemes\(Han,[1989](https://arxiv.org/html/2607.29044#bib.bib30); Huang,[1999](https://arxiv.org/html/2607.29044#bib.bib37)\), this framework is simpler and more applicable to downstream philological tasks\. Notably, because referential exegesis historically developed from lexical exegesis, the boundary between the two is sometimes blurred\. In this study, any commentary involving the external referent of a named entity is classified asRE\.

### 3\.4Commentary Integration Criteria

Although collected commentaries have been widely discussed in Chinese exegetics and philology\(Huang,[2001](https://arxiv.org/html/2607.29044#bib.bib34); Guo,[2005](https://arxiv.org/html/2607.29044#bib.bib28); Ying,[2005](https://arxiv.org/html/2607.29044#bib.bib104); Zhang,[2011](https://arxiv.org/html/2607.29044#bib.bib108); Xu,[2014](https://arxiv.org/html/2607.29044#bib.bib96); Guoet al\.,[2024](https://arxiv.org/html/2607.29044#bib.bib29)\), little attention has been paid to the criteria for commentary integration\.Feng \([2003](https://arxiv.org/html/2607.29044#bib.bib25)\)argues that commentaries collected under ajijieentry should represent alternative interpretations of the same main\-text segment, whileMa \([2022](https://arxiv.org/html/2607.29044#bib.bib57)\)notes that integration also requires resolving differences in commentary placement as well as redundancy and contradiction across versions\. The compilation of works such asShiji\([Sima,](https://arxiv.org/html/2607.29044#bib.bib71)\)andLunyu Jijie\([He,](https://arxiv.org/html/2607.29044#bib.bib31)\)likewise suggest that commentaries are generally integrated by a shared textual segment\. However, as discussed above, the contextual dependency of inline notes makes such segments difficult to identify\. In practice, editorial convenience sometimes leads compilers to group commentaries on similar rather than identical segments\. For example,Shanhaijing Jishiplaces Wang’s annotation on “南山經之首曰” together with Bi’s annotation on “南山經之首”\(Zhou,[2019](https://arxiv.org/html/2607.29044#bib.bib112)\), sacrificing precision\.

Operationally, our analysis of theSJcommentaries and other representativejijieworks suggests that commentary integration is primarily determined by three factors: \(1\) shared main\-text segments; \(2\) commentary placement; and \(3\) semantic similarity among commentaries\. As inline notes usually follow the passages they explain, commentary placement is largely determined by the main text, although editorial conventions also affect placement \(e\.g\., the different placement preferences between Guo and Wang\)\. Semantic similarity is less informative for some commentaries because of their diverse content, but serves as a useful supplementary criterion foryishuthat quote or build upon earlier annotations\.

## 4Computational Framework

Based on the above analysis, our framework adopts shared main\-text segments and semantic similarity between first\-layer commentaries as the primary integration criteria, using commentary placement only during proofreading\. Semantic similarity is included because Edition B\-E areyishuand have absorbed Edition A\. Accordingly, the task is decomposed into two subtasks: \(1\) identifying the main\-text segment associated with each commentary, and \(2\) grouping commentaries with the same main\-text segment or highly similar first\-layer commentary, based on semantic similarity\. We address the first using LLMs and formulate the second as a cross\-source mention clustering problem\. To support downstream applications, we also identify each commentary’s exegetical function in the first subtask\.

### 4\.1Two\-step Prompt Chaining

Because a commentary may involve multiple exegetical targets and itsanchor\_textis not always the ultimate target, we adopt a two\-step prompt chaining strategy to identify the associated main\-text segments\. Prompt chaining has been shown to outperform a single prompt on complex tasks\(Sunet al\.,[2024](https://arxiv.org/html/2607.29044#bib.bib72)\)\. Figure[2](https://arxiv.org/html/2607.29044#S4.F2)illustrates the pipeline\.

![Refer to caption](https://arxiv.org/html/2607.29044v1/Fig2-prompt.png)Figure 2:Two\-step prompt chaining pipeline\.Each paragraph of each edition is processed independently\. After paragraph IDs are assigned, the main text and commentaries are separated and converted into JSON as the pipeline input\. Step 1 \(Anchoring\) identifies each commentary’sanchor\_textand correspondingentry\_unit\(primarily named entities\)\. Step 2 \(Refining\) further determines the actual exegetical target and classifies or, when necessary, decomposes each commentary according to the threefold exegetical classification\. For post hoc verification, auxiliary fields \(e\.g\.,context\_leftandcontext\_right\) are retained in the output\. Commentaries with two layers are separated after Step 1 so that Step 2 operates only on the c\-commentary\. The example at the bottom of Figure[2](https://arxiv.org/html/2607.29044#S4.F2)illustrates the two challenges addressed by the framework: multiple exegetical targets within a single commentary \(DW9\.6\) and the mismatch between theanchor\_textand the actual exegetical target \(DW9\.7\)\.

Because the five editions differ in commentary hierarchy and placement, we designed two prompt variants \(a and b\)\. Variant b adds additional fields \(marked with asterisks in Figure[2](https://arxiv.org/html/2607.29044#S4.F2)\) for edition\-specific features, including commentary splitting and b\-commentary\-dependent interpretation\.

### 4\.2Cross\-source Mention Clustering

Clustering is based on the semantic similarity of either the associated main\-text segment or the first\-layer commentary\. The associated main\-text segment may be theanchor\_text,entry\_unitor target identified in Subtask 1\. Partly due to the high density of named entities, commentaries like DW9\.7 that are attached to one segment but target another are very common inSJ\. We therefore use theanchor\_textfor clustering\. Using theentry\_unitor target alone may merge commentaries with different exegetical focuses or decomposed commentaries\.

Algorithm[1](https://arxiv.org/html/2607.29044#alg1)presents the cross\-source mention clustering procedure\. Commentaries from the sameSJparagraph across the five editions are processed as a single batch\. For each commentary, the top\-k candidates are retrieved according to the semantic similarity ofanchor\_textcontextual embeddings or first\-layer commentary embeddings\. Clustering is then performed using a Union\-Find algorithm with three constraints: \(1\) cluster\-size limits, \(2\) complete linkage for commentary\-similarity merging, and \(3\) source uniqueness\. Similar strategies have been adopted in record linkage\(Guoet al\.,[2010](https://arxiv.org/html/2607.29044#bib.bib27)\), entity resolution\(Bhattacharya and Getoor,[2007](https://arxiv.org/html/2607.29044#bib.bib4); Kardeset al\.,[2013](https://arxiv.org/html/2607.29044#bib.bib43)\), and mention clustering\(Zhanget al\.,[2018](https://arxiv.org/html/2607.29044#bib.bib109)\)\. In our framework, these constraints primarily prevent bridging effects\.

Algorithm 1Two\-Route Union\-Find Clustering\.1:Initialize each row as a singleton cluster\.

2:Phase 1: Commentary Merge

3:foreach candidate pair

\(i,j\)\(i,j\)do

4:

Ci←cluster​\(i\)C\_\{i\}\\leftarrow\\mathrm\{cluster\}\(i\),

Cj←cluster​\(j\)C\_\{j\}\\leftarrow\\mathrm\{cluster\}\(j\)
5:if

Ci≠CjC\_\{i\}\\neq C\_\{j\}and

comm​\_​sim​\(i,j\)≥Tc\\mathrm\{comm\\\_sim\}\(i,j\)\\geq T\_\{c\}then

6:if

∀x∈Ci,∀y∈Cj:comm​\_​sim​\(x,y\)≥Tc\\forall x\\in C\_\{i\},\\forall y\\in C\_\{j\}:\\mathrm\{comm\\\_sim\}\(x,y\)\\geq T\_\{c\}and

source​\(Ci\)∩source​\(Cj\)=∅\\mathrm\{source\}\(C\_\{i\}\)\\cap\\mathrm\{source\}\(C\_\{j\}\)=\\emptysetthen

7:merge

\(Ci,Cj\)\(C\_\{i\},C\_\{j\}\)
8:endif

9:endif

10:endfor

11:Phase 2: Anchor\-text Merge

12:foreach candidate pair

\(i,j\)\(i,j\)do

13:

Ci←cluster​\(i\)C\_\{i\}\\leftarrow\\mathrm\{cluster\}\(i\),

Cj←cluster​\(j\)C\_\{j\}\\leftarrow\\mathrm\{cluster\}\(j\)
14:if

Ci≠CjC\_\{i\}\\neq C\_\{j\}and

anchor​\_​sim​\(i,j\)≥Tt\\mathrm\{anchor\\\_sim\}\(i,j\)\\geq T\_\{t\}then

15:if

\(source​\(Ci\)∩source​\(Cj\)\)⊆\{C\}\(\\mathrm\{source\}\(C\_\{i\}\)\\cap\\mathrm\{source\}\(C\_\{j\}\)\)\\subseteq\\\{C\\\}then

16:merge

\(Ci,Cj\)\(C\_\{i\},C\_\{j\}\)
17:endif

18:endif

19:endfor

20:Assign each row the ID of its final cluster\.

21:Output clustered rows withgroup\_id\.

## 5Experiments

Since all five editions annotate theSJ, we compare them at multiple levels using character counts and character\-frequency distributions to characterize their similarity and inform experiments\. The main texts and Guo’s commentaries exhibit over 96% similarity in character\-frequency distributions\. Similarity decreases as later commentaries are incorporated, but remains above 75%\. This high similarity makes model selection and clustering thresholds particularly important\.

### 5\.1Model Selection

We evaluate three LLMs222DeepSeek\-Chat, Kimi\-K2\-0905\-Preview, and GPT\-5\-mini\.for Subtask 1 and seven Classical Chinese BERT models333SikuBERT, SikuRoBERTa\(Wanget al\.,[2022](https://arxiv.org/html/2607.29044#bib.bib84)\), BERT\-Ancient\-Chinese\(Wang and Ren,[2022](https://arxiv.org/html/2607.29044#bib.bib90)\), RoBERTa\-Classical\-Chinese\-Base\-Char\(Yasuoka,[2022a](https://arxiv.org/html/2607.29044#bib.bib101)\), RoBERTa\-Classical\-Chinese\-Large\-Char\(Yasuoka,[2022b](https://arxiv.org/html/2607.29044#bib.bib102)\), GuwenBERT\-Base\(Ethanyt,[2020a](https://arxiv.org/html/2607.29044#bib.bib22)\), GuwenBERT\-Large\(Ethanyt,[2020b](https://arxiv.org/html/2607.29044#bib.bib23)\)\.for Subtask 2 on manually curated datasets\. Based on their performance, we select DeepSeek and SikuRoBERTa, respectively\. For SikuRoBERTa, we average the last five hidden layers foranchor\_textcontextual embeddings, the last two for commentary embeddings, and use the correspondingpos\_p05andpos\_meanvalues as clustering thresholds\.

### 5\.2Evaluation

The extraction and clustering results for theClassic of Northern Mountainssection \(18\.044% of the dataset\) were manually verified and used as the gold standard for evaluation\. For Subtask 1, we report the accuracy of each step , their weighted average \(E\-E Acc\.\), and the mean Jaccard score for Step 2, which evaluates the identification of multiple exegetical functions within a single commentary\. Table[3](https://arxiv.org/html/2607.29044#S5.T3)summarizes the results\. The overall accuracy is 84%, with errors mainly caused by the inaccurateanchor\_textextraction and incorrect decomposition of exegetical functions\.

S1 Acc\.S2 Acc\.S2 Jacc\.E2E Acc\.0\.8270\.8530\.8480\.840Table 3:Extraction performance of Subtask 1\.We evaluate clustering performance using standard coreference metrics: MUC\(Vilainet al\.,[1995](https://arxiv.org/html/2607.29044#bib.bib82)\),B3B^\{3\}\(BAGGA,[1998](https://arxiv.org/html/2607.29044#bib.bib1)\), and CEAFϕ4\{\}\_\{\\phi\_\{4\}\}\(Luo,[2005](https://arxiv.org/html/2607.29044#bib.bib53)\), and report the CoNLL score\(Pradhanet al\.,[2011](https://arxiv.org/html/2607.29044#bib.bib67),[2012](https://arxiv.org/html/2607.29044#bib.bib68)\)as the their average F1\. Pairwise scores are reported to evaluate pairwise linking\. Table[4](https://arxiv.org/html/2607.29044#S5.T4)shows that precision and recall exceed 95% across all metrics, and the CoNLL F1exceeds 97%\. Errors are mainly caused by the textual variants inancho\_textand Wang’s commentaries\.

Table 4:Clustering performance of Subtask 2\.
### 5\.3Results

Table[5](https://arxiv.org/html/2607.29044#S5.T5)summarizes the extraction and clustering statistics\. More than 78% ofSJparagraphs contain commentaries\. The numbers ofComm\.andC\-comm\.reflect the accumulation of exegetical knowledge over time, while the distributions ofTC,LE, andREfurther reveal a shift in exegetical practice\. Guo’s commentary is dominated byLE, whereas Qing\-dynasty commentaries contain fewer pureLEcases and many moreREorTC, likely refelcting the rise of evidential scholarship, in whichLEincreasingly servedREandTCrather than functioning independently\. TheSplitrate indicates that each c\-commentary contains an average of 1\.16\-1\.26 exegetical functions\. Finally, the c\-commentaries \(5,463 in total\) are grouped into 2,440 clusters, averaging 2\.24 commentaries per semantic anchor\.

Table 5:Extraction and clustering statistics\.*Note:*Ann\.denotes the percentage ofSJparagraphs with commentaries\.Comm\.andC\-comm\.denote the numbers of commentaries and c\-commentaries\. For Editions A and C, they would theoretically be identical, butC\-comm\.is larger in Edition C because some commentaries are split during Step 1 of prompt chaining\.TC,LE, andREreport the distribution of the three exegetical functions among c\-commentaries\.Splitrate = \(TC\+LE\+RE\) /C\-comm\., indicating the average diversity of exegetical functions per c\-commentary\.

## 6Implications

The implications of this study extend beyond theSJ\. Figure[3](https://arxiv.org/html/2607.29044#S6.F3)illustrates the experimental results for the “Qiuru Mountain” paragraph in Table[2](https://arxiv.org/html/2607.29044#S3.T2)\. The left panel shows commentary integration results444Commentaries consisting solely of Guo’s commentary \(without a c\-commentary\) or repeated occurrences of Guo’s commentary are marked in light gray\., the middle panel the threefold classification of exegetical functions555Rectangles represent commentaries and circles represent their corresponding exegetical targets; green, blue, and yellow denote TC, LE, and RE, respectively\., and the right panel potential philological downstream applications\. In the main text above, commentary\-group markers \(comm\_group marker\) indicate both commentary placement and clustering assignments\. Using this example, we discuss the broader implications of our framework from four perspectives: context\-preserving commentary integration, exegetical function classification for philological downstream tasks, edition genealogy through variant character analysis, and potential NLP applications\.

![Refer to caption](https://arxiv.org/html/2607.29044v1/Fig3-implications.png)Figure 3:Experimental results from the “Mountain Qiuru” paragraph and their possible applications\.### 6\.1Context\-Preserving Integration

Our framework groups commentaries that share the sameanchor\_textor have similar first\-layer annotations and orders them chronologically by edition\. Even after duplicate b\-commentaries are removed, the interpretive context formed by themain text \- b\-commentary \- c\-commentaryhierarchy is preserved\. For example, all commentaries in group \[2\.9\] of Figure[3](https://arxiv.org/html/2607.29044#S6.F3)were originally attached to the same segment \(see Table[2](https://arxiv.org/html/2607.29044#S3.T2)\), and the c\-commentary of EN2\.7 remains fully interpretable through AN2\.5\. Cases that do not satisfy these criteria are resolved through manual philological verification\.

### 6\.2Philological Downstream Tasks

The proposed threefold classification of exegetical functions, combined with data linking, supports downstream tasks such as textual collation and variant collection, lexicography and lost\-text reconstruction, and historical encyclopedia compilation and historical entity identification\.

TCcommentaries support textual collation and variant collection by recording textual variants and editorial corrections\. For example, DN2\.1 and EN2\.1 identify*huashui*\(滑水\) as a variant of*xushui*\(湑水\), while EN2\.6 relates 疣 to 肬 in theShuowen Jiezi\. Such information is essential for reconstructing reliable texts, as demonstrated in recent studies on theShuijingzhu\(Liet al\.,[2021](https://arxiv.org/html/2607.29044#bib.bib50)\)\.

LEcommentaries support lexicography and lost\-text reconstruction by explaining word meanings and pronunciations while preserving quotations from other philological works\. For example, AN2\.4 and CN2\.2 explain the meaning and pronunciation of 疣, offering evidence for the diachronic development of Classical Chinese vocabulary\. BN2\.4 cites the lostZishu\. Collecting such quotations at scale facilitates the reconstruction of this work, as illustrated byWang \([2016](https://arxiv.org/html/2607.29044#bib.bib93)\)\. Similar quotation\-based approaches have also been used to medieval fragments\(Bertiet al\.,[2014](https://arxiv.org/html/2607.29044#bib.bib3)\)\.

REcommentaries support historical encyclopedia compilation and historical entity identification by preserving investigations of historical entities and institutions\. Examples include CN2\.1 on the relationship between*Qiuru Mountain*and*Buzhou Mountain*, and CN2\.3 on the identity of*shuima*and the geography of*huashui*\. Such evidence can be integrated with other classical sources for large\-scale historical knowledge reconstruction\. Comparable uses of commentarial evidence have been reported for named\-entity disambiguation in theZuozhuan\(Liet al\.,[2020](https://arxiv.org/html/2607.29044#bib.bib48)\)and LLM\-based geocoding of Early China toponyms\(Chenet al\.,[2026](https://arxiv.org/html/2607.29044#bib.bib9)\)\.

### 6\.3Edition Genealogy Analysis

The organization of traditional commentaries often requires careful comparison across textual editions and commentary layers\(Ni,[2004](https://arxiv.org/html/2607.29044#bib.bib62)\)\. As Table[2](https://arxiv.org/html/2607.29044#S3.T2)shows, the five editions exhibit a number of character variants in the main text\. Based on the character frequency analysis in Section 5, we further cluster the five editions using the 20 most frequent variant\-character pairs as features\. The results suggest that variant\-character patterns can serve as a quantitative indicator of edition genealogy\. The resulting dendrogram in Figure[4](https://arxiv.org/html/2607.29044#S6.F4)closely agrees with the philological analysis ofZhang \([2007](https://arxiv.org/html/2607.29044#bib.bib110)\): Editions D and E derive from the Ming\-dynasty woodblock edition collated by Xiangyin \(項絪\), Editions B and C represent two other branches of the Ming woodblock tradition, and Edition A was reproduced from a Ming imperial edition\.

![Refer to caption](https://arxiv.org/html/2607.29044v1/Fig4-Genealogy.png)Figure 4:Edition clustering by variant pairs\.
### 6\.4Potential NLP Applications

As discussed above, the organized exegetical knowledge produced by our framework can provide foundational knowledge for many NLP tasks, such as CWS and NER\. It also enables the adaptation of other NLP tasks, such as text summarization and textual entailment, to collected commentaries, supporting applications including commentary summarization and cross\-commentator consistency analysis\.

## 7Conclusion

This paper formulates collected commentary compilation as an NLP task and proposes a computational framework for preserving the contextual dependency of inline notes and organizing historical exegetical knowledge\. The main limitation of this work is that our experiments are limited to theSJ, although the proposed framework is generalizable to other ancient texts\. Future work includes extending the framework to a broader range of ancient texts, investigating its contribution to related NLP tasks, and reconstructing the natural\-historical landscape of theSJbased on the resulting structured knowledge\.

## References

- Algorithms for scoring coreference chains\.Proc\. Linguistic Coreference Workshop at the first Conf\. on Language Resources and Evaluation \(LREC\), Granada, Spain, May 1998\.Cited by:[§5\.2](https://arxiv.org/html/2607.29044#S5.SS2.p2.4)\.
- M\. Berti, B\. Almas, and G\. R\. Crane \(2016\)The Leipzig Open Fragmentary Texts Series \(LOFTS\)\.\.DHQ: Digital Humanities Quarterly10\(2\),pp\. 1\.External Links:ISSN 1938\-4122Cited by:[§2\.2](https://arxiv.org/html/2607.29044#S2.SS2.p2.1)\.
- M\. Berti, B\. Almas, D\. Dubin, G\. Franzini, S\. Stoyanova, and G\. R\. Crane \(2014\)The Linked Fragment: TEI and the Encoding of Text Reuses of Lost Authors\.Journal of the Text Encoding Initiative\(Issue 8\)\.External Links:ISSN 2162\-5603,[Document](https://dx.doi.org/10.4000/jtei.1218)Cited by:[§2\.2](https://arxiv.org/html/2607.29044#S2.SS2.p2.1),[§6\.2](https://arxiv.org/html/2607.29044#S6.SS2.p3.1)\.
- I\. Bhattacharya and L\. Getoor \(2007\)Collective entity resolution in relational data\.ACM Transactions on Knowledge Discovery from Data \(TKDD\)1\(1\),pp\. 5–es\.External Links:ISSN 1556\-4681,[Document](https://dx.doi.org/10.1145/1217299.1217304)Cited by:[§4\.2](https://arxiv.org/html/2607.29044#S4.SS2.p2.1)\.
- \[5\]CAGBCommentaria in Aristotelem Graeca et Byzantina \(CAGB\)\.Note:https://digitalhumanities\.de/en/projekt/cagb/Cited by:[§2\.2](https://arxiv.org/html/2607.29044#S2.SS2.p2.1)\.
- L\. Chen \(2012\)Studies in the Scholarly History of the Classic of Mountains and Seas \(《山海經》學術史考論\)\.Peking University Press,Beijing\.External Links:ISBN 978\-7\-301\-20302\-6Cited by:[§3\.1](https://arxiv.org/html/2607.29044#S3.SS1.p2.1),[footnote 1](https://arxiv.org/html/2607.29044#footnote1)\.
- X\. Chen, M\. Xu, and R\. Xu \(2013\)Information Processing for Pre\-Qin Textual Sources \(先秦文獻信息處理\)\.World Publishing Corporation,Beijing\.External Links:ISBN 978\-7\-5100\-5661\-1Cited by:[§2\.2](https://arxiv.org/html/2607.29044#S2.SS2.p1.1)\.
- Y\. Chen, W\. Shang, H\. Wang, S\. Zhang, and P\. K\. Bol \(2026\)Geocoding the past world: unearthing coordinates of early China from texts using generative AI\.International Journal of Geographical Information Science40\(3\),pp\. 781–807\.External Links:ISSN 1365\-8816,[Document](https://dx.doi.org/10.1080/13658816.2025.2491711)Cited by:[§6\.2](https://arxiv.org/html/2607.29044#S6.SS2.p4.1)\.
- E\. Choi, J\. Palomaki, M\. Lamm, T\. Kwiatkowski, D\. Das, and M\. Collins \(2021\)Decontextualization: Making Sentences Stand\-Alone\.Transactions of the Association for Computational Linguistics9,pp\. 447–461\.External Links:ISSN 2307\-387X,[Document](https://dx.doi.org/10.1162/tacl%5Fa%5F00377)Cited by:[§1](https://arxiv.org/html/2607.29044#S1.p5.1)\.
- Z\. Deng, H\. Yang, and J\. Wang \(2024\)Diachronic Named Entity Disambiguation for Ancient Chinese Historical Records\.InNeural Information Processing,B\. Luo, L\. Cheng, Z\. Wu, H\. Li, and C\. Li \(Eds\.\),Singapore,pp\. 305–319\.External Links:[Document](https://dx.doi.org/10.1007/978-981-99-8145-8%5F24),ISBN 978\-981\-99\-8145\-8Cited by:[§2\.1](https://arxiv.org/html/2607.29044#S2.SS1.p1.1),[§2\.3](https://arxiv.org/html/2607.29044#S2.SS3.p2.1)\.
- J\. Devlin, M\. Chang, K\. Lee, and K\. Toutanova \(2019\)BERT: Pre\-training of Deep Bidirectional Transformers for Language Understanding\.InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 \(Long and Short Papers\),J\. Burstein, C\. Doran, and T\. Solorio \(Eds\.\),Minneapolis, Minnesota,pp\. 4171–4186\.External Links:[Document](https://dx.doi.org/10.18653/v1/N19-1423)Cited by:[§1](https://arxiv.org/html/2607.29044#S1.p5.1)\.
- I\. Dobcheva and C\. Mackert \(2018\)Manuscript Fragments in the University Library, Leipzig: Types and Cataloguing Patterns\.Fragmentology1,pp\. 83–110\.External Links:ISSN 2624\-9340,[Document](https://dx.doi.org/10.24446/rx89)Cited by:[§2\.2](https://arxiv.org/html/2607.29044#S2.SS2.p2.1)\.
- H\. Dong \(1993\)On the Similarities and Differences between Annotation and Chinese Exegetics \(註釋與訓詁異同辨\)\.Chinese Classics & Culture\(1\),pp\. 89–93\.External Links:ISSN 1004\-3241,[Document](https://dx.doi.org/10.16093/j.cnki.ccc.1993.01.021)Cited by:[§1](https://arxiv.org/html/2607.29044#S1.p1.1)\.
- Ethanyt \(2020a\)Guwenbert Base\.Cited by:[footnote 3](https://arxiv.org/html/2607.29044#footnote3)\.
- Ethanyt \(2020b\)Guwenbert Large\.Cited by:[footnote 3](https://arxiv.org/html/2607.29044#footnote3)\.
- H\. Feng \(2003\)A Study to the Style and Form of Annotated and Collated Works of Chinese Ancient Books （中國古籍整理體式研究）\.Higher Education Press,Beijing\.External Links:ISBN 978\-7\-04\-012277\-0Cited by:[§3\.4](https://arxiv.org/html/2607.29044#S3.SS4.p1.1)\.
- D\. K\. Gardner \(1998\)Confucian Commentary and Chinese Intellectual History\.The Journal of Asian Studies57\(2\),pp\. 397–422\.External Links:ISSN 1752\-0401, 0021\-9118,[Document](https://dx.doi.org/10.2307/2658830)Cited by:[§1](https://arxiv.org/html/2607.29044#S1.p2.1)\.
- S\. Guo, X\. L\. Dong, D\. Srivastava, and R\. Zajac \(2010\)Record linkage with uniqueness constraints and erroneous values\.Proceedings of the VLDB Endowment3\(1\-2\),pp\. 417–428\.External Links:ISSN 2150\-8097,[Document](https://dx.doi.org/10.14778/1920841.1920897)Cited by:[§4\.2](https://arxiv.org/html/2607.29044#S4.SS2.p2.1)\.
- Y\. Guo, X\. Yu, and X\. Li \(2024\)Theories and Methods of Chinese Classical Philology \(2nd ed\.\) \(中國古典文獻學的理論與方法\)\.Beijing Normal University Press,Beijing\.External Links:ISBN 978\-7\-303\-24183\-5Cited by:[§3\.4](https://arxiv.org/html/2607.29044#S3.SS4.p1.1)\.
- Z\. Guo \(2005\)Chinese Exegetics \(訓詁學\)\.Zhonghua shu ju\.External Links:ISBN 978\-7\-101\-04796\-7Cited by:[§3\.4](https://arxiv.org/html/2607.29044#S3.SS4.p1.1)\.
- G\. Han \(1989\)Can Chinese Exegetics Evolve into a Discipline of Annotation for Ancient Chinese Texts?\.Journal of Ancient Books Collation\(5\),pp\. 42–46, 97\.External Links:ISSN 1009\-1017Cited by:[§3\.3](https://arxiv.org/html/2607.29044#S3.SS3.p2.1)\.
- \[22\]Y\. HeCollected Explanations to the Analects \(Lunyu jijie, 論語集解\)\.Sibu Congkan edition edition\.Cited by:[§3\.2](https://arxiv.org/html/2607.29044#S3.SS2.p2.1),[§3\.4](https://arxiv.org/html/2607.29044#S3.SS4.p1.1)\.
- \[23\]HMTThe Homer Multitext project\.Note:https://www\.homermultitext\.org/Cited by:[§2\.2](https://arxiv.org/html/2607.29044#S2.SS2.p2.1)\.
- R\. Hu, S\. Li, and Y\. Zhu \(2021\)Knowledge Representation and Sentence Segmentation of Ancient Chinese Based on Deep Language Models\.Journal of Chinese Information Processing35\(4\),pp\. 8–15\.External Links:ISSN 1003\-0077Cited by:[§2\.1](https://arxiv.org/html/2607.29044#S2.SS1.p1.1)\.
- H\. Huang, C\. Sun, and H\. Chen \(2010\)Classical Chinese Sentence Segmentation\.InCIPS\-SIGHAN Joint Conference on Chinese Language Processing,Cited by:[§2\.1](https://arxiv.org/html/2607.29044#S2.SS1.p1.1),[§2\.3](https://arxiv.org/html/2607.29044#S2.SS3.p2.1)\.
- S\. Huang and J\. Wu \(2018\)A Pragmatic Approach for Classical Chinese Word Segmentation\.InProceedings of the Eleventh International Conference on Language Resources and Evaluation \(LREC 2018\),N\. Calzolari, K\. Choukri, C\. Cieri, T\. Declerck, S\. Goggi, K\. Hasida, H\. Isahara, B\. Maegaard, J\. Mariani, H\. Mazo, A\. Moreno, J\. Odijk, S\. Piperidis, and T\. Tokunaga \(Eds\.\),Miyazaki, Japan\.Cited by:[§2\.1](https://arxiv.org/html/2607.29044#S2.SS1.p1.1)\.
- S\. Huang, D\. Wang, and L\. He \(2015\)Research on Constructing Automatic Recognition Model for Ancient Chinese Place Names Based on Pre\-Qin Corpus\.Library and Information Service59\(12\),pp\. 135–140\.External Links:ISSN 0252\-3116,[Document](https://dx.doi.org/10.13266/j.issn.0252-3116.2015.12.020)Cited by:[§2\.1](https://arxiv.org/html/2607.29044#S2.SS1.p1.1),[§2\.3](https://arxiv.org/html/2607.29044#S2.SS3.p2.1)\.
- Y\. Huang \(1999\)A Preliminary Discussion of the Types of Annotations in Ancient Chinese Texts \(古籍註釋類型芻議\)\.Journal of the Northwest Normal University\(Social Sciences\)\(3\),pp\. 91–94\.External Links:ISSN 1001\-9162,[Document](https://dx.doi.org/10.16783/j.cnki.nwnus.1999.03.020)Cited by:[§3\.3](https://arxiv.org/html/2607.29044#S3.SS3.p2.1)\.
- Y\. Huang \(2001\)An Introduction to the Collation of Ancient Chinese Texts \(古籍整理概論\)\.Shanghai Bookstore Publishing House,Shanghai\.External Links:ISBN 978\-7\-80622\-464\-9Cited by:[§3\.4](https://arxiv.org/html/2607.29044#S3.SS4.p1.1)\.
- F\. Jia \(2015\)A Knowledge\-Based Clustering Method for Constructing Knowledge Bases of Ancient Chinese Texts\.Journal of Library Science37\(5\),pp\. 45–48\.External Links:ISSN 1002\-1884,[Document](https://dx.doi.org/10.14037/j.cnki.tsgxk.2015.05.015)Cited by:[§2\.2](https://arxiv.org/html/2607.29044#S2.SS2.p1.1)\.
- W\. Kang, J\. Zuo, Q\. Dai, Y\. Hu, and M\. Wang \(2025\)Research on Named Entity Recognition in Ancient Chinese Based on Incremental Pre\-training and Domain Lexicon\.InNatural Language Processing and Chinese Computing,D\. F\. Wong, Z\. Wei, and M\. Yang \(Eds\.\),Singapore,pp\. 483–503\.External Links:[Document](https://dx.doi.org/10.1007/978-981-97-9431-7%5F37),ISBN 978\-981\-97\-9431\-7Cited by:[§2\.1](https://arxiv.org/html/2607.29044#S2.SS1.p1.1),[§2\.3](https://arxiv.org/html/2607.29044#S2.SS3.p2.1)\.
- H\. Kardes, D\. Konidena, S\. Agrawal, M\. Huff, and A\. Sun \(2013\)Graph\-based Approaches for Organization Entity Resolution in MapReduce\.InProceedings of TextGraphs\-8 Graph\-based Methods for Natural Language Processing,Z\. Kozareva, I\. Matveeva, G\. Melli, and V\. Nastase \(Eds\.\),Seattle, Washington, USA,pp\. 70–78\.Cited by:[§4\.2](https://arxiv.org/html/2607.29044#S4.SS2.p2.1)\.
- B\. Li, L\. Wang, X\. Chen, and D\. Wang \(2020\)Digital Humanity Based Ancient Text Annotation and Visualization——A Case Study on Zuozhuan Knowledgebase\.Journal of Academic Libraries38\(5\),pp\. 72–80, 90\.External Links:ISSN 1002\-1027,[Document](https://dx.doi.org/10.16603/j.issn1002-1027.2020.05.011)Cited by:[§6\.2](https://arxiv.org/html/2607.29044#S6.SS2.p4.1)\.
- S\. Li, M\. Li, Y\. Xu, Z\. Bao, L\. Fu, and Y\. Zhu \(2018\)Capsules Based Chinese Word Segmentation for Ancient Chinese Medical Books\.IEEE Access6,pp\. 70874–70883\.External Links:ISSN 2169\-3536,[Document](https://dx.doi.org/10.1109/ACCESS.2018.2881280)Cited by:[§2\.1](https://arxiv.org/html/2607.29044#S2.SS1.p1.1),[§2\.3](https://arxiv.org/html/2607.29044#S2.SS3.p2.1)\.
- W\. Li, Y\. Li, Y\. Shao, and M\. Bi \(2025\)A Structure\-aware Approach Leveraging Semantic Relevance Graph for Annotation Alignment of Chinese Classics\.ACM Transactions on Asian and Low\-Resource Language Information Processing24\(5\),pp\. 49:1–49:21\.External Links:ISSN 2375\-4699,[Document](https://dx.doi.org/10.1145/3725733)Cited by:[§2\.2](https://arxiv.org/html/2607.29044#S2.SS2.p1.1),[§2\.3](https://arxiv.org/html/2607.29044#S2.SS3.p1.1)\.
- X\. Li, C\. Yang, Y\. Wang, and K\. Qu \(2021\)Original and Revised Books: A Comprehensive Study on the Existing Editions of Notes to the Book of Rivers \(古本與今本：現存⟨水經注⟩版本彙考\)\.Fudan University Press,Shanghai\.External Links:ISBN 978\-7\-309\-15866\-3Cited by:[§6\.2](https://arxiv.org/html/2607.29044#S6.SS2.p2.1)\.
- K\. Liang, C\. Huang, and X\. Jiang \(2024\)From Text to Historical Ecological Knowledge: The Construction and Application of the Shan Jing Knowledge Base\.InProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation \(LREC\-COLING 2024\),N\. Calzolari, M\. Kan, V\. Hoste, A\. Lenci, S\. Sakti, and N\. Xue \(Eds\.\),Torino, Italia,pp\. 7521–7530\.Cited by:[§2\.1](https://arxiv.org/html/2607.29044#S2.SS1.p2.1),[§2\.3](https://arxiv.org/html/2607.29044#S2.SS3.p2.1)\.
- S\. Liang \(2021\)Information Processing of the Mencius Based on Commentary Literature \(基於註疏文獻的《孟子》信息處理研究\)\.Peking University Press,Beijing\.External Links:ISBN 978\-7\-301\-32098\-3Cited by:[§2\.2](https://arxiv.org/html/2607.29044#S2.SS2.p1.1)\.
- L\. Ling \(2021\)A Systematic Research Methodology towards Explanations in the Context:Taking Maoshi Guxun Zhuan\(《毛诗故训传》\)as an example\.Folklore and Research on the History of Linguistics Documentation 民俗典籍文字研究\(2\),pp\. 11–21, 256–257\.Cited by:[§1](https://arxiv.org/html/2607.29044#S1.p1.1)\.
- X\. Luo \(2005\)On Coreference Resolution Performance Metrics\.InProceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing,R\. Mooney, C\. Brew, L\. Chien, and K\. Kirchhoff \(Eds\.\),Vancouver, British Columbia, Canada,pp\. 25–32\.Cited by:[§5\.2](https://arxiv.org/html/2607.29044#S5.SS2.p2.4)\.
- C\. Ma, X\. Chen, W\. Qu, and P\. Lu \(2012a\)The Construction of The Analects of Confucius And Its Commentary Literatures Alignment Corpus\.Modern Educational Technology22\(7\),pp\. 109–113\.External Links:ISSN 1009\-8097Cited by:[§2\.2](https://arxiv.org/html/2607.29044#S2.SS2.p1.1)\.
- C\. Ma, X\. Chen, and W\. Qu \(2012b\)Automatic Analysis of Comments in Commentary Literatures\.Computer Science39\(10\),pp\. 220–223\.External Links:ISSN 1002\-137XCited by:[§2\.2](https://arxiv.org/html/2607.29044#S2.SS2.p1.1)\.
- Y\. Ma \(2022\)A Re\-examination of the Origin of the Two\-Annotation Edition of Historical Records\.Journal of Weinan Normal University37\(3\),pp\. 39–43, 64\.External Links:ISSN 1009\-5128,[Document](https://dx.doi.org/10.15924/j.cnki.1009-5128.2022.03.006)Cited by:[§1](https://arxiv.org/html/2607.29044#S1.p3.1),[§3\.4](https://arxiv.org/html/2607.29044#S3.SS4.p1.1)\.
- K\. Mo, Z\. Qiu, R\. Hu, and Y\. Wang \(2024\)Construction and Application of Ancient Chinese Allusion Resources\.Journal of Chinese Information Processing38\(11\),pp\. 27–34\.External Links:ISSN 1003\-0077Cited by:[§2\.1](https://arxiv.org/html/2607.29044#S2.SS1.p2.1)\.
- Q\. Ni \(2004\)An Outline of Textual Collation \(校勘學大綱\)\.Peking University Press,Beijing\.External Links:ISBN 978\-7\-301\-06918\-9Cited by:[§3](https://arxiv.org/html/2607.29044#S3.p1.1),[§6\.3](https://arxiv.org/html/2607.29044#S6.SS3.p1.1)\.
- Y\. Nie, Y\. Zhang, Y\. Peng, and L\. Yang \(2022\)Borrowing wisdom from world: modeling rich external knowledge for Chinese named entity recognition\.Neural Computing and Applications34\(6\),pp\. 4905–4922\.External Links:ISSN 1433\-3058,[Document](https://dx.doi.org/10.1007/s00521-021-06680-6)Cited by:[§2\.3](https://arxiv.org/html/2607.29044#S2.SS3.p2.1)\.
- A\. Paparnakis and C\. Domouchtsis \(2017\)Digital Greek Patristic Catena \(DGPC\)\. A brief presentation\.Journal of Data Mining & Digital HumanitiesSpecial Issue on Computer\-Aided Processing of Intertextuality in Ancient Languages\(Project presentations\)\.External Links:ISSN 2416\-5999,[Document](https://dx.doi.org/10.46298/jdmdh.4001)Cited by:[§2\.2](https://arxiv.org/html/2607.29044#S2.SS2.p2.1)\.
- M\. E\. Peters, M\. Neumann, M\. Iyyer, M\. Gardner, C\. Clark, K\. Lee, and L\. Zettlemoyer \(2018\)Deep Contextualized Word Representations\.InProceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 \(Long Papers\),M\. Walker, H\. Ji, and A\. Stent \(Eds\.\),New Orleans, Louisiana,pp\. 2227–2237\.External Links:[Document](https://dx.doi.org/10.18653/v1/N18-1202)Cited by:[§1](https://arxiv.org/html/2607.29044#S1.p5.1)\.
- S\. Pradhan, A\. Moschitti, N\. Xue, O\. Uryupina, and Y\. Zhang \(2012\)CoNLL\-2012 Shared Task: Modeling Multilingual Unrestricted Coreference in OntoNotes\.InJoint Conference on EMNLP and CoNLL \- Shared Task,S\. Pradhan, A\. Moschitti, and N\. Xue \(Eds\.\),Jeju Island, Korea,pp\. 1–40\.Cited by:[§5\.2](https://arxiv.org/html/2607.29044#S5.SS2.p2.4)\.
- S\. Pradhan, L\. Ramshaw, M\. Marcus, M\. Palmer, R\. Weischedel, and N\. Xue \(2011\)CoNLL\-2011 Shared Task: Modeling Unrestricted Coreference in OntoNotes\.InProceedings of the Fifteenth Conference on Computational Natural Language Learning: Shared Task,S\. Pradhan \(Ed\.\),Portland, Oregon, USA,pp\. 1–27\.Cited by:[§5\.2](https://arxiv.org/html/2607.29044#S5.SS2.p2.4)\.
- X\. Shen, Z\. Wang, S\. Li, C\. Huang, and G\. Zhou \(2026\)New Compendium of a Myriad of Plants: A New Dataset Describing Ancient Chinese Plants\.InFindings of the Association for Computational Linguistics: ACL 2026,M\. Liakata, V\. P\. Moreira, J\. Zhang, and D\. Jurgens \(Eds\.\),San Diego, California, United States,pp\. 1483–1498\.External Links:[Document](https://dx.doi.org/10.18653/v1/2026.findings-acl.73),ISBN 979\-8\-89176\-395\-1Cited by:[§2\.1](https://arxiv.org/html/2607.29044#S2.SS1.p2.1)\.
- L\. Shu, Y\. Guo, H\. Wang, X\. Zhang, and R\. Hu \(2022\)The Construction and Application of Ancient Chinese Corpus with Word Sense Annotation\.Journal of Chinese Information Processing36\(5\),pp\. 21–30\.External Links:ISSN 1003\-0077Cited by:[§2\.1](https://arxiv.org/html/2607.29044#S2.SS1.p2.1)\.
- \[53\]Q\. SimaRecords of the Grand Historian \(Shiji, 史記\)\.Baina eidtion edition\.Cited by:[§3\.2](https://arxiv.org/html/2607.29044#S3.SS2.p2.1),[§3\.4](https://arxiv.org/html/2607.29044#S3.SS4.p1.1)\.
- S\. Sun, R\. Yuan, Z\. Cao, W\. Li, and P\. Liu \(2024\)Prompt Chaining or Stepwise Prompt? Refinement in Text Summarization\.InFindings of the Association for Computational Linguistics: ACL 2024,L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Bangkok, Thailand,pp\. 7551–7558\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.findings-acl.449)Cited by:[§4\.1](https://arxiv.org/html/2607.29044#S4.SS1.p1.1)\.
- X\. Tang, Z\. Deng, Q\. Su, H\. Yang, and J\. Wang \(2024\)CHisIEC: An Information Extraction Corpus for Ancient Chinese History\.InProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation \(LREC\-COLING 2024\),N\. Calzolari, M\. Kan, V\. Hoste, A\. Lenci, S\. Sakti, and N\. Xue \(Eds\.\),Torino, Italia,pp\. 3192–3202\.Cited by:[§2\.1](https://arxiv.org/html/2607.29044#S2.SS1.p1.1)\.
- X\. Tang and Q\. Su \(2022\)That Slepen Al the Nyght with Open Ye\! Cross\-era Sequence Segmentation with Switch\-memory\.InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),S\. Muresan, P\. Nakov, and A\. Villavicencio \(Eds\.\),Dublin, Ireland,pp\. 7830–7840\.External Links:[Document](https://dx.doi.org/10.18653/v1/2022.acl-long.540)Cited by:[§2\.1](https://arxiv.org/html/2607.29044#S2.SS1.p1.1),[§2\.3](https://arxiv.org/html/2607.29044#S2.SS3.p2.1)\.
- X\. Tang, L\. Wang, and J\. Wang \(2026\)Language model collaboration for relation extraction from classical Chinese historical documents\.Information Processing & Management63\(1\),pp\. 104286\.External Links:ISSN 03064573,[Document](https://dx.doi.org/10.1016/j.ipm.2025.104286)Cited by:[§2\.1](https://arxiv.org/html/2607.29044#S2.SS1.p1.1)\.
- Y\. Tang \(2013\)Research of Automatically Recognizing Name in Pre\-Qin Ancient Chinese Classics\.New Technology of Library and Information Service\(Z1\),pp\. 63–68\.External Links:ISSN 1003\-3513Cited by:[§2\.1](https://arxiv.org/html/2607.29044#S2.SS1.p1.1)\.
- A\. Vaswani, N\. Shazeer, N\. Parmar, J\. Uszkoreit, L\. Jones, A\. N\. Gomez, Ł\. ukasz Kaiser, and I\. Polosukhin \(2017\)Attention is All you Need\.InAdvances in Neural Information Processing Systems,Vol\.30\.Cited by:[§1](https://arxiv.org/html/2607.29044#S1.p5.1)\.
- M\. Vilain, J\. Burger, J\. Aberdeen, D\. Connolly, and L\. Hirschman \(1995\)A Model\-Theoretic Coreference Scoring Scheme\.InSixth Message Understanding Conference \(MUC\-6\): Proceedings of a Conference Held in Columbia, Maryland, November 6\-8, 1995,Cited by:[§5\.2](https://arxiv.org/html/2607.29044#S5.SS2.p2.4)\.
- B\. Wang, X\. Shi, Z\. Tan, Y\. Chen, and W\. Wang \(2016\)A Sentence Segmentation Method for Ancient Chinese Texts Based on NNLM\.InChinese Lexical Semantics,M\. Dong, J\. Lin, and X\. Tang \(Eds\.\),Cham,pp\. 387–396\.External Links:[Document](https://dx.doi.org/10.1007/978-3-319-49508-8%5F36),ISBN 978\-3\-319\-49508\-8Cited by:[§2\.1](https://arxiv.org/html/2607.29044#S2.SS1.p1.1)\.
- D\. Wang, C\. Liu, Z\. Zhu, J\. Liu, H\. Hu, S\. Shen, and B\. Li \(2022\)Construction and Application of Pre\-trained Models of Siku Quanshu in Orientation to Digital Humanities\.Library Tribune42\(6\),pp\. 31–43\.External Links:ISSN 1002\-1167Cited by:[§2\.1](https://arxiv.org/html/2607.29044#S2.SS1.p1.1),[footnote 3](https://arxiv.org/html/2607.29044#footnote3)\.
- N\. Wang \(1996\)Principles of Chinese Exegetics \(訓詁學原理\)\.China International Broadcasting Press,Beijing\.External Links:ISBN 978\-7\-5078\-1378\-4Cited by:[§3\.2](https://arxiv.org/html/2607.29044#S3.SS2.p1.1),[§3\.3](https://arxiv.org/html/2607.29044#S3.SS3.p1.1)\.
- P\. Wang and Z\. Ren \(2022\)The Uncertainty\-based Retrieval Framework for Ancient Chinese CWS and POS\.InProceedings of the Second Workshop on Language Technologies for Historical and Ancient Languages,R\. Sprugnoli and M\. Passarotti \(Eds\.\),Marseille, France,pp\. 164–168\.Cited by:[§2\.1](https://arxiv.org/html/2607.29044#S2.SS1.p1.1),[footnote 3](https://arxiv.org/html/2607.29044#footnote3)\.
- S\. Wang \(2018\)The Study of the Sentence Alignment between Shijing with Its Commentations and Annotations\.Master’s Thesis,Nanjing Agricultural University\.External Links:[Document](https://dx.doi.org/10.27244/d.cnki.gnjnu.2018.000860)Cited by:[§2\.2](https://arxiv.org/html/2607.29044#S2.SS2.p1.1)\.
- Y\. Wang \(2016\)On the Printing Time of “Zishu”\.Journal of Henan University of Science & Technology\(Social Science\)34\(1\),pp\. 23–25\.External Links:ISSN 1672\-3910,[Document](https://dx.doi.org/10.15926/j.cnki.hkdsk.2016.01.004)Cited by:[§6\.2](https://arxiv.org/html/2607.29044#S6.SS2.p3.1)\.
- Z\. Wang, S\. Zhang, R\. Hu, and X\. Zhang \(2024\)Ancient Chinese Language Resource on Tongjiazi: Construction and Application\.Journal of Chinese Information Processing38\(3\),pp\. 152–162\.External Links:ISSN 1003\-0077Cited by:[§2\.1](https://arxiv.org/html/2607.29044#S2.SS1.p2.1)\.
- E\. P\. Wilkinson \(2018\)Chinese History: A New Manual\.Endymion Wilkinson, c/o Harvard University Asia Center\.External Links:ISBN 978\-0\-9988883\-0\-9Cited by:[§1](https://arxiv.org/html/2607.29044#S1.p2.1)\.
- R\. Xu and X\. Chen \(2012\)A Method of Segmentation on ”Zuo Zhuan” by Using Commentaries\.Journal of Chinese Information Processing26\(2\),pp\. 13–17, 45\.External Links:ISSN 1003\-0077Cited by:[§2\.3](https://arxiv.org/html/2607.29044#S2.SS3.p2.1)\.
- R\. Xu and S\. Liang \(2019\)Study on the Automatic Alignment of the Contents Between Zuo Zhuan and its Annotated Documents\.Journal of Jinling Institute of Technology\(Social Science\)33\(2\),pp\. 84–88\.External Links:ISSN 1673\-131X,[Document](https://dx.doi.org/10.16515/j.cnki.32-1745/c.2019.02.019)Cited by:[§2\.2](https://arxiv.org/html/2607.29044#S2.SS2.p1.1),[§2\.3](https://arxiv.org/html/2607.29044#S2.SS3.p1.1)\.
- Y\. Xu \(2014\)Illustrative Cases in the Collation of Ancient Chinese Texts \(Revised and Expanded Edition\) \(古籍整理釋例\)\.Zhonghua shu ju,Beijing\.External Links:ISBN 978\-7\-101\-10079\-2Cited by:[§3\.4](https://arxiv.org/html/2607.29044#S3.SS4.p1.1)\.
- Y\. Xu, C\. Mao, Z\. Wang, G\. Jin, L\. Zhong, and T\. Qian \(2024\)Semantic\-enhanced graph neural network for named entity recognition in ancient Chinese books\.Scientific Reports14\(1\),pp\. 17488\.External Links:ISSN 2045\-2322,[Document](https://dx.doi.org/10.1038/s41598-024-68561-x)Cited by:[§2\.1](https://arxiv.org/html/2607.29044#S2.SS1.p1.1)\.
- F\. Yang, Q\. Zhang, W\. Xing, C\. Liu, and L\. Liu \(2026\)Research on graph\-retrieval augmented generation based on historical text knowledge graphs\.Digital Scholarship in the Humanities,pp\. fqag006\.External Links:ISSN 2055\-7671,[Document](https://dx.doi.org/10.1093/llc/fqag006)Cited by:[§2\.1](https://arxiv.org/html/2607.29044#S2.SS1.p2.1)\.
- K\. Yasuoka \(2022a\)Roberta Classical Chinese Base Char\.Cited by:[footnote 3](https://arxiv.org/html/2607.29044#footnote3)\.
- K\. Yasuoka \(2022b\)Roberta Classical Chinese Large Char\.Cited by:[footnote 3](https://arxiv.org/html/2607.29044#footnote3)\.
- K\. Yasuoka \(2023\)Sequence\-Labeling RoBERTa Model for Dependency\-Parsing in Classical Chinese and Its Application to Vietnamese and Thai\.In2023 8th International Conference on Business and Industrial Research \(ICBIR\),pp\. 169–173\.External Links:[Document](https://dx.doi.org/10.1109/ICBIR57571.2023.10147628)Cited by:[§2\.1](https://arxiv.org/html/2607.29044#S2.SS1.p1.1)\.
- S\. Ying \(2005\)On the study of Shiji Jijie\.Journal of Ancient Books Collation\(2\),pp\. 42–47\.External Links:ISSN 1009\-1017Cited by:[§1](https://arxiv.org/html/2607.29044#S1.p2.1),[§3\.4](https://arxiv.org/html/2607.29044#S3.SS4.p1.1)\.
- C\. Zhang \(2007\)Studies on the Classic of Mountains and Seas \(《山海經》研究\)\.Shanghai Academy of Social Sciences Press,Shanghai\.External Links:ISBN 978\-7\-80745\-071\-9Cited by:[§6\.3](https://arxiv.org/html/2607.29044#S6.SS3.p1.1)\.
- R\. Zhang, C\. Nogueira dos Santos, M\. Yasunaga, B\. Xiang, and D\. Radev \(2018\)Neural Coreference Resolution with Deep Biaffine Attention by Joint Mention Detection and Mention Clustering\.InProceedings of the 56th Annual Meeting of the Association for Computational Linguistics \(Volume 2: Short Papers\),I\. Gurevych and Y\. Miyao \(Eds\.\),Melbourne, Australia,pp\. 102–107\.External Links:[Document](https://dx.doi.org/10.18653/v1/P18-2017)Cited by:[§4\.2](https://arxiv.org/html/2607.29044#S4.SS2.p2.1)\.
- X\. Zhang \(2011\)The Definition of Collected Annotations\.Journal of Tangshan Teachers College33\(4\),pp\. 4–7\.External Links:ISSN 1009\-9115Cited by:[§1](https://arxiv.org/html/2607.29044#S1.p2.1),[§3\.4](https://arxiv.org/html/2607.29044#S3.SS4.p1.1)\.
- M\. Zhou \(2019\)Shanhai Jing Jishi \(山海經集釋\)\.Bashu Publishing House,Chengdu\.External Links:ISBN 978\-7\-5531\-1134\-6Cited by:[§3\.4](https://arxiv.org/html/2607.29044#S3.SS4.p1.1)\.
- D\. Zhu, Z\. Zhixiao, N\. Wu, and X\. Wang \(2024\)Research on Word Segmentation of Ancient Books Based on Domain Large Language Model\.Journal of Scientific Information Research6\(2\)\.Cited by:[§2\.1](https://arxiv.org/html/2607.29044#S2.SS1.p1.1),[§2\.3](https://arxiv.org/html/2607.29044#S2.SS3.p2.1)\.
- \[83\]Z\. Zhuang, X\. Guo, D\. Lu, and Y\. SunZhuangzi \(莊子\)\.Cited by:[§3\.2](https://arxiv.org/html/2607.29044#S3.SS2.p3.1)\.