InteChar: A Unified Oracle Bone Character List for Ancient Chinese Language Modeling

Daqian Shi; Hao Xu; Lida Shi; Ruihua Qi; Ting Wang; Xiaolei Diao; Zhihan Zhou

REVIEW 4 major objections 3 minor 1 cited by

InteChar: A Unified Oracle Bone Character List for Ancient Chinese Language Modeling

T0 review · 4 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper introduces InteChar, a unified character list that gives unencoded oracle bone characters a digital representation alongside traditional and modern Chinese, and reports that models trained with it on the new OracleCS corpus impro

desk verdict The InteChar abstract describes a plausibly useful encoding resource for oracle bone inscriptions, but the submitted full text is a different paper (KG-o1), so the actual work is not reviewable in this submission. read the letter →

arxiv 2508.15791 v1 pith:LI2O6LO5 submitted 2025-08-12 cs.CL cs.AI

Xiaolei Diao , Zhihan Zhou , Lida Shi , Ting Wang , Ruihua Qi , Hao Xu , Daqian Shi This is my paper

classification cs.CLcs.AI

keywords oracleboneinscriptionsancientChineseNLPcharacterencodinglanguagemodelingdigitizationhistoricaltextrepresentationcorpusconstructionLLMdataaugmentation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that a key reason ancient Chinese language models lag is that oracle bone characters, most of which have no standard encoding, cannot be digitized consistently enough to build usable text corpora. Its proposed fix is InteChar, a unified and extensible character list that slots oracle bone characters together with traditional and modern Chinese into one consistent token space. To test this, the authors construct OracleCS, a corpus of oracle bone inscriptions that combines expert-annotated samples with LLM-assisted data augmentation, and report that models trained with InteChar on OracleCS improve substantially on historical language understanding tasks. If the evaluation holds, the contribution is a reusable encoding standard and a training corpus that lower a barrier to computational study of early Chinese writing.

What carries the argument

InteChar itself: a character list (an encoding inventory) that gives unencoded oracle bone characters explicit representational slots alongside traditional and modern Chinese characters, so that a tokenizer can map an entire historical text into a single consistent vocabulary. The accompanying OracleCS corpus supplies the training signal, pairing expert-annotated examples with LLM-assisted augmentation to compensate for data scarcity.

What would settle it

Construct a held-out test set of oracle bone inscriptions transcribed solely by epigraphers who never saw the LLM-generated data, train a model on OracleCS, and compare performance against a model trained only on the expert-annotated portion. If the gains vanish on this external set, the improvement is an artifact of the augmentation distribution rather than a gain in understanding genuine historical texts.

Watch

Extended reading notes

Core claim

The paper claims that the absence of a comprehensive encoding scheme for oracle bone characters, rather than model architecture or raw data volume alone, is a primary bottleneck for ancient Chinese language modeling. InteChar is introduced as a unified, extensible character list that integrates unencoded oracle bone characters with traditional and modern Chinese, enabling historical texts to be digitized and represented consistently. To evaluate this, the authors construct OracleCS, a corpus centered on oracle bone inscriptions that combines expert-annotated samples with LLM-assisted data augmentation. Experiments reported in the paper show that models trained with InteChar on OracleCS achie

Load-bearing premise

The load-bearing premise is that the test data in the experiments is independent of the LLM-assisted augmentation used to build the training corpus, so the reported improvements measure genuine historical language understanding rather than memorization of the augmentation distribution.

Editorial extensions

If this is right

Ancient Chinese texts containing oracle bone characters can be digitized and processed with standard NLP tools, instead of remaining as images or hand-encoded fragments.
Language models trained on OracleCS with InteChar can perform historical language understanding tasks such as classification, dating, or provenance-related questions better than models without it.
The unified character list gives future ancient Chinese NLP datasets a common token space, avoiding duplicated ad-hoc character mappings.
The corpus construction strategy of expert annotation plus LLM-assisted augmentation offers a path to create training data for other low-resource historical scripts.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the independence of the test data is confirmed, the same recipe could extend to bronze inscriptions, bamboo slips, and other pre-modern Chinese forms: build an encoding list spanning the unencoded historical glyphs and modern characters, seed with expert-annotated text, and augment with an LLM.
A testable extension is to use the InteChar token space as an evaluation standard for comparing different ancient Chinese LMs, making model outputs on unencoded characters directly comparable across systems.
A caution: if the LLM-generated augmentation contains anachronistic or normalized readings of oracle bone characters, models may appear to improve on the augmented distribution while failing on raw expert transcriptions; an external held-out set of fully expert-annotated inscriptions would be the decisive check.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The submitted full text is a different paper — KG-o1, a knowledge-graph QA paper, arXiv:2508.15790. So the InteChar work can only be assessed from its abstract, and that is a serious submission defect. The editor should send this back to the authors to fix the file before any substantive review.

What the abstract describes is worth taking seriously. A unified, extensible character list that maps unencoded oracle bone characters onto traditional and modern Chinese fills a real gap: digital resources for oracle bone inscriptions are fragmented and lack a consistent encoding standard. If InteChar actually ships, it would be an enabling resource for ancient Chinese NLP and epigraphy. Combining expert-annotated samples with LLM-assisted augmentation is also a reasonable strategy for a low-resource domain, and the claim of downstream LM gains is the right kind of empirical test.

That said, the abstract alone gives no way to verify any of this. There are no numbers, baselines, error bars, or details about how the character list was constructed or validated. There is no comparison to existing encoding efforts, such as Unicode's oracle bone blocks or prior epigraphic databases, so the novelty claim is provisional. The biggest red flag is the evaluation. OracleCS combines expert annotations with LLM-assisted augmentation, and the same corpus is used both to train and to evaluate the models. If the test split shares distribution with the LLM-generated training data, the reported 'substantial improvements' could reflect memorization of the synthetic distribution rather than genuine historical understanding. The abstract gives no evidence that the test set is independent. This is not yet a proven flaw, but it is exactly the question the full paper must answer — ideally with an external test set of independently sourced inscriptions.

My recommendation: return the submission to the authors to correct the file mismatch. The abstract is promising enough that a corrected version would deserve a serious referee. This version does not, because the actual InteChar manuscript is absent. I would not cite it or bring it to reading group until the real paper appears.

Referee Report

4 major / 3 minor

Summary. The submission, identified as arXiv:2508.15791, presents an abstract for a paper called 'InteChar: A Unified Oracle Bone Character List for Ancient Chinese Language Modeling.' The abstract claims that InteChar integrates unencoded oracle bone characters with traditional and modern Chinese, that the authors construct an Oracle Corpus Set (OracleCS) combining expert-annotated samples with LLM-assisted data augmentation, and that models trained with InteChar on OracleCS achieve 'substantial improvements' on historical language understanding tasks. However, the full text supplied for review is not this paper: it is an entirely different manuscript, 'KG-o1: Enhancing Multi-hop Question Answering in Large Language Models via Knowledge Graph Integration,' with no mention of InteChar, OracleCS, oracle bone characters, or ancient Chinese NLP. Consequently, the technical content of the claimed InteChar paper cannot be inspected, and none of the abstract's claims are verifiable from the submitted manuscript.

Significance. If the abstract's claims were substantiated, the contribution could be valuable: a unified, extensible encoding list for unencoded oracle bone characters would directly address a recognized digitization bottleneck in historical Chinese NLP, and a carefully constructed corpus with expert-annotated and LLM-augmented samples could support future modeling. The significance cannot be assessed from the submitted manuscript, however, because the full text contains no description of InteChar's structure, no details of OracleCS construction, no evaluation protocol, no baselines, and no quantitative results. No code, data, or reproducibility artifacts are present in the manuscript. In its current form, the paper provides an abstract-only claim without supporting evidence.

major comments (4)

[Full Text] The submitted full text is a different paper, 'KG-o1: Enhancing Multi-hop Question Answering in Large Language Models via Knowledge Graph Integration,' with no content related to InteChar, OracleCS, oracle bone characters, ancient Chinese, or historical language models. The central claims of the abstract therefore have no inspectable derivation, no experimental section, and no supporting description. This is a load-bearing defect: the paper as submitted cannot be evaluated for correctness, reproducibility, or contribution to ancient Chinese NLP.
[Abstract] The abstract claims 'substantial improvements across various historical language understanding tasks' but reports no metrics, no baselines, no dataset statistics, no error bars, and no external validation. Even under the assumption that the correct full text exists, the abstract alone does not provide enough information to assess the magnitude or validity of the claimed improvements. The manuscript needs a full experimental section before any soundness judgment is possible.
[Abstract (OracleCS construction and evaluation)] The abstract states that OracleCS 'combines expert-annotated samples with LLM-assisted data augmentation.' This raises a specific circularity risk: if the LLM-augmented data is used to construct both training and test splits, or if the test set is generated by the same augmentation pipeline, then the reported gains may reflect memorization of the synthetic distribution rather than improved modeling of genuine historical texts. The manuscript does not state how test splits were isolated from the augmentation process, nor does it report evaluation on independently sourced oracle bone inscriptions. This is a load-bearing point for the headline claim of improved historical language understanding.
[Abstract (resource availability)] The abstract introduces InteChar and OracleCS as named resources but provides no definition, schema, statistics, or availability statement. Without the actual character list, corpus, or a description of the augmentation and expert-annotation protocols, the paper cannot be reproduced or independently verified. The current manuscript also does not state whether InteChar mappings and the OracleCS corpus will be released.

minor comments (3)

[Abstract] The abstract cites no related work on oracle bone character encoding, historical Chinese corpora, or ancient-script LM resources. It should position the contribution against existing attempts at oracle bone character digitization and existing historical Chinese datasets.
[Title/Abstract] The term 'extensible' for InteChar is asserted without explanation. The paper should specify the intended extension mechanism (e.g., Unicode private-use areas, user-defined character registry, versioning policy).
[General] If the full-text mismatch is a submission error, the correct manuscript must be supplied before review can proceed; the current submission cannot be treated as the claimed paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity in the supplied manuscript (KG-o1, arXiv:2508.15790); external benchmarks support the central claim.

full rationale

The supplied full text is the KG-o1 paper (arXiv:2508.15790), not the InteChar paper named in the header (arXiv:2508.15791); I flag this internal inconsistency, but it is a manuscript-integrity issue rather than a circularity step. The InteChar abstract alone contains no derivation chain, equations, or train/test split details, so no specific circular reduction can be exhibited for that portion. For the KG-o1 manuscript actually provided, the central claim is that integrating knowledge graphs into long-term reasoning training improves multi-hop question answering. The method is trained on the KG-MHQA SFT dataset, but it is evaluated on three external benchmarks (HotpotQA, 2WikiMultiHopQA, MINTQA) plus a MedQA transfer set. These external evaluations show consistent gains over the base models, so the improvement does not reduce by construction to the training distribution. The self-built KG-MHQA test set is generated by the same ChatGPT-4o pipeline as the training data, which is a potential data-quality concern, but the paper describes entity-based partitioning and manual annotation, and it is not the only evidence. There are no load-bearing self-citations, no imported uniqueness theorems, no ansatz smuggled in via citation, and no renaming of a known result as a new derivation. Therefore no specific circular step can be quoted, and the paper is self-contained with respect to external validation.

Assumptions & free parameters 0 free parameters · 3 assumptions · 2 invented entities

The central claim depends on the feasibility and correctness of the unified character list and the corpus; these are introduced by the paper without external evidence visible in the abstract. No free parameters are visible, but the evaluation assumptions carry the risk of circularity.

assumptions (3)

domain assumption Oracle bone characters can be consistently encoded in a single list with traditional and modern Chinese characters.
The entire InteChar contribution rests on the feasibility of a unified encoding that maps the ancient script onto the modern glyph system without losing meaning; this is asserted, not demonstrated in the abstract.
domain assumption The expert annotations and LLM-assisted augmentations in OracleCS accurately represent historical Chinese.
The corpus is claimed to combine expert-annotated samples with LLM-assisted augmentation; the abstract does not provide evidence that LLM-generated variations preserve historical linguistic authenticity.
domain assumption The evaluation tasks in OracleCS are valid measures of historical language understanding.
Reported improvements are relative to models trained on the same augmented corpus; without an externally validated benchmark, the evaluation's validity is assumed.

invented entities (2)

InteChar
purpose: Unified character list for digitizing oracle bone inscriptions alongside traditional and modern Chinese.
No external validation or usage outside the paper is shown in the abstract; its utility is the paper's claim.
OracleCS
purpose: Ancient Chinese corpus for training and evaluating language models on oracle bone texts.
The abstract does not indicate that the corpus is independently verified against real historical documents; it is constructed within the paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of InteChar: A Unified Oracle Bone Character List for Ancient Chinese Language Modeling." pith.science (2026). https://pith.science/paper/LI2O6LO5

@misc{pith2026250815791,
  author       = {Pith},
  title        = {Pith review of: InteChar: A Unified Oracle Bone Character List for Ancient Chinese Language Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LI2O6LO5}},
  note         = {Machine review of arXiv:2508.15791}
}

read the original abstract

Constructing historical language models (LMs) plays a crucial role in aiding archaeological provenance studies and understanding ancient cultures. However, existing resources present major challenges for training effective LMs on historical texts. First, the scarcity of historical language samples renders unsupervised learning approaches based on large text corpora highly inefficient, hindering effective pre-training. Moreover, due to the considerable temporal gap and complex evolution of ancient scripts, the absence of comprehensive character encoding schemes limits the digitization and computational processing of ancient texts, particularly in early Chinese writing. To address these challenges, we introduce InteChar, a unified and extensible character list that integrates unencoded oracle bone characters with traditional and modern Chinese. InteChar enables consistent digitization and representation of historical texts, providing a foundation for robust modeling of ancient scripts. To evaluate the effectiveness of InteChar, we construct the Oracle Corpus Set (OracleCS), an ancient Chinese corpus that combines expert-annotated samples with LLM-assisted data augmentation, centered on Chinese oracle bone inscriptions. Extensive experiments show that models trained with InteChar on OracleCS achieve substantial improvements across various historical language understanding tasks, confirming the effectiveness of our approach and establishing a solid foundation for future research in ancient Chinese NLP.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Do Language Models Agree with Human Perceptions of Suspense in Stories?
cs.CL 2025-08 unverdicted novelty 5.0 of 10

Language models can classify suspenseful stories but cannot reproduce human ratings of suspense magnitude or arc across story segments, and their judgments diverge from humans under text permutation.

Reference graph

Works this paper leans on

5 extracted references · 5 linked inside Pith · cited by 1 Pith paper

[2]

Guan, X.; Zhang, L

Deliberative Alignment: Reasoning Enables Safer Language Models.OpenAI Blog. Guan, X.; Zhang, L. L.; Liu, Y .; Shang, N.; Sun, Y .; Zhu, Y .; Yang, F.; and Yang, M. 2025. rStar-Math: Small LLMs Can Master Math Reasoning with Self-Evolved Deep Thinking. arXiv preprint arXiv:2501.04519. Guo, D.; Yang, D.; Zhang, H.; Song, J.; Zhang, R.; Xu, R.; Zhu, Q.; Ma,...

arXiv 2025
[4]

InProceedings of the 28th International Conference on Computational Linguistics, 6609–6625

Constructing A Multi-hop QA Dataset for Comprehen- sive Evaluation of Reasoning Steps. InProceedings of the 28th International Conference on Computational Linguistics, 6609–6625. Jaech, A.; Kalai, A.; Lerer, A.; Richardson, A.; El-Kishky, A.; Low, A.; Helyar, A.; Madry, A.; Beutel, A.; Carney, A.; et al. 2024. Openai o1 system card.arXiv preprint arXiv:24...

arXiv 2024
[5]

Shinichiro_Mik

KG-CoT: Chain-of-thought prompting of large lan- guage models over knowledge graphs for knowledge-aware question answering. InProceedings of the Thirty-Third Inter- national Joint Conference on Artificial Intelligence (IJCAI- 24), 6642–6650. International Joint Conferences on Artificial Intelligence. Zhao, Y .; Yin, H.; Zeng, B.; Wang, H.; Shi, T.; Lyu, C...

arXiv 2024
[2020]

arXiv preprint arXiv:2010.14701

Scaling laws for autoregressive generative modeling. arXiv preprint arXiv:2010.14701. Ho, X.; Nguyen, A.-K. D.; Sugawara, S.; and Aizawa, A

arXiv 2010
[2024]

Frisoni, G.; Cocchieri, A.; Presepi, A.; Moro, G.; and Meng, Z

Break the Chain: Large Language Models Can be Shortcut Reasoners.arXiv preprint arXiv:2406.06580. Frisoni, G.; Cocchieri, A.; Presepi, A.; Moro, G.; and Meng, Z. 2024. To Generate or to Retrieve? On the Effectiveness of Artificial Contexts for Medical Open-Domain Question Answering. InProceedings of the 62nd Annual Meeting of the Association for Computati...

arXiv 2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [2]

Guan, X.; Zhang, L

Deliberative Alignment: Reasoning Enables Safer Language Models.OpenAI Blog. Guan, X.; Zhang, L. L.; Liu, Y .; Shang, N.; Sun, Y .; Zhu, Y .; Yang, F.; and Yang, M. 2025. rStar-Math: Small LLMs Can Master Math Reasoning with Self-Evolved Deep Thinking. arXiv preprint arXiv:2501.04519. Guo, D.; Yang, D.; Zhang, H.; Song, J.; Zhang, R.; Xu, R.; Zhu, Q.; Ma,...

arXiv 2025

[2] [4]

InProceedings of the 28th International Conference on Computational Linguistics, 6609–6625

Constructing A Multi-hop QA Dataset for Comprehen- sive Evaluation of Reasoning Steps. InProceedings of the 28th International Conference on Computational Linguistics, 6609–6625. Jaech, A.; Kalai, A.; Lerer, A.; Richardson, A.; El-Kishky, A.; Low, A.; Helyar, A.; Madry, A.; Beutel, A.; Carney, A.; et al. 2024. Openai o1 system card.arXiv preprint arXiv:24...

arXiv 2024

[3] [5]

Shinichiro_Mik

KG-CoT: Chain-of-thought prompting of large lan- guage models over knowledge graphs for knowledge-aware question answering. InProceedings of the Thirty-Third Inter- national Joint Conference on Artificial Intelligence (IJCAI- 24), 6642–6650. International Joint Conferences on Artificial Intelligence. Zhao, Y .; Yin, H.; Zeng, B.; Wang, H.; Shi, T.; Lyu, C...

arXiv 2024

[4] [2020]

arXiv preprint arXiv:2010.14701

Scaling laws for autoregressive generative modeling. arXiv preprint arXiv:2010.14701. Ho, X.; Nguyen, A.-K. D.; Sugawara, S.; and Aizawa, A

arXiv 2010

[5] [2024]

Frisoni, G.; Cocchieri, A.; Presepi, A.; Moro, G.; and Meng, Z

Break the Chain: Large Language Models Can be Shortcut Reasoners.arXiv preprint arXiv:2406.06580. Frisoni, G.; Cocchieri, A.; Presepi, A.; Moro, G.; and Meng, Z. 2024. To Generate or to Retrieve? On the Effectiveness of Artificial Contexts for Medical Open-Domain Question Answering. InProceedings of the 62nd Annual Meeting of the Association for Computati...

arXiv 2024