Pith. sign in

REVIEW 3 major objections 4 minor 111 references

This paper claims that collected commentaries on classical Chinese texts can be compiled automatically by linking annotations to their true main-text targets and clustering them across editions, preserving interpretive context.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 14:43 UTC pith:BKJK7VYS

load-bearing objection Novel task formulation with solid philological grounding, but the 97% clustering F1 rest on an evaluation loop that needs an external split before it can be trusted. the 3 major comments →

arxiv 2607.29044 v1 pith:BKJK7VYS submitted 2026-07-31 cs.CL

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

Ke Liang , Qi Su , Churen Huang This is my paper
classification cs.CL
keywords classical Chinese exegesiscollected commentariesinline notesprompt chainingcross-source mention clusteringcontextual exegesisClassic of Mountainsexegetical functions
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that the centuries-old practice of compiling collected commentaries (jijie) can be automated for classical Chinese texts. Inline notes in five editions of the Classic of Mountains are attached to specific spans of main text, but often actually explain a different span; the authors argue that a two-step LLM prompt chain can recover the true target and classify each note's exegetical function, and that cross-edition clustering can then group notes about the same passage. If correct, this would produce structured, context-preserving exegetical knowledge at scale, supporting textual collation, lexicography, and historical entity research, plus NLP tasks on classical Chinese. The central quantitative claim is a CoNLL F1 above 97% on clustering for the tested section.

Core claim

The central discovery is that collected-commentary compilation reduces to two coupled subtasks: pinpointing the main-text segment each note actually explains (not merely the segment it is attached to) and clustering semantically linked notes across editions. The framework introduced here—two-step prompt chaining for anchoring and refining, followed by union-find clustering constrained by source uniqueness and complete linkage—achieves a CoNLL F1 of 97.7% on five editions of the Classic of Mountains, and produces 2,440 clusters from 5,463 core commentaries. The authors further show that a threefold classification of exegetical functions (textual criticism, lexical exegesis, referential exeges

What carries the argument

The load-bearing mechanism is the distinction between a commentary's anchor_text (the span it is physically attached to) and its actual exegetical target, which may lie elsewhere. Unpacking this requires a two-step prompt chain: an Anchoring step that identifies the anchor span and entry unit, and a Refining step that identifies the true target and decomposes the commentary into exegetical functions (textual criticism, lexical exegesis, referential exegesis). Clustering then embeds the anchor_text with SikuRoBERTa and merges commentaries across editions via a two-phase Union-Find algorithm with three constraints (cluster-size limits, complete linkage for commentary similarity, and source uni

Load-bearing premise

The 97% result rests on the assumption that the single manually verified section—about 18% of the dataset—is representative of all five editions and that model and threshold choices were not tuned on that same gold data.

What would settle it

Run the same pipeline on the rest of the Classic of Mountains with gold clusters created independently, and compare CoNLL F1; if it falls well below 97%, the reported accuracy does not transfer. Alternatively, check whether the clusters agree with an existing scholarly collected commentary (e.g., Shanhaijing Jishi) at the level claimed.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the framework generalizes, scholars can automatically assemble collected commentaries for other classical Chinese texts from multiple editions, replacing time-consuming manual collation.
  • The structured clusters, together with the exegetical-function labels, provide a data foundation for textual criticism, variant collection, lost-text reconstruction, and historical named-entity identification.
  • Variant-character patterns among editions can serve as a quantitative signal for edition genealogy, as demonstrated by the dendrogram that matches traditional philological analysis.
  • The labeled exegetical knowledge can be used to adapt modern NLP tasks—word segmentation, NER, summarization, entailment—to classical Chinese and to analyze consistency across commentators.
  • The classification statistics reveal a historical shift in exegetical practice: early commentary is dominated by lexical exegesis, while Qing-era commentary is dominated by referential and textual criticism, consistent with the rise of evidential scholarship.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the anchor-vs-target distinction is as pervasive as the paper suggests, similar two-step identification may improve automatic annotation alignment for other traditions with embedded glosses, such as Talmudic or biblical catenae, where a gloss's placement and its referent also diverge.
  • The reported 97% F1 may depend on the unusually high textual similarity (above 96%) among the five editions; testing on editions with larger divergence would reveal whether the clustering constraints still hold.
  • The union-find constraints—especially source uniqueness—encode a strong assumption that no two commentaries from the same edition should be clustered together; relaxing this could allow detection of later commentaries that explicitly reference multiple earlier notes.
  • A direct validation would compare the automatically generated groups against an existing modern collected commentary such as Shanhaijing Jishi; high agreement would independently confirm the integration criteria.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper introduces a new NLP task—the automatic compilation of collected commentaries (jijie) from inline notes in classical Chinese texts—and proposes a two-stage framework: (1) two-step LLM prompt chaining to identify the main-text segment anchored by each commentary and to classify its exegetical function (TC/LE/RE), and (2) cross-source mention clustering over five editions of the Classic of Mountains (SJ) using contextual embeddings and a Union-Find algorithm. The authors report 84% end-to-end extraction accuracy and a CoNLL F1 of 0.977 in a case study on the Classic of Northern Mountains section, and discuss philological downstream tasks.

Significance. The task formulation is a genuine contribution: existing computational work on Chinese commentaries has focused on alignment within one edition, whereas this paper targets the historically important problem of integrating commentary across editions while preserving the main-text–b-commentary–c-commentary hierarchy. The use of traditional exegetics to motivate the distinction between anchor_text and exegetical target is thoughtful, and the resource built from five historical editions is potentially valuable. If the reported performance is robust and independently reproducible, the framework would be a useful foundation for digital-humanities applications. However, the load-bearing empirical claim rests on an evaluation whose independence from model and threshold selection is not established; as reported, the evidence is not fully convincing.

major comments (3)
  1. [§5.1–§5.2] The central claim of CoNLL F1 > 97% (Table 4) is not fully supported as reported because the independence between model selection and the gold evaluation is not demonstrated. Section 5.1 states that DeepSeek and SikuRoBERTa were selected, and that clustering thresholds (pos_p05/pos_mean) and layer choices were taken from “manually curated datasets,” after evaluating 3 LLMs and 7 BERT models. Section 5.2 evaluates on the Classic of Northern Mountains section (18.044% of the dataset), which was “manually verified and used as the gold standard.” The manuscript never states that the §5.1 manually curated datasets are disjoint from this gold section. If they overlap, the reported 97.7% is optimistically biased. Please specify an explicit train/development/test split, state clearly which portions were used for model selection, threshold setting, and final evaluation, and provide per-section re
  2. [Table 3 vs. Table 4] The 84% end-to-end extraction accuracy (Table 3), with S1 anchor_text accuracy of 82.7%, is difficult to reconcile with a clustering CoNLL F1 of 97.7% when clustering is performed on the predicted anchor_text. Either the clustering is highly insensitive to anchor_text errors, or the evaluation protocol uses oracle/human-corrected anchors for clustering, or the extraction and clustering evaluations are applied to different data. The paper should report clustering metrics computed on (a) predicted anchor_text and (b) gold anchor_text, and provide error analysis showing how extraction errors propagate (or do not propagate) to clustering. Without this, Table 4 may not reflect the actual end-to-end pipeline claimed in the abstract.
  3. [§4.2 / Algorithm 1] The clustering method relies on several hyperparameters—top-k candidate size, thresholds T_c and T_t, similarity function, and the source-uniqueness constraint—but no sensitivity analysis or ablation is reported. In particular, the choice of anchor_text as the clustering key, rather than entry_unit or exegetical target, is a central design decision that is only argued qualitatively (the DW9.7 example). Since this assumption directly drives the reported F1, a small ablation comparing clustering variants (anchor_text vs. entry_unit vs. target, and with/without the commentary-similarity route) would substantially strengthen the empirical contribution.
minor comments (4)
  1. [§5.1] The threshold symbols pos_p05 and pos_mean are used without definitions. Please define them precisely (e.g., quantile of pairwise anchor-text similarities in the development set) and report the actual values used.
  2. [Algorithm 1] The source-uniqueness condition in Phase 2, line 15, uses `source(Ci) ∩ source(Cj) ⊆ {C}`. The notation is unclear; if C refers to Edition C, the condition should be explained explicitly. This constraint appears to allow one particular edition to appear in multiple clusters, which is not intuitive.
  3. [§5.2] The gold section is described only as “18.044% of the dataset.” Please report the number of paragraphs, commentaries, c-commentaries, and clusters in this section, and state how the remaining sections are distributed; the representativeness of this one section for the full SJ is otherwise hard to judge.
  4. [Throughout] There are several typos and rendering artifacts: “facilitate the understand” (§1), “areasi” (§2), “refelcting” (§5.3), “ancho_text” (§5.2), and stray OCR markers such as “oQiuru Mountain p”. These should be cleaned before publication.

Circularity Check

0 steps flagged

No significant circularity in the derivation; the 97% F1 is an empirical result, though the §5.1/§5.2 evaluation split is not documented.

full rationale

The paper does not present a mathematical derivation in which a predicted quantity is equal to an input by construction. The framework (§4) defines subtasks and evaluates them empirically. The only self-citation (Liang et al., 2024) appears in related work and is not load-bearing. A methodological caveat exists: §5.1 selects models and clustering thresholds on 'manually curated datasets' and §5.2 evaluates on a 'manually verified' section, without stating that these data are disjoint; and the gold standard is described as manually verified results, which can introduce optimism. However, under the hard rules, this is a missing independence statement, not an exhibited reduction. Without evidence that the §5.1 curated data coincide with the §5.2 gold section, or that the gold was generated directly from the predicted output without independent annotation, the 97% F1 cannot be shown to be forced by construction. The paper's theoretical foundations and task formulation are self-contained; no uniqueness theorem or ansatz is imported from the authors' prior work.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

This is an empirical system paper, so the axiom burden is not about mathematical axioms. It is about data representativeness, clustering criteria, and in-sample selection of models and thresholds. The central 97% number inherits all of these assumptions.

free parameters (4)
  • Clustering thresholds Tc and Tt = Not reported numerically; set from pos_p05/pos_mean similarity values
    Used in Algorithm 1 Phases 1–2. §5.1 says thresholds are chosen based on model performance on manually curated data; the same data feeds the gold evaluation in §5.2, so the F1 is partly in-sample.
  • Top-k candidate retrieval size k = Not reported
    Algorithm 1 retrieves top-k candidates before union-find; without k the exact pipeline cannot be reproduced.
  • SikuRoBERTa embedding layer choices = Last 5 hidden layers for anchor_text; last 2 for commentaries
    Chosen in §5.1 without ablation or held-out validation; these choices directly affect all similarity scores and cluster boundaries.
  • Cluster-size limits in Union-Find = Not reported
    Algorithm 1 constraint (1) is not quantified; the limits can materially change whether clusters merge.
axioms (4)
  • domain assumption Collected-commentary integration is primarily determined by shared main-text segments, with semantic similarity as a supplementary criterion for first-layer commentaries.
    §3.4 operationalizes integration criteria. The paper itself gives a counterexample where Shanhaijing Jishi groups commentaries on similar-but-not-identical segments; the model does not handle that case.
  • domain assumption The Northern Mountains section's manual gold standard is representative of the full SJ and is reliable.
    §5.2 evaluates on 18.044% of the dataset; no inter-annotator agreement, random sampling, or section-wise breakdown is given.
  • domain assumption High character-frequency similarity across editions means anchor_text spans can be compared with fixed embedding thresholds.
    §5 reports >96% character-frequency similarity for Guo's text/commentary; if textual variants or Wang's placement conventions distort embeddings, the fixed thresholds would over- or under-merge.
  • ad hoc to paper A commentary's anchor_text is a better clustering key than entry_unit or exegetical target.
    §4.2 argues anchor_text avoids merging different exegetical focuses; this is a design choice supported only by examples and not by an ablation.
invented entities (1)
  • anchor_text vs. exegetical target distinction no independent evidence
    purpose: Data model for separating where a note is physically attached from what it actually explains
    Introduced in §3.2 and illustrated with AN2.5 and DW9.7. It is a useful operational construct, but no external benchmark validates it as ground truth.

pith-pipeline@v1.3.0-daily-deepseek · 15180 in / 15336 out tokens · 151533 ms · 2026-08-03T14:43:55.039182+00:00 · methodology

0 comments
read the original 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 F1 score above 97% in a case study on the Classic 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.

Figures

Figures reproduced from arXiv: 2607.29044 by Churen Huang, Ke Liang, Qi Su.

Figure 1
Figure 1. Figure 1: The Xuncao passage of Edition D. such as “任臣案”,“沅曰”and “懿行案”( “Renchen/Yuan/Yixing comments”). By con￾trast, Edition C integrates Guo’s annotations with Wang’s own interpretation into a single layer, due to the underdevelopment of eviden￾tial scholarship in Wang’s time (Chen, 2012). Accordingly, Editions A and C contain only c-commentary, while the other three contain both. The commentary placement in Edit… view at source ↗
Figure 2
Figure 2. Figure 2: Two-step prompt chaining pipeline. main-text segment may be the anchor_text, entry_unit or 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 in SJ. We therefore use the anchor_text for clustering. Using the entry_unit or target alone may merge com￾mentaries with different exegetical f… view at source ↗
Figure 3
Figure 3. Figure 3: Experimental results from the “Mountain Qiuru”paragraph and their possible applications. 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 sug￾gest that variant-character patterns can serve as a quantitative indicator of edition geneal￾ogy. The resulting dendrogram in [PITH_FUL… view at source ↗
Figure 4
Figure 4. Figure 4: Edition clustering by variant pairs. computational framework for preserving the contextual dependency of inline notes and or￾ganizing historical exegetical knowledge. The main limitation of this work is that our ex￾periments are limited to the SJ, although the proposed framework is generalizable to other ancient texts. Future work includes extending the framework to a broader range of ancient texts, invest… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

111 extracted references · 34 canonical work pages

  1. [1]

    Algorithms for Scoring Coreference Chains , author =. Proc. Linguistic Coreference Workshop at the first Conf. on Language Resources and Evaluation (LREC), Granada, Spain, May 1998 , urldate =

    1998

  2. [2]

    , year = 2016, month = apr, journal =

    Berti, Monica and Almas, Bridget and Crane, Gregory R. , year = 2016, month = apr, journal =. The

    2016

  3. [3]

    , year = 2014, journal =

    Berti, Monica and Almas, Bridget and Dubin, David and Franzini, Greta and Stoyanova, Simona and Crane, Gregory R. , year = 2014, journal =. The. doi:10.4000/jtei.1218 , urldate =

    work page doi:10.4000/jtei.1218 2014

  4. [4]

    ACM Transactions on Knowledge Discovery from Data (TKDD) , volume =

    Collective Entity Resolution in Relational Data , author =. ACM Transactions on Knowledge Discovery from Data (TKDD) , volume =. doi:10.1145/1217299.1217304 , urldate =

    arXiv

  5. [5]

    Bingenheimer, Marcus , year = 2015, month = nov, journal =. The. doi:10.1186/s40655-015-0007-3 , urldate =

    work page doi:10.1186/s40655-015-0007-3 2015

  6. [6]

    Language

    Brown, Tom and Mann, Benjamin and Ryder, Nick and Subbiah, Melanie and Kaplan, Jared D and Dhariwal, Prafulla and Neelakantan, Arvind and Shyam, Pranav and Sastry, Girish and Askell, Amanda and Agarwal, Sandhini and. Language. Advances in

  7. [7]

    Commentaria in

    CAGB , journal =. Commentaria in

  8. [8]

    doi:10.13530/j.cnki.jlis.2022013 , urldate =

    Chen, Li , year = 2022, journal =. doi:10.13530/j.cnki.jlis.2022013 , urldate =

    work page doi:10.13530/j.cnki.jlis.2022013 2022

  9. [9]

    , year = 2026, month = mar, journal =

    Chen, Yuqi and Shang, Wenyi and Wang, Hongsu and Zhang, Sophia and Bol, Peter K. , year = 2026, month = mar, journal =. Geocoding the Past World: Unearthing Coordinates of Early. doi:10.1080/13658816.2025.2491711 , urldate =

    arXiv 2026

  10. [10]

    doi:10.13365/j.jirm.2025.01.139 , urldate =

    Chen, Tao and Zhao, Xiaofei and Yang, Xin and Lin, Lixin , year = 2025, journal =. doi:10.13365/j.jirm.2025.01.139 , urldate =

    work page doi:10.13365/j.jirm.2025.01.139 2025

  11. [11]

    Chen, Lianshan , year = 2012, publisher =

    2012

  12. [12]

    Chen, Xiaohe and Xu,

  13. [13]

    Decontextualization:

    Choi, Eunsol and Palomaki, Jennimaria and Lamm, Matthew and Kwiatkowski, Tom and Das, Dipanjan and Collins, Michael , year = 2021, month = apr, journal =. Decontextualization:. doi:10.1162/tacl_a_00377 , urldate =

    doi:10.1162/tacl_a_00377 2021

  14. [14]

    Linking Past Insights with Contemporary Understanding: An Ontological and Knowledge Graph Approach to the Transmission of Ancient

    Cui, Yu and Yao, Shenjun and Wu, Jianping and Lv, Muhan , year = 2024, month = nov, journal =. Linking Past Insights with Contemporary Understanding: An Ontological and Knowledge Graph Approach to the Transmission of Ancient. doi:10.1186/s40494-024-01504-x , urldate =

    work page doi:10.1186/s40494-024-01504-x 2024

  15. [15]

    Diachronic

    Deng, Zekun and Yang, Hao and Wang, Jun , editor =. Diachronic. Neural. doi:10.1007/978-981-99-8145-8_24 , abstract =

    work page doi:10.1007/978-981-99-8145-8_24

  16. [16]

    Proceedings of the 2019

    Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina , editor =. Proceedings of the 2019. doi:10.18653/v1/N19-1423 , urldate =

    doi:10.18653/v1/n19-1423 2019

  17. [17]

    Disentangling the Cultural Evolution of Ancient

    Duan, Siyu and Wang, Jun and Yang, Hao and Su, Qi , year = 2023, journal =. Disentangling the Cultural Evolution of Ancient. doi:10.1057/s41599-023-01811-x , urldate =

    work page doi:10.1057/s41599-023-01811-x 2023

  18. [18]

    Manuscript

    Dobcheva, Ivana and Mackert, Christoph , year = 2018, month = dec, journal =. Manuscript. doi:10.24446/rx89 , urldate =

    work page doi:10.24446/rx89 2018

  19. [19]

    Character-

    Dong, Chuanhai and Zhang, Jiajun and Zong, Chengqing and Hattori, Masanori and Di, Hui , editor =. Character-. Natural. doi:10.1007/978-3-319-50496-4_20 , abstract =

    work page doi:10.1007/978-3-319-50496-4_20

  20. [20]

    doi:10.16093/j.cnki.ccc.1993.01.021 , urldate =

    Dong, Hongli , year = 1993, journal =. doi:10.16093/j.cnki.ccc.1993.01.021 , urldate =

    work page doi:10.16093/j.cnki.ccc.1993.01.021 1993

  21. [21]

    Restoring

    Duan, Siyu and Wang, Jun and Su, Qi , editor =. Restoring. Proceedings of the 2024

    2024

  22. [22]

    Ethayarajh, Kawin , editor =. How. Proceedings of the 2019. doi:10.18653/v1/D19-1006 , urldate =

    doi:10.18653/v1/d19-1006 2019

  23. [23]

    Feng, Haofei , year = 2003, publisher =

    2003

  24. [24]

    , year = 1998, month = may, journal =

    Gardner, Daniel K. , year = 1998, month = may, journal =. Confucian. doi:10.2307/2658830 , urldate =

    work page doi:10.2307/2658830 1998

  25. [25]

    Proceedings of the VLDB Endowment , volume =

    Record Linkage with Uniqueness Constraints and Erroneous Values , author =. Proceedings of the VLDB Endowment , volume =. doi:10.14778/1920841.1920897 , urldate =

    arXiv

  26. [26]

    Guo, Zaiyi , year = 2005, publisher =

    2005

  27. [27]

    Theories and

    Guo, Yingde and Yu, Xuetang and Li, Xiaolong , year = 2024, publisher =. Theories and

    2024

  28. [28]

    Han, Geping , year = 1989, journal =

    1989

  29. [29]

    Collected

    He, Yan , edition =. Collected

  30. [30]

    HMT , journal =. The

  31. [31]

    Classical

    Huang, Hen-Hsen and Sun, Chuen-Tsai and Chen, Hsin-Hsi , year = 2010, urldate =. Classical

    2010

  32. [32]

    Huang, Yongnian , year = 2001, publisher =

    2001

  33. [33]

    Huang, Shilei and Wu, Jiangqin , editor =. A. Proceedings of the

  34. [34]

    doi:10.13266/j.issn.0252-3116.2015.12.020 , urldate =

    Huang, Shuiqing and Wang, Dongbo and He, Lin , year = 2015, journal =. doi:10.13266/j.issn.0252-3116.2015.12.020 , urldate =

    arXiv 2015

  35. [35]

    doi:10.16783/j.cnki.nwnus.1999.03.020 , urldate =

    Huang, Yaping , year = 1999, journal =. doi:10.16783/j.cnki.nwnus.1999.03.020 , urldate =

    work page doi:10.16783/j.cnki.nwnus.1999.03.020 1999

  36. [36]

    Hu, Renfen and Li, Shen and Zhu, Yuchen , year = 2021, journal =

    2021

  37. [37]

    Hu, Puan and Wang, Qihe , year = 2025, publisher =

    2025

  38. [38]

    Jawahar, Ganesh and Sagot, Beno. What. Proceedings of the 57th. doi:10.18653/v1/P19-1356 , urldate =

    doi:10.18653/v1/p19-1356

  39. [39]

    doi:10.14037/j.cnki.tsgxk.2015.05.015 , urldate =

    Jia, Fengxu , year = 2015, journal =. doi:10.14037/j.cnki.tsgxk.2015.05.015 , urldate =

    work page doi:10.14037/j.cnki.tsgxk.2015.05.015 2015

  40. [40]

    Research on

    Kang, Wenjun and Zuo, Jiali and Dai, Qili and Hu, Yiyu and Wang, Mingwen , editor =. Research on. Natural. doi:10.1007/978-981-97-9431-7_37 , abstract =

    work page doi:10.1007/978-981-97-9431-7_37

  41. [41]

    Graph-Based

    Kardes, Hakan and Konidena, Deepak and Agrawal, Siddharth and Huff, Micah and Sun, Ang , editor =. Graph-Based. Proceedings of

  42. [42]

    Liang, Ke and Huang, Chu-Ren and Jiang, Xin-Lan , editor =. From. Proceedings of the 2024

    2024

  43. [43]

    Liang, Shehui , year = 2021, month = may, publisher =

    2021

  44. [44]

    Capsules

    Li, Si and Li, Mingzheng and Xu, Yajing and Bao, Zuyi and Fu, Lu and Zhu, Yan , year = 2018, journal =. Capsules. doi:10.1109/ACCESS.2018.2881280 , urldate =

    arXiv 2018

  45. [45]

    Li, Wei and Shao, Yanqiu and Bi, Mengxi and Cui, Xiaoya , year = 2024, journal =

    2024

  46. [46]

    doi:10.16603/j.issn1002-1027.2020.05.011 , urldate =

    Li, Bin and Wang, Lu and Chen, Xiaohe and Wang, Dongbo , year = 2020, journal =. doi:10.16603/j.issn1002-1027.2020.05.011 , urldate =

    arXiv 2020

  47. [47]

    Ling, Lijun , year = 2021, journal =

    2021

  48. [48]

    Li, Xiaojie and Yang, Changyu and Wang, Yuhai and Qu, Kale , year = 2021, publisher =

    2021

  49. [49]

    Li, Wei and Li, Yi and Shao, Yanqiu and Bi, Mengxi , year = 2025, month = apr, journal =. A. doi:10.1145/3725733 , urldate =

    work page doi:10.1145/3725733 2025

  50. [50]

    Liu, Shi and Li, Feiyue , year = 2021, journal =

    2021

  51. [51]

    Luo, Xiaoqiang , editor =. On. Proceedings of

  52. [52]

    Ma, Chuangxin and Chen, Xiaohe and Qu, Weiguang , year = 2012, journal =

    2012

  53. [53]

    Ma, Chuangxin and Chen, Xiaohe and Qu, Weiguang and Lu, Pengfei , year = 2012, journal =

    2012

  54. [54]

    doi:10.13366/j.dik.2013.01.011 , urldate =

    Ma, Chuangxin and Chen, Xiaohe , year = 2013, journal =. doi:10.13366/j.dik.2013.01.011 , urldate =

    work page doi:10.13366/j.dik.2013.01.011 2013

  55. [55]

    doi:10.15924/j.cnki.1009-5128.2022.03.006 , urldate =

    Ma, Yingjie , year = 2022, journal =. doi:10.15924/j.cnki.1009-5128.2022.03.006 , urldate =

    arXiv 2022

  56. [56]

    doi:10.13663/j.cnki.lj.2017.08.010 , urldate =

    Ma, Chuangxin and Chen, Xiaohe , year = 2017, journal =. doi:10.13663/j.cnki.lj.2017.08.010 , urldate =

    work page doi:10.13663/j.cnki.lj.2017.08.010 2017

  57. [57]

    Ma, Chuangxin and Chen, Xiaohe and Qu, Weiguang , year = 2013, journal =

    2013

  58. [58]

    Mo, Kaijie and Qiu, Ziliang and Hu, Renfen and Wang, Yupei , year = 2024, journal =

    2024

  59. [59]

    Borrowing Wisdom from World: Modeling Rich External Knowledge for

    Nie, Yu and Zhang, Yilai and Peng, Yongkang and Yang, Lisha , year = 2022, month = mar, journal =. Borrowing Wisdom from World: Modeling Rich External Knowledge for. doi:10.1007/s00521-021-06680-6 , urldate =

    work page doi:10.1007/s00521-021-06680-6 2022

  60. [60]

    Ni, Qixin , year = 2004, publisher =

    2004

  61. [61]

    Information

    Orlandi, Silvia and. Information

  62. [62]

    Pan, Xiaomeng and Wang, Hongfei and Oka, Teruaki and Komachi, Mamoru , editor =. Zuo. Proceedings of the 2022. doi:10.18653/v1/2022.naacl-srw.17 , urldate =

    work page doi:10.18653/v1/2022.naacl-srw.17 2022

  63. [63]

    Paparnakis, Athanasios and Domouchtsis, Constantinos , year = 2017, month = oct, journal =. Digital. doi:10.46298/jdmdh.4001 , urldate =

    work page doi:10.46298/jdmdh.4001 2017

  64. [64]

    and Neumann, Mark and Iyyer, Mohit and Gardner, Matt and Clark, Christopher and Lee, Kenton and Zettlemoyer, Luke , editor =

    Peters, Matthew E. and Neumann, Mark and Iyyer, Mohit and Gardner, Matt and Clark, Christopher and Lee, Kenton and Zettlemoyer, Luke , editor =. Deep. Proceedings of the 2018. doi:10.18653/v1/N18-1202 , urldate =

    doi:10.18653/v1/n18-1202 2018

  65. [65]

    Proceedings of the

    Pradhan, Sameer and Ramshaw, Lance and Marcus, Mitchell and Palmer, Martha and Weischedel, Ralph and Xue, Nianwen , editor =. Proceedings of the

  66. [66]

    Pradhan, Sameer and Moschitti, Alessandro and Xue, Nianwen and Uryupina, Olga and Zhang, Yuchen , editor =. Joint

  67. [67]

    Shen, Xiaobin and Wang, Zhongqing and Li, Shichen and Huang, Chu-Ren and Zhou, Guodong , editor =. New. Findings of the. doi:10.18653/v1/2026.findings-acl.73 , urldate =

    work page doi:10.18653/v1/2026.findings-acl.73 2026

  68. [68]

    Shu, Lei and Guo, Yiluan and Wang, Huiping and Zhang, Xuetao and Hu, Renfen , year = 2022, journal =

    2022

  69. [69]

    Records of the

    Sima, Qian , edition =. Records of the

  70. [70]

    Sun, Shichao and Yuan, Ruifeng and Cao, Ziqiang and Li, Wenjie and Liu, Pengfei , editor =. Prompt. Findings of the. doi:10.18653/v1/2024.findings-acl.449 , urldate =

    doi:10.18653/v1/2024.findings-acl.449 2024

  71. [71]

    Proceedings of the 2024

    Tang, Xuemei and Deng, Zekun and Su, Qi and Yang, Hao and Wang, Jun , editor =. Proceedings of the 2024

    2024

  72. [72]

    Tang, Xuemei and Wang, Jun and Su, Qi and Huang, Chu-Ren and Gu, Jinghang , editor =. An. Proceedings of the 63rd. doi:10.18653/v1/2025.acl-short.38 , urldate =

    doi:10.18653/v1/2025.acl-short.38 2025

  73. [73]

    Language Model Collaboration for Relation Extraction from Classical

    Tang, Xuemei and Wang, Linxu and Wang, Jun , year = 2026, month = jan, journal =. Language Model Collaboration for Relation Extraction from Classical. doi:10.1016/j.ipm.2025.104286 , urldate =

    arXiv 2026

  74. [74]

    Tang, Yafen , year = 2013, journal =

    2013

  75. [75]

    Tang, Xuemei and Su, Qi , editor =. That. Proceedings of the 60th. doi:10.18653/v1/2022.acl-long.540 , urldate =

    work page doi:10.18653/v1/2022.acl-long.540 2022

  76. [76]

    A Temporal Fusion Model and Automated Annotation Framework for Computing

    Tan, Zelin and Tang, Chao and Zhang, Baoyu and Zhao, Xiaobing , year = 2026, month = may, journal =. A Temporal Fusion Model and Automated Annotation Framework for Computing. doi:10.1038/s40494-026-02643-z , urldate =

    work page doi:10.1038/s40494-026-02643-z 2026

  77. [77]

    Proceedings of the 57th

    Tenney, Ian and Das, Dipanjan and Pavlick, Ellie , editor =. Proceedings of the 57th. doi:10.18653/v1/P19-1452 , urldate =

    doi:10.18653/v1/p19-1452

  78. [78]

    Tong, Yu and Chen, Ge and Zheng, Guokai and Li, Rui and Dazhi, Jiang , editor =. When. Proceedings of the 2024. doi:10.18653/v1/2024.emnlp-main.593 , urldate =

    work page doi:10.18653/v1/2024.emnlp-main.593 2024

  79. [79]

    Attention Is

    Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N and ukasz Kaiser,. Attention Is. Advances in

  80. [80]

    Vilain, Marc and Burger, John and Aberdeen, John and Connolly, Dennis and Hirschman, Lynette , year = 1995, urldate =. A. Sixth

    1995

Showing first 80 references.