AlphaOracle: Oracle bone script decipherment via human-workflow-inspired deep learning

Bang Li; Haisu Guan; Handong Zheng; Huanxin Yang; Jinpeng Wan; Kaile Zhang; Lianwen Jin; Pengjie Wang; Xiang Bai; Xingchen Liu

REVIEW 4 major objections 5 minor 13 references

AlphaOracle: Oracle bone script decipherment via human-workflow-inspired deep learning

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read AlphaOracle claims to be the first AI system to decipher oracle bone script by emulating the full expert workflow, and its evidence chain resolves the contested character '勞' as a toponym or clan name.

desk verdict Oversold 'Lao' resolution, but a solid engineering contribution that deserves refereeing once framing is corrected. read the letter →

arxiv 2607.17849 v1 pith:J2KT47OR submitted 2026-07-20 cs.HC cs.CLcs.CV

Yuliang Liu , Haisu Guan , Pengjie Wang , Xinyu Wang , Jinpeng Wan , Kaile Zhang , Handong Zheng , Xingchen Liu

show 6 more authors

Zhebin Kuang Huanxin Yang Bang Li Yongge Liu Lianwen Jin Xiang Bai

This is my paper

classification cs.HCcs.CLcs.CV

keywords oraclebonescriptdeciphermentpaleographydeeplearningmorphologicalanalysiscontextualalignmentphilologicalgroundingShangdynasty

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

This paper introduces AlphaOracle, a computational pipeline that tries to decipher unread oracle bone script by mirroring how human paleographers work: first analyze a glyph's shape and evolution, then check how it is used across thousands of inscriptions, then test proposed readings against ancient texts and modern scholarship. The authors claim this is the first AI system to attempt all three steps together rather than treating characters as isolated images for recognition. On withheld known characters used as a proxy for unknown glyphs, the system outperforms general language models, and in a study of 86 specialists it reduced reported analysis time by 64 percent, with about 79 percent rating it helpful or very helpful. The paper's showcase result is a disputed character, '勞', which AlphaOracle reads as a toponym or clan designation, a reading the authors say is supported by converging morphological, contextual, and philological evidence. If correct, the work would give paleographers a scalable, transparent assistant for the roughly 3,000 oracle bone characters that remain undeciphered.

What carries the argument

The load-bearing mechanism is the sequential integration of three evidence modules: morphological analysis (a confidence-weighted ensemble of networks modeling glyph evolution, radical decomposition, and cross-period variants), contextual alignment (a masked-language model over the digitized oracle-bone corpus plus modern-character semantic prediction), and philological grounding (retrieval over 25 transmitted texts and 23,755 scholarly studies, with retrieved passages weighted by relevance and source authority). The final report performs no additional adjudication; it assembles pre-computed scores from all three stages into candidate readings with confidence levels. That design is what make

What would settle it

Look up every attested occurrence of the glyph the paper reads as '勞' (the ten specimens it lists: 33473, 33535, 34656, 34732, 34781, 34800, 35554, 35560, 35561, and 35562) and check whether any appears in a syntactic position where a place-name reading is impossible; a single clear counterexample would falsify the toponym conclusion.

Watch

Extended reading notes

Core claim

AlphaOracle's central claim is that oracle-bone decipherment can be formalized as a sequential, evidence-producing pipeline rather than a single visual classification step. Morphological analysis generates ranked candidate readings using a confidence-weighted ensemble; contextual alignment re-ranks those candidates and independently predicts readings from masked-language modeling over a large corpus of digitized inscriptions; philological grounding scores candidates against retrieved passages from transmitted classics and modern scholarly literature. The pipeline produces a transparent report with explicit confidence weights, and the paper reports that across multiple test characters its rea

Load-bearing premise

The load-bearing premise is that a system trained on already-deciphered characters and validated against the same scholarly tradition it retrieves from can stand in for deciphering genuinely unknown glyphs.

Editorial extensions

If this is right

If reliable, the pipeline gives scholars a scalable way to generate candidate readings for the roughly 3,000 undeciphered oracle bone characters, with each reading backed by a citable evidence chain.
The '勞' resolution, if accepted, would shift interpretations of certain Shang divination records from event-completion statements to references to a place or clan, revising how those administrative and social contexts are read.
The reported 64 percent reduction in expert analysis time suggests scarce paleographic expertise could be redirected from manual corpus collation to adjudicating hard cases.
The same three-stage structure—form analysis, contextual modeling, philological retrieval—can in principle be carried to other partially documented ancient scripts that have a continuous graphic lineage and surviving textual tradition.
Explicit confidence weighting and low-confidence flags may make scholarly disagreement more visible and reduce the risk that a single visual resemblance is mistaken for a secure decipherment.

Reading between the lines

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

Beyond the paper: the approach's transferability to other undeciphered scripts is bounded by whether those scripts have a continuous visual lineage and a body of later literature to retrieve; for scripts with no living descendants, the philological-grounding stage would have little to search.
Beyond the paper: because the withheld-known-character test, the retrieved scholarship, and the validating experts all come from the same scholarly tradition, the strongest future test would be a pre-registered, blind decipherment of characters that currently lack any scholarly consensus.
Beyond the paper: the 64 percent time saving is based on experts' self-reported efficiency estimates rather than measured task durations; a controlled timing study would be a natural next check.
Beyond the paper: if the model were retrained with modern readings randomly permuted and still scored well above chance on the 88-character holdout, the simulated-decipherment result would indicate leakage or memorization rather than genuine decipherment.

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

Short version: the abstract promises more than the system actually shows. The genuinely new piece is an end-to-end workflow-inspired pipeline that chains morphological, contextual, and philological evidence into explicit, auditable decipherment reports, and the engineering behind it is substantial. But the showcase 'Lao' result is, on close reading, a weighted summary of existing scholarship rather than an independent decipherment. The quantitative evidence supports an assistive-tool claim, not the word 'resolves.'

What is new and done well: the corpus is the largest assembled for this problem (73,883 rubbings, 45,364 scans, 23,755 scholarly studies), the three-stage pipeline is described concretely enough to reimplement, and the authors ship code, pretrained weights, and data-processing scripts. The human evaluation is real: 86 specialists, reported 64% time reduction, 79% rating helpful, plus a blinded human-arena comparison where AlphaOracle received 45.2% of expert preferences over GPT-5, Gemini, and DeepSeek — a credible signal, even with the evaluators drawn from the same community.

Now the soft spots. Most importantly, the abstract's central claim — 'resolves the character 勞' — is not supported by the evidence in Appendix C. The conclusion is assembled from two of three morphological models favoring 勞, ten retrieved fragments all pointing to '在勞卜', and a philological grounding stage that retrieves and weights sixteen authoritative studies through Eq. (6). Several of those studies already argue the toponym reading. So the system is doing a well-organized retrieval and re-weighting of an existing scholarly position, not producing an independent inference. If the corpus had contained a different consensus, the output would mirror it. The authors acknowledge exactly this in Appendix I ('biases in the existing philological literature may shape retrieval results'), which makes the abstract's wording hard to defend. A corpus-exclusion test or an ablation without the philological module would be the natural control, and none is reported.

The quantitative evaluation is also thinner than the headline numbers suggest. Morphological top-1 accuracy on withheld characters is 23.1%; the LLM baselines are zero-shot, not domain-trained; there are no error bars anywhere; the expert time reduction is self-reported without controls; and the 95.2% reranking number applies to candidate sets the system itself generated. These are not fatal to the assistive-tool claim — the component tasks are hard and the system does beat baselines — but they should be presented with these caveats. The '22 disputed variant forms' are supported only by case-study prose, not by independent adjudication.

On the whole, this is a serious piece of work with transparent limitations stated in the supplement. The self-citations are to released datasets, so no concern there. Who should read it: computational paleographers, digital humanities researchers, and anyone building multi-stage evidence-chain systems for degraded historical scripts. I would send it to peer review — it deserves referee time — but I'd send it back for major framing changes: drop 'resolves' from the abstract, and require the authors to add a corpus-exclusion or ablation test that demonstrates the system can reach a reading the source corpus does not already endorse.

Referee Report

4 major / 5 minor

Summary. The paper introduces AlphaOracle, a multi-stage pipeline for oracle bone script (OBS) decipherment that combines rubbing parsing, morphological analysis (an ensemble of three networks), contextual alignment (masked-language modeling), and philological grounding (retrieval over classical texts and modern scholarship). The authors report top-1 morphological accuracy of 23.1% on 88 withheld character classes, contextual restoration accuracy of 56.1%, reranking accuracy of 95.2%, and BLEU/ROUGE/METEOR scores of 0.491/0.586/0.659 on an interpretation task, all exceeding zero-shot general LLM baselines. An expert study with 86 participants reports 78.7% approval and self-reported 64% time savings, and a separate blind human arena with 21 senior scholars gives AlphaOracle a 45.2% preference rate. The central showcase is the character '勞' (Lao), which AlphaOracle interprets as a toponym/clan designation, and the authors claim this 'resolves' a contested reading and offers 'concrete revisions' to Shang interpretations. The paper frames AlphaOracle as the first AI system to emulate the expert human workflow for OBS decipherment.

Significance. If the claims were fully supported, AlphaOracle would be a valuable assistive tool for OBS research: it assembles the largest digitized OBS corpus to date (73,883 transcriptions, 45,364 rubbings, 23,755 studies), provides transparent evidence chains, releases code and pretrained weights, and attempts a blinded expert evaluation. The component-wise metrics are plausible, and the system's ability to aggregate morphological, contextual, and philological evidence is a useful engineering contribution. However, the significance as stated in the abstract—'resolves' the character '勞' and 'offers concrete revisions' to Shang interpretations—is not supported by the evidence presented. The Lao conclusion is a weighted summary of existing scholarship retrieved via Eq. (6), not an independent decipherment, and the quantitative proxy tasks do not establish that the system can decipher truly unattested glyphs. With appropriate reframing and additional controls, the contribution is a solid assistive-tool paper, not a decipherment-resolution paper.

major comments (4)

[Appendix C, Eq. (6), Appendix I] The central claim—that AlphaOracle 'resolves' 勞 as a toponym/clan designation—rests on a retrieval-and-weighting process over a curated scholarly corpus. Appendix C shows that the philological grounding module retrieves 'the top sixteen relevant studies', several of which (e.g., Qiu Xigui's work) already argue the toponym reading. The score in Eq. (6) aggregates similarity and authority weights over these passages. If the corpus had contained a different dominant reading, the system would have retrieved that instead. The paper acknowledges this in Appendix I ('biases in the existing philological literature may shape retrieval results'), but the abstract and Section 3.3 present the outcome as a resolution. A corpus-exclusion test is needed: rerun the pipeline with the passages that already assert the toponym reading removed from the retrieval index, and show that the morphological and con
[Abstract, Section 3.2, Table 1] The abstract's claim that AlphaOracle 'resolves' contested characters and offers 'concrete revisions' is disproportionate to the reported quantitative results. The headline morphological accuracy is 23.1% top-1 on 88 withheld characters (Table 1), meaning roughly three-quarters of test characters are not correctly identified. The LLM baselines (GPT-5, Gemini 2.5 Pro, DeepSeek) are explicitly zero-shot and not domain-trained, so the large gaps in the interpretation task largely reflect the absence of domain-adapted baselines, not a demonstrated decipherment capability. The paper should provide (i) a domain-adapted baseline, e.g., fine-tune an LLM on the same training data, and (ii) significance tests or confidence intervals for the accuracy differences. Without these, the wording 'strongly agreed with expert interpretations' and 'resolves' overstates the evidence.
[Appendix D, Table S1, Section 3.3] The expert evaluation does not support the strength of the claims made from it. The main survey (N=86) includes 38.3% Computer Science/AI researchers (Table S1, Q1), who are not the primary target domain experts; the reported '78.7% approval' counts ratings 4 and 5, while 19.15% rated the system only 'somewhat helpful'. The 64% time reduction is self-reported (Q3), not measured against a control condition. The blind human arena (21 senior scholars, 210 judgments) gives AlphaOracle 45.2% preference vs. DeepSeek 34.8%—a modest gap, and the paper does not report inter-rater reliability or statistical significance. At minimum, the abstract and results should be reworded to reflect these magnitudes, and the time-saving claim should be labeled as a subjective estimate.
[Appendix D, Morphological analysis; Appendix I] The evaluation's core proxy assumption—that withholding known characters from training simulates deciphering truly unknown glyphs—needs a direct justification and a stress test. The test set (88 characters, 8,032 samples) is drawn from the same distribution as the training set, with clean, segmented glyph images. Appendix I itself notes that many real unknown characters are hapax, damaged, or unevenly represented. The paper should show performance broken down by frequency and by degree of physical damage, and should demonstrate that the system's confidence scores are calibrated on these harder cases. As written, the 23.1% top-1 accuracy on held-out known characters is not evidence that the system can decipher the ~3,000 unattested or poorly attested characters the introduction motivates.

minor comments (5)

[Table 1] The rows for Pix2Pix and CycleGAN report 'NAN' for glyph analysis accuracy, which is uninformative. Either provide the actual scores or state why these baselines are not applicable.
[Figure S1 caption] Typo: 'Contextual alignment metwork' should be 'network'.
[Appendix F, Morphological analysis models] The abbreviation 'OBSD' is used without definition; presumably it refers to one of the OBS datasets listed in Table S3. Please clarify.
[Appendix D, Human evaluation] The phrase 'Erta Essence Interpretations' appears to be a misnomer; likely 'Erta' should be 'Erta' or another transliteration. Also, the parenthetical reflection that reference transcriptions contain errors is important but not incorporated into the reported metrics; consider excluding those samples or reporting a sensitivity analysis.
[Appendix I] The limitation section is candid and should be retained, but its content directly contradicts the framing in the abstract. Please align the abstract and public summary with the limitations stated here.

Circularity Check

1 steps flagged · score 6.0 of 10

The 'Lao' 'resolution' is a weighted retrieval of existing scholarship (Eq. 6), not an independent decipherment; the abstract's strongest claim reduces to its corpus input.

fitted input called prediction [Abstract; Section 3.3; Appendix C (Figure S2D); Eq. (6)]
"AlphaOracle resolves the character “勞” as a toponymic or clan designation, offering concrete revisions to Shang administrative and social interpretations. ... The philological grounding module ... retrieving the top sixteen relevant studies from the comprehensive literature database. ... The retrieved literature suggests that “勞” in oracle bone inscriptions primarily functioned as a toponym."

The headline 'resolution' is the output of Eq. (6), a weighted aggregation of retrieved passages (sim, alpha, auth) from the curated scholarly corpus. Those passages include Guwenzi Gulin and Qiu Xigui's work, which already assert the toponym reading and that the graph cannot simply be identified as '卒'. The system does not independently infer the reading; it re-weights the same scholarly consensus that the abstract presents as a concrete revision. Appendix I concedes that 'biases in the existing philological literature may shape retrieval results' and that outputs are 'hypotheses for expert adjudication rather than definitive conclusions,' yet the abstract calls the retrieval result a 'resolution.' No corpus-exclusion test or ablation removing philological grounding is reported, so the ce

full rationale

The paper has substantial non-circular evaluation: the morphological analysis is tested on 88 held-out character categories (TRAIN 1577, TEST 88), contextual restoration is measured on masked characters, and the Human Arena is a blind preference test by 21 senior scholars. These give independent signal and the self-citations to earlier datasets (HUST-OBC, EVOBC) are not load-bearing. However, the abstract's strongest claim—that AlphaOracle 'resolves' the contested character Lao and offers 'concrete revisions'—is not supported by an independent inference for that case. Appendix C shows the Lao conclusion emerges from retrieving ten inscriptional parallels and sixteen scholarly studies, several of which already argue the toponym reading; Eq. (6) simply aggregates retrieved passages by relevance and authority. The paper itself admits in Appendix I that literature bias can shape retrieval and that outputs should be treated as hypotheses, not definitive conclusions. Because no control is reported to show the Lao result is not merely an artifact of the retrieval corpus, the central showcase reduces by construction to its scholarly input. This is a partial circularity: the framework and quantitative evaluations are independent, but the marquee 'resolution' is a weighted summary of existing scholarship presented as a discovery.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim does not rest on new physical entities or mathematical postulates. It rests on learned weights, curation choices, and domain assumptions about script evolution and the authority of the scholarly corpus. The most consequential free parameter is the philological weighting scheme, since the 'Lao' resolution is essentially a retrieval and re-weighting of existing scholarly arguments.

free parameters (4)

Morphological ensemble weights w_k = learned, not reported
Eq. (2) combines three expert models with learned weights; fitted on validation data and directly determines candidate ranking for decipherment.
Edge-loss weight in reading-order loss = 2
Eq. (1) manually assigns double weight to adjacency prediction; affects sentence reconstruction and therefore all downstream contextual evidence.
Masked-language-modeling hyperparameters = 15% mask, 80/10/10 split
Standard but chosen values; they determine contextual restoration accuracy and the BERT-style re-ranking results.
Authority/relevance weights in philological score = not specified
Eq. (6) uses auth(T) and alpha_c,q to weight retrieved passages, but their values and tuning are not reported; these weights shape the final 'Lao' evidence chain.

assumptions (4)

domain assumption Chinese character morphology preserves semantic content across script periods
Stated in Section 2.1 ('graphic form often encodes semantic content'); underpins the glyph and radical evolution networks.
ad hoc to paper Withholding known characters from training is a valid proxy for deciphering unknown characters
Evaluation protocol in Section 3.1 and Appendix D assumes a model trained on held-out known characters transfers to truly unattested glyphs; no direct evidence supports this transfer.
domain assumption Expert-curated transcriptions and modern scholarly references used for training and retrieval are correct
All supervised components and the 'Lao' conclusion depend on the correctness of the curated corpus; Appendix I acknowledges corpus bias and expert disagreement.
domain assumption Later transmitted texts can validate earlier oracle bone readings
The philological grounding module in Section 2.3 assumes diachronic continuity between oracle bone script and bronze/seal/classical texts.

how reviews work

0 comments

Cite this review

Pith. "Pith review of AlphaOracle: Oracle bone script decipherment via human-workflow-inspired deep learning." pith.science (2026). https://pith.science/paper/J2KT47OR

@misc{pith2026260717849,
  author       = {Pith},
  title        = {Pith review of: AlphaOracle: Oracle bone script decipherment via human-workflow-inspired deep learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J2KT47OR}},
  note         = {Machine review of arXiv:2607.17849}
}

read the original abstract

Approximately 3,000 of the 4,500 oracle bone script (OBS) characters remain undeciphered due to fragmentary inscriptions and sparse evidence. Current AI approaches fail to replicate expert workflows that integrate form analysis, contextual semantics, and philological reasoning. We introduce AlphaOracle, a human-workflow-inspired framework that systematizes OBS decipherment using the largest digitized corpus to date. Its multi-stage pipeline comprises: (i) rubbing parsing; (ii) radical-based morphological analysis with diachronic modeling; (iii) contextual retrieval with semantic alignment; and (iv) philological validation against classical sources. Each stage generates explicit, confidence-weighted evidence chains, culminating in interpretable reports for scholarly verification. Across multiple test characters, AlphaOracle's readings strongly agreed with expert interpretations. In a study of 86 domain specialists, it reduced analysis time by 64% and 79% of participants rated it highly useful. Notably, AlphaOracle resolves the character "Lao" as a toponymic or clan designation, offering concrete revisions to Shang administrative and social interpretations. These results suggest that computational methods aligned with philological practice can facilitate OBS research and provide a conceptual reference for studies of other undeciphered scripts.

Figures

Figures reproduced from arXiv: 2607.17849 by the authors.

**Figure 2.** Overview of character evolution and decipherment workflow [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

13 extracted references · 1 canonical work pages

[1]

Carion N., Massa F., Synnaeve G. et al. (2020). End-to-end object detection with transformers. In Proc. Eur. Conf. Comput. Vis., pp. 213–229. DOI:10.1007/978-3-030-58452-8_13

doi:10.1007/978-3-030-58452-8_13 2020
[2]

Liu Z., Mao H., Wu C.Y. et al. (2022). A convnet for the 2020s. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., pp. 11976–11986. DOI:10.1109/CVPR52688.2022.01167

arXiv 2022
[3]

and Nichol A

Dhariwal P. and Nichol A. (2021). Diffusion models beat gans on image synthesis.Adv. Neural Inf. Process. Syst.pp. 8780–8794. DOI:10.5555/3540261.3540933

arXiv 2021
[4]

Devlin J., Chang M.W., Lee K. et al. (2019). Bert: Pre-training of deep bidirectional transformers for language understanding. In Proc. Conf. North Am. Chapter Assoc. Comput. Linguist., pp. 4171–4186. DOI:10.18653/v1/N19-1423

doi:10.18653/v1/n19-1423 2019
[5]

Vaswani A., Shazeer N., Parmar N. et al. (2017). Attention is all you need. Guyon, I., von Luxburg, U., Bengio, S. et al. (eds.). Adv. Neural Inf. Process. Syst., pp. 5998–6008. DOI:10.5555/3295222. 3295349

doi:10.5555/3295222 2017
[6]

Papineni K., Roukos S., Ward T. et al. (2002). Bleu: a method for automatic evaluation of machine translation. In Proc. Annu. Meet. Assoc. Comput. Linguist., pp. 311–318. DOI:10.3115/1073083. 1073135

doi:10.3115/1073083 2002
[7]

DOI:10.48550/arXiv.2506.05176

ZhangY.,LiM.,LongD.etal.(2025).Qwen3embedding: Advancingtextembeddingandreranking through foundation models.arXiv. DOI:10.48550/arXiv.2506.05176

arXiv doi:10.48550/arxiv.2506.05176 2025
[8]

Oquab M., Darcet T., Moutakanni T. et al. (2023). Dinov2: Learning robust visual features without supervision.arXiv. DOI:10.48550/arXiv.2304.07193

arXiv doi:10.48550/arxiv.2304.07193 2023

Show all 13 references

[9]

DOI:10.48550/arXiv.2506.05218

LiZ.,LiuY.,LiuQ.etal.(2025).Monkeyocr: Documentparsingwithastructure-recognition-relation triplet paradigm.arXiv. DOI:10.48550/arXiv.2506.05218

2025 doi
[10]

Cui C., Sun T., Lin M. et al. (2025). Paddleocr 3.0 technical report.arXiv. DOI:10.48550/arXiv.250 7.05595

2025 doi
[11]

Li B., Luo D., Liang Y. et al. (2024). Oracle bone inscriptions multi-modal dataset.arXiv. DOI:10.48550/arXiv.2407.03900

2024 doi arXiv
[12]

Wang P., Zhang K., Wang X. et al. (2024). An open dataset for oracle bone character recognition and decipherment.Sci. Data11:976. DOI:10.1038/s41597-024-03807-x

2024 doi
[13]

Guan H., Wan J., Liu Y. et al. (2024). An open dataset for the evolution of oracle bone characters: Evobc.arXiv. DOI:10.48550/arXiv.2401.12467 31

2024 doi arXiv

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

Carion N., Massa F., Synnaeve G. et al. (2020). End-to-end object detection with transformers. In Proc. Eur. Conf. Comput. Vis., pp. 213–229. DOI:10.1007/978-3-030-58452-8_13

doi:10.1007/978-3-030-58452-8_13 2020

[2] [2]

Liu Z., Mao H., Wu C.Y. et al. (2022). A convnet for the 2020s. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., pp. 11976–11986. DOI:10.1109/CVPR52688.2022.01167

arXiv 2022

[3] [3]

and Nichol A

Dhariwal P. and Nichol A. (2021). Diffusion models beat gans on image synthesis.Adv. Neural Inf. Process. Syst.pp. 8780–8794. DOI:10.5555/3540261.3540933

arXiv 2021

[4] [4]

Devlin J., Chang M.W., Lee K. et al. (2019). Bert: Pre-training of deep bidirectional transformers for language understanding. In Proc. Conf. North Am. Chapter Assoc. Comput. Linguist., pp. 4171–4186. DOI:10.18653/v1/N19-1423

doi:10.18653/v1/n19-1423 2019

[5] [5]

Vaswani A., Shazeer N., Parmar N. et al. (2017). Attention is all you need. Guyon, I., von Luxburg, U., Bengio, S. et al. (eds.). Adv. Neural Inf. Process. Syst., pp. 5998–6008. DOI:10.5555/3295222. 3295349

doi:10.5555/3295222 2017

[6] [6]

Papineni K., Roukos S., Ward T. et al. (2002). Bleu: a method for automatic evaluation of machine translation. In Proc. Annu. Meet. Assoc. Comput. Linguist., pp. 311–318. DOI:10.3115/1073083. 1073135

doi:10.3115/1073083 2002

[7] [7]

DOI:10.48550/arXiv.2506.05176

ZhangY.,LiM.,LongD.etal.(2025).Qwen3embedding: Advancingtextembeddingandreranking through foundation models.arXiv. DOI:10.48550/arXiv.2506.05176

arXiv doi:10.48550/arxiv.2506.05176 2025

[8] [8]

Oquab M., Darcet T., Moutakanni T. et al. (2023). Dinov2: Learning robust visual features without supervision.arXiv. DOI:10.48550/arXiv.2304.07193

arXiv doi:10.48550/arxiv.2304.07193 2023