Pith. sign in

REVIEW 2 major objections 1 minor 37 references

Chengyu-Bench: Benchmarking Large Language Models for Chinese Idiom Understanding and Use

T0 review · 2 major / 1 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Large language models reliably tell whether a Chinese idiom is positive or negative but still fail to use idioms correctly in context, a new 2,937-example benchmark with three tasks shows.

desk verdict Chengyu-Bench is a solid new benchmark, but the open cloze exact-match scoring makes the headline 40% number a lower bound that needs re-validation. read the letter →

arxiv 2506.18105 v1 pith:4S7UTKYX submitted 2025-06-22 cs.CL

Yicheng Fu , Zhemin Huang , Liuxin Yang , Yumeng Lu , Zhongdongming Dai This is my paper
classification cs.CL
keywords Chineseidiomschengyubenchmarklargelanguagemodelsidiomcomprehensionopenclozetasksentimentpolaritycontextualappropriateness
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 claims that current large language models are competent at one superficial facet of Chinese idioms—their positive or negative connotation—but not at the deeper task of using idioms correctly in realistic contexts. To make this case, it introduces Chengyu-Bench, a human-verified benchmark of 2,937 examples covering 1,765 common idioms, with three tasks: labeling an idiom as positive or negative, judging whether a highlighted idiom fits its sentence, and generating five idioms to fill a blank in a long passage with no options. Evaluated models score over 95% on the sentiment task, around 85% on the appropriateness task, and about 40% top-1 on the open cloze task; error analysis attributes most mistakes to fundamental misunderstanding of idiom meanings. If the benchmark is right, multiple-choice idiom tests are flattering models, and real control of an idiom's cultural and contextual constraints remains a largely unsolved capability.

What carries the argument

The load-bearing object is the Chengyu-Bench dataset. Its construction pipeline samples naturally occurring texts (webpages, news, academic papers, essays, exam materials), extracts idioms and the sentences or paragraphs containing them, filters to idioms with adequate document frequency in a public Chinese lexicon, and manually labels polarity, appropriateness, and target completions. The decisive component is the Open Cloze subtask: it provides only a long passage with a #idiom# placeholder, demands five ranked candidate idioms, and averages about 600 context tokens per instance—nearly three times the length of the standard ChID cloze benchmark. This removes candidates and forces generative recall, which is what turns the benchmark from a recognition test into a genuine test of idiom use.

What would settle it

Ask a panel of native Chinese speakers to complete the Open Cloze passages without knowing the original idiom, and measure how often their top choice matches it (human top-1 agreement with the original). If human agreement with the original is close to the models' ~40%, then the benchmark mostly measures how predictable a single author's phrasing is, not general idiom competence, and the sentiment-versus-usage gap would need to be reinterpreted accordingly.

Watch

Extended reading notes

Core claim

The paper's central finding is a performance split: every tested LLM exceeds 95% accuracy on Evaluative Connotation, yet Appropriateness accuracy stays below 85% and Open Cloze top-1 accuracy is 40% or less, with the best open-weight model (DeepSeek-R1) reaching only 83.27% on Appropriateness and the best closed-source model (Gemini-2.5-Pro) reaching 40.05% top-1 on Open Cloze. Error analysis on the top Appropriateness performer shows 57.3% of its mistakes are meaning misinterpretations, followed by domain-adaptation errors. The authors interpret the pattern as evidence that LLMs can recognize the evaluative valence embedded in idioms—a relatively shallow signal—without mastering the semantic, collocational, and presuppositional knowledge that governs their appropriate use in extended discourse.

Load-bearing premise

The benchmark's Open Cloze task scores a prediction as correct only when it exactly matches the idiom that originally appeared in the passage, even though the paper acknowledges that multiple idioms can express similar meaning; if human-accepted alternative completions are common, the reported 40% top-1 accuracy understates true idiom-use ability.

Editorial extensions

If this is right

  • Sentiment classification of idioms is essentially saturated for current LLMs, so future Chinese idiom benchmarks should concentrate on usage tasks like appropriateness judgment and open generation.
  • Closed-form cloze benchmarks overstate idiom competence: models score far higher on ChID multiple-choice cloze than on the option-free open cloze, showing that selection among candidates is much easier than recall.
  • The errors that remain are not subtle: more than half of the best model's appropriateness mistakes come from basic meaning misinterpretation, which means idiom understanding—not just context modeling—is the bottleneck for LLMs.
  • For practical NLP, these results imply that LLM-generated Chinese text cannot be assumed to use idioms correctly, so idiom-misuse checking would be needed before high-stakes publication.
  • Best models on Appropriateness still score only around 83%, confirming that even state-of-the-art LLMs are far from fluent in contextual idiom usage.

Reading between the lines

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

  • The exact-match scoring on Open Cloze likely makes the reported 40% top-1 accuracy conservative: the paper itself notes that several idioms can convey similar nuances, so human raters might accept valid alternatives that are currently counted as errors.
  • The sentiment-vs-usage gap invites a causal probe: if a model that has never seen Chinese data still scores above 95% on Evaluative Connotation through character-level valence statistics, then the task is not measuring idiomatic knowledge at all; comparing such a control model would isolate the benchmark's signal.
  • The same three-task design could be exported to other culturally loaded formulaic language, such as Japanese yojijukugo or English idioms with opaque origins; a similar wide gap across languages would suggest that sentiment recognition and contextual usage are universally decoupled in LLMs.
  • The dominance of meaning misinterpretation in the error analysis suggests a concrete training intervention: idiom-focused pretraining objectives or retrieval-augmented reference to idiom dictionaries should improve Appropriateness more than scaling alone; this is a testable prediction of the paper's error taxonomy.
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.

Referee Report

2 major / 1 minor

Summary. The paper introduces Chengyu-Bench, a benchmark for Chinese idiom understanding and use with three tasks: Evaluative Connotation, Appropriateness, and Open Cloze. The benchmark is built from diverse sources (webpages, news, academic papers, essays, exam materials) and contains 2,937 human-verified examples covering 1,765 idioms. The authors evaluate eight LLMs and report above 95% accuracy on Evaluative Connotation, roughly 85% or less on Appropriateness (best 83.27%), and around 40% top-1 accuracy on Open Cloze for the best model. An error analysis of the Appropriateness task, using Gemini-2.5-Pro to label errors made by DeepSeek-R1, attributes most errors to meaning misinterpretation. The central claim is that LLMs reliably recognize idiom sentiment but still struggle with contextual and cultural nuances required for appropriate use.

Significance. If the measurements hold, Chengyu-Bench is a useful and well-motivated contribution: it addresses a real gap in existing Chinese idiom benchmarks by combining sentiment classification, misuse detection, and generative cloze filling in longer contexts. The dataset is constructed from external, human-annotated sources rather than fitted to model behavior, so there is little risk of circularity. The public release of data and code is a concrete strength that supports reproducibility. The qualitative finding that connotation accuracy far exceeds appropriateness and open-cloze accuracy is plausible and interesting. However, the quantitative magnitude of the open-cloze result, and to a lesser degree the appropriateness result, depends on scoring and annotation choices that currently are not fully validated.

major comments (2)
  1. [Section 3.2, Section 3.3] The paper repeatedly claims that the dataset is human-verified and human-annotated, but it reports no annotation protocol, no number of annotators, and no inter-annotator agreement. This is particularly important for the Appropriateness task, where labels require subjective judgments about whether an idiom's use is acceptable; Table 8's example involving '山高水低' shows that some benchmark labels are contestable. Without an agreement measure, the reliability of the 2,937 labels is not established. Please report IAA on a representative sample (e.g., Cohen's kappa or Krippendorff's alpha) and provide detailed annotation guidelines in the appendix or supplementary materials.
  2. [Section 3.2, Section 3.3] The negative examples for the Appropriateness task are drawn from exam materials and educational sites rather than from naturally occurring misuse. The paper motivates the task by saying such misuse is common among human writers, but exam materials are typically constructed to illustrate clear, textbook errors and may not reflect the distribution of real-world misuse. This limits the external validity of the reported ~85% appropriateness accuracy as a measure of detecting real misuse. Please either supplement with naturally occurring misuse (e.g., from learner corpora or web text, then human-verified) or provide evidence that the exam-sourced negatives are representative of natural misuse patterns.
minor comments (1)
  1. [Section 3.3, Table 4] The paper does not report a human performance baseline on the benchmark tasks. Human baselines would help calibrate whether ~40% top-1 accuracy is in fact low, and would strengthen the claim that the benchmark reflects real-world difficulty.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the benchmark is human-annotated from external corpora and all reported scores are independent measurements.

full rationale

The paper's derivation chain is a data-construction pipeline (Section 3.2): idioms are sampled from the Xinhua dictionary and filtered by document frequency; sentences and paragraphs are extracted from webpages, exam materials, news articles, academic papers, and essays; annotations are manually labeled and human-verified; models are then evaluated on fixed prompts. No parameter is fitted to model outputs, and no benchmark score is used to define the tasks or labels. The headline gap (Table 7: over 95% on Connotation, roughly 85% on Appropriateness, and roughly 40% on Open Cloze) is thus an independent empirical measurement, not a quantity forced by construction. The main validity caveat is the exact-match Open Cloze scoring, where Acc.@k counts whether the original idiom appears in the model's top-k predictions; if multiple idioms are acceptable, this may understate generative ability. That is a measurement-validity concern, not circularity, because the reference answer is the original corpus idiom rather than a model-dependent or fitted target. The error analysis uses Gemini-2.5-Pro to categorize DeepSeek-R1 errors without human validation, but this post-hoc categorization does not feed back into any benchmark score. Self-citations (Fu et al. 2024, 2025) appear only in the Limitations section as examples of compound AI systems and are not load-bearing. The Limitations section also explicitly acknowledges scope restrictions, including binary polarity, exclusion of non-chengyu forms, and the focus on standalone LLMs, so there is no hidden self-referential support. Verdict: no significant circularity.

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

The benchmark is a new dataset, not a new physical or mathematical entity. No free parameters are fitted to data; the main caveats are annotation assumptions and scoring choices, not fitted constants.

assumptions (4)
  • domain assumption Human annotations are treated as ground truth despite no reported inter-annotator agreement.
    Section 3.2 labels every example manually; if labels are idiosyncratic, all task scores inherit noise.
  • ad hoc to paper Each Open Cloze passage has a single correct idiom, the one originally in the source text.
    Section 3.2 replaces the target idiom with a placeholder, and Section 4 scores exact matches, while Section 3.1 acknowledges several idioms may be acceptable.
  • domain assumption Idiom sentiment is reducible to binary positive or negative for the selected items.
    Section 3.2 discards neutral cases; the paper itself acknowledges this simplification in the Limitations section.
  • domain assumption The Xinhua Dictionary and THUOCL document frequencies define common idiom validity.
    Section 3.2 filters idioms using this vocabulary and uses it for the Valid Idiom metric, so dictionary coverage determines what counts as a valid idiom.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Chengyu-Bench: Benchmarking Large Language Models for Chinese Idiom Understanding and Use." pith.science (2026). https://pith.science/paper/4S7UTKYX

@misc{pith2026250618105,
  author       = {Pith},
  title        = {Pith review of: Chengyu-Bench: Benchmarking Large Language Models for Chinese Idiom Understanding and Use},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4S7UTKYX}},
  note         = {Machine review of arXiv:2506.18105}
}
read the original abstract

Chinese idioms (Chengyu) are concise four-character expressions steeped in history and culture, whose literal translations often fail to capture their full meaning. This complexity makes them challenging for language models to interpret and use correctly. Existing benchmarks focus on narrow tasks - multiple-choice cloze tests, isolated translation, or simple paraphrasing. We introduce Chengyu-Bench, a comprehensive benchmark featuring three tasks: (1) Evaluative Connotation, classifying idioms as positive or negative; (2) Appropriateness, detecting incorrect idiom usage in context; and (3) Open Cloze, filling blanks in longer passages without options. Chengyu-Bench comprises 2,937 human-verified examples covering 1,765 common idioms sourced from diverse corpora. We evaluate leading LLMs and find they achieve over 95% accuracy on Evaluative Connotation, but only ~85% on Appropriateness and ~40% top-1 accuracy on Open Cloze. Error analysis reveals that most mistakes arise from fundamental misunderstandings of idiom meanings. Chengyu-Bench demonstrates that while LLMs can reliably gauge idiom sentiment, they still struggle to grasp the cultural and contextual nuances essential for proper usage. The benchmark and source code are available at: https://github.com/sofyc/ChengyuBench.

Figures

Figures reproduced from arXiv: 2506.18105 by the authors.

Figure 1
Figure 1. Subtask example. In the Evaluative Connotation subtask, the model must classify the sentiment polarity of a single idiom. In the Appropriateness subtask, it must decide whether the highlighted idiom fits the given context. In the Open Cloze subtask, it generates five idiom candidates, ranked by confidence, to complete the paragraph. Purple text highlights the idiom or placeholder in the prompt, and green text shows … view at source ↗
Figure 2
Figure 2. Overview of the benchmark generation pipeline. The process consists of four phases: (1) Sampling [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Distribution of error types made by Deepseek [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 18 canonical work pages

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Anthropic. 2024. https://www-cdn.anthropic.com/de8ba9b01c9ab7cbabf5c33b80b7bbc618857627/Model_Card_Claude_3.pdf The claude 3 model family: Opus, sonnet, haiku

    2024

  4. [4]

    Jiahuan Cao, Yang Liu, Yongxin Shi, Kai Ding, and Lianwen Jin. 2024. Wenmind: A comprehensive benchmark for evaluating large language models in chinese classical literature and language arts. Advances in Neural Information Processing Systems, 37:51358--51410

    work page 2024

  5. [5]

    DeepSeek-AI, Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Daya Guo, Dejian Yang, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, and 1 others. 2025. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437

    arXiv 2025

  6. [6]

    Yicheng Fu, Raviteja Anantha, and Jianpeng Cheng. 2024. Camphor: Collaborative agents for multi-input planning and high-order reasoning on device. arXiv preprint arXiv:2410.09407

    arXiv 2024

  7. [7]

    Yicheng Fu, Zikui Wang, Liuxin Yang, Meiqing Huo, and Zhongdongming Dai. 2025. Conquer: A framework for concept-based quiz generation. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 4: Student Research Workshop), pages 92--104

    work page 2025

  8. [8]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948

    arXiv 2025

Show all 37 references
  1. [9]

    Zishan Guo, Yufei Huang, and Deyi Xiong. 2024. Ctooleval: a chinese benchmark for llm-powered agent evaluation in real-world api interactions. In Findings of the Association for Computational Linguistics ACL 2024, pages 15711--15724

    2024

  2. [10]

    Shiyi Han, Yuhui Zhang, Yunshan Ma, Cunchao Tu, Zhipeng Guo, Zhiyuan Liu, and Maosong Sun. 2016. Thuocl: Tsinghua open chinese lexicon. Tsinghua University

    2016

  3. [11]

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2020. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300

    2020 arXiv

  4. [12]

    Yuzhen Huang, Yuzhuo Bai, Zhihao Zhu, Junlei Zhang, Jinghan Zhang, Tangjun Su, Junteng Liu, Chuancheng Lv, Yikai Zhang, Yao Fu, and 1 others. 2023. C-eval: A multi-level multi-discipline chinese evaluation suite for foundation models. Advances in Neural Information Processing ...

    2023

  5. [13]

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, and 1 others. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276

    2024 arXiv

  6. [14]

    Zhiying Jiang, Boliang Zhang, Lifu Huang, and Heng Ji. 2018. Chengyu cloze test. In Proceedings of the Thirteenth Workshop on Innovative Use of NLP for Building Educational Applications, pages 154--158

    2018

  7. [15]

    Jon Jonz. 1991. Cloze item types and second language comprehension. Language testing, 8(1):1--22

    1991

  8. [16]

    Hongmei Kang and Yang Yang. 2022. A study on english translation of chinese four-character idioms: Strategies and problems. Linguistics and Culture Review, 6(1):200--213

    2022

  9. [17]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K \"u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \"a schel, and 1 others. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural informati...

    2020

  10. [18]

    Haonan Li, Yixuan Zhang, Fajri Koto, Yifei Yang, Hai Zhao, Yeyun Gong, Nan Duan, and Timothy Baldwin. 2023. Cmmlu: Measuring massive multitask language understanding in chinese. arXiv preprint arXiv:2306.09212

    2023 arXiv

  11. [19]

    Shuang Li, Jiangjie Chen, Siyu Yuan, Xinyi Wu, Hao Yang, Shimin Tao, and Yanghua Xiao. 2024 a . Translate meanings, not just words: Idiomkb’s role in optimizing idiomatic translation with language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume...

    2024

  12. [20]

    Yuanchun Li, Hao Wen, Weijun Wang, Xiangyu Li, Yizhen Yuan, Guohong Liu, Jiacheng Liu, Wenxing Xu, Xiang Wang, Yi Sun, and 1 others. 2024 b . Personal llm agents: Insights and survey about the capability, efficiency and security. arXiv preprint arXiv:2401.05459

    2024 arXiv

  13. [21]

    Junwei Liao, Shuai Cheng, and Minghuan Tan. 2023. Text polishing with chinese idiom: Task, datasets and pre-trained baselines. ACM Transactions on Asian and Low-Resource Language Information Processing, 22(6):1--24

    2023

  14. [22]

    Xiao Liu, Xuanyu Lei, Shengyuan Wang, Yue Huang, Zhuoer Feng, Bosi Wen, Jiale Cheng, Pei Ke, Yifan Xu, Weng Lam Tam, and 1 others. 2023. Alignbench: Benchmarking chinese alignment of large language models. arXiv preprint arXiv:2311.18743

    2023 arXiv

  15. [23]

    Jipeng Qiang, Yang Li, Chaowei Zhang, Yun Li, Yi Zhu, Yunhao Yuan, and Xindong Wu. 2023. https://doi.org/10.1162/tacl_a_00572 Chinese idiom paraphrasing . Transactions of the Association for Computational Linguistics, 11:740--754

    2023 doi

  16. [25]

    Yutong Shao, Rico Sennrich, Bonnie Webber, and Federico Fancellu. 2017. Evaluating machine translation performance on chinese idioms with a blacklist method. arXiv preprint arXiv:1711.07646

    2017 arXiv

  17. [26]

    Minghuan TAN. 2022. Chinese idiom understanding with transformer-based pretrained language models

    2022

  18. [27]

    Minghuan Tan and Jing Jiang. 2021. Learning and evaluating chinese idiom embeddings. In Proceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP 2021), pages 1387--1396

    2021

  19. [28]

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, and 1 others. 2025. Gemini: A family of highly capable multimodal models. arXiv preprint arXiv:2312.11805

    2025 arXiv

  20. [29]

    Annie Tremblay. 2011. Proficiency assessment standards in second language acquisition research:“clozing” the gap. Studies in Second Language Acquisition, 33(3):339--372

    2011

  21. [30]

    Andrea W Wen-Yi, Unso Eun Seo Jo, and David Mimno. 2025. Do chinese models speak chinese languages? arXiv preprint arXiv:2504.00289

    2025 arXiv

  22. [31]

    Mingmin Wu, Yuxue Hu, Yongcheng Zhang, Zeng Zhi, Guixin Su, and Ying Sha. 2024. Mitigating idiom inconsistency: A multi-semantic contrastive learning method for chinese idiom reading comprehension. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pa...

    2024

  23. [32]

    Beilei Xiang, Changbing Yang, Yu Li, Alex Warstadt, and Katharina Kann. 2021. Climp: A benchmark for chinese language model evaluation. arXiv preprint arXiv:2101.11131

    2021 arXiv

  24. [33]

    Liang Xu, Hai Hu, Xuanwei Zhang, Lu Li, Chenjie Cao, Yudong Li, Yechen Xu, Kai Sun, Dian Yu, Cong Yu, and 1 others. 2020. Clue: A chinese language understanding evaluation benchmark. arXiv preprint arXiv:2004.05986

    2020 arXiv

  25. [34]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, and 1 others. 2024. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115

    2024 arXiv

  26. [35]

    Yu Yang, Stephen J Read, and Lynn C Miller. 2006. A taxonomy of situations from chinese idioms. Journal of Research in Personality, 40(5):750--778

    2006

  27. [36]

    Alex Young, Bei Chen, Chao Li, Chengen Huang, Ge Zhang, Guanwei Zhang, Guoyin Wang, Heng Li, Jiangcheng Zhu, Jianqun Chen, and 1 others. 2024. Yi: Open foundation models by 01. ai. arXiv preprint arXiv:2403.04652

    2024 arXiv

  28. [37]

    Wei Zeng, Xiaozhe Ren, Teng Su, Hui Wang, Yi Liao, Zhiwei Wang, Xin Jiang, ZhenZhang Yang, Kaisheng Wang, Xiaoda Zhang, Chen Li, Ziyan Gong, Yifan Yao, Xinjing Huang, Jun Wang, Jianfeng Yu, Qi Guo, Yue Yu, Yan Zhang, and 19 others. 2021. https://arxiv.org/abs/2104.12369 Pangu-...

    2021 arXiv

  29. [38]

    Chujie Zheng, Minlie Huang, and Aixin Sun. 2019. https://doi.org/10.18653/v1/P19-1075 C h ID : A large-scale C hinese ID iom dataset for cloze test . In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 778--787, Florence, Italy. As...

    2019 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.