Pith. sign in

REVIEW 3 major objections 5 minor 47 references

Read it in Two Steps: Translating Extremely Low-Resource Languages with Code-Augmented Grammar Books

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Grammar rules written as code functions beat full grammar books for low-resource translation

desk verdict A useful dataset and a plausible two-step framing, but the headline 'code format works' claim is confounded with added verbosity, and the abstract mislabels the reported metric. read the letter →

arxiv 2506.01796 v1 pith:24I54TBJ submitted 2025-06-02 cs.CL

Chen Zhang , Jiuheng Lin , Xiao Liu , Zekai Zhang , Yansong Feng This is my paper
classification cs.CL
keywords extremelylow-resourcemachinetranslationgrammarbookslargelanguagemodelscode-formattedrulesruleretrievalapplicationZhuangin-contextlearning
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 argues that using a grammar book for machine translation should be thought of as two separate skills—finding the rule that applies to a sentence, then carrying that rule out—and that the first is the main obstacle for large language models. To test this, the authors build a modular dataset of 109 Zhuang grammar rules paired with 608 test sentences plus lexicons, so that lexical knowledge is not confused with grammar understanding. They find that giving the model only the required rule helps far more than feeding it the whole book, and that rewriting grammar rules as pseudo-code functions improves both retrieval and application. The full recipe—code-formatted rules retrieved one by one—outperforms the whole-book approach by up to 13.1% in chrF++. If correct, this gives a practical way to make grammar resources usable for extremely low-resource translation without training data.

What carries the argument

The load-bearing device is the code-formatted grammar rule: each rule is converted into a pseudo-code function whose docstring states the rule and its steps and whose body simulates the operation—reordering words, adding affixes, choosing between alternatives with if/else, concatenating phrase pieces. A companion strategy, rule-by-rule retrieval, feeds the model one candidate rule and the sentence at a time and asks for a binary relevance judgment, instead of making the model search the whole book. Together they convert a long-context text problem into short procedural checks plus function-style application, which the authors argue matches how large language models reason with code.

What would settle it

Run the same Zhuang test set with three rule formats—code functions, plain text, and plain text expanded with the same numbered steps and comparable length; if the expanded text matches the code rule results on retrieval recall and chrF++, the paper's attribution of the gain to code format is refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that grammar-based translation by large language models is bottlenecked by rule retrieval, not by rule application, and that representing grammar rules as code functions relieves both bottlenecks. The authors report that converting textual rules to Python-style pseudo-code raises retrieval recall by 8.8% on average, improves rule application by an average of 8.5% chrF++, and that combining code rules with a rule-by-rule relevance check surpasses the standard 'give the model the entire grammar book' baseline by up to 13.1% chrF++. They also show the code-rule advantage on a second language, Kalamang, and on multi-rule sentences, where composing two rule functions helps more than providing the rules as prose.

Load-bearing premise

The argument assumes that the code format itself—not the extra step-by-step detail and the roughly sevenfold longer text that comes with it—is what improves retrieval and translation, since the paper does not test equally detailed plain-text rules.

Editorial extensions

If this is right

  • The main payoff of future grammar-book methods should go into retrieval: finding the right rule, not explaining it, is what currently caps translation quality.
  • Grammar resources written once as code rules can be reused across models and languages: the same code-rule advantage appears on Zhuang and Kalamang.
  • For complex rules with several actions, code format helps most; when parallel examples are already provided, the format matters less.
  • Multiple rules can be combined by merging their code bodies, and letting the model write the merged function works best among the tested strategies.
  • Synthetic interlinear glosses do not add value on top of code rules, so code rules may replace a costly annotation step.

Reading between the lines

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

  • An untested possibility is that the gain is not the code format but the extra procedural detail: code rules in the paper are roughly seven times longer than the text rules and contain explicit step-by-step comments, so a matched plain-text control would be needed to isolate the format effect.
  • If the format effect is real, grammar books for many low-resource languages could be converted to code once and served through retrieval, turning static books into executable language resources; this is testable on any language with a published grammar.
  • The two-step decomposition suggests a practical engineering recipe: a cheap classifier over rules plus a code-augmented translator may outperform a single end-to-end prompt, even for stronger models.
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

3 major / 5 minor

Summary. The paper studies how large language models use grammar books for extremely low-resource machine translation, focusing on Zhuang and Kalamang. It introduces ZhuangRules, a modular dataset of 109 Zhuang grammar rules paired with 608 parallel examples and lexicons, and decomposes grammar-book-based translation into two steps: grammar rule retrieval and rule application. Through pilot studies and probing experiments, the authors find that retrieval is a major bottleneck and that LLMs struggle to apply complex textual rules. They propose converting grammar rules into code-format functions using GPT-4o and a rule-by-rule retrieval strategy that checks each rule's relevance individually. Experiments across three LLMs and two languages show that code rules improve retrieval recall and application quality, and that combining code rules with rule-by-rule retrieval yields a headline improvement of up to 13.1% in translation quality over a full-book textual baseline. The paper also reports a smaller multi-rule experiment and validates rule application on the MTOB Kalamang benchmark.

Significance. If the reported improvements hold, the paper makes a practical contribution to a real problem: using grammar books for extremely low-resource machine translation. The strengths are the controlled ZhuangRules dataset with public data and code, the decomposition of grammar-based MT into retrieval and application, the use of three open-source LLMs and two typologically distinct languages, and the external validation on MTOB. The rule-by-rule retrieval strategy is simple and effective, and the observation that code-format rules improve both retrieval and application is potentially useful to practitioners. However, the central mechanism claim—that the gains come specifically from representing rules as code rather than from the additional procedural detail those representations contain—is not convincingly established, and the headline metric and baseline are misreported. These issues require revision before the paper's conclusions can be accepted at face value.

major comments (3)
  1. [Abstract; §4.4, Table 6] The headline number is misreported. The abstract claims a 13.1% BLEU improvement, but §4.4 reports a 13.1% increase in chrF++, and Table 6 reports chrF++ values. Moreover, the 13.1% figure corresponds to comparing the RULE-BY-RULE(code) condition with Code Rule application (55.7 chrF++ for za2zh) to the w/o Retrieval condition with Text Rule application (42.6), not to the stated baseline of feeding the whole grammar book (FULL-BOOK Text Rule, 43.1 for za2zh and 66.5 for zh2za). The maximum differences against the stated full-book baseline are 12.6 and 7.8 chrF++ for the two directions. Please restate the headline with the correct metric and the correct baseline, and ensure the abstract matches the experimental section.
  2. [§3.2, §3.3; Table 8; Tables 11–12] The attribution of the observed gains to 'code format' is confounded with added procedural detail and instruction verbosity. Code rules total 32,231 tokens versus 4,386 for textual rules (a factor of roughly 7.3, Table 8), and the examples in Tables 11–12 contain a human-written docstring with numbered step-by-step instructions plus a pseudocode function with explicit output templates, whereas textual rules are single descriptive sentences averaging 57.1 Chinese characters. The paper does not test a text-only condition with equally detailed step-by-step instructions or output templates. The improvements in Tables 1, 2, and 6 are therefore equally consistent with increased explicitness—a known driver of LLM instruction-following—as with the claimed 'code' mechanism. Please add a text control matched for length and instruction detail, or substantially soften the mechanistic conclusions and frame the contribution as a practical representation choice.
  3. [§3.2] Only 10 of the 109 code rules were manually quality-checked, and the paper states that all sampled rules were complete and accurate except one that omitted minor information. Because the code rules are a core resource for both retrieval and application, a 10% sample is too small to rule out systematic conversion errors that could differentially affect the comparison against textual rules. Please provide a larger random sample (ideally a full audit) and release the conversion prompts and the complete generated code rule set so that the quality and reproducibility of this resource can be independently verified.
minor comments (5)
  1. [§4.4] The sentence reporting the 13.1% increase should specify the exact baseline condition and metric, since the number is not currently reproducible from Table 6 as stated.
  2. [Appendix D.3] The dataset is referred to as ZHUANGBENCH in Section D.3; it should be ZHUANGRULES for consistency with the rest of the paper.
  3. [Tables 2 and 3] Several numeric entries are missing spacing, e.g., '78.9/90.463.9' in Table 2 and '43.867.3' in Table 3; these should be corrected.
  4. [Limitations] The Limitations section discusses the higher token count of code rules only as an efficiency issue; it should also acknowledge the potential confound between code format and added procedural detail, which is the main interpretive risk for the paper's central claim.
  5. [Appendix B.2] The IGT quality check is based on 20 generated samples and reports 72% morpheme-level correctness; the paper should report how many Zhuang sentences contain at least one wrong gloss and whether the negative interaction between IGTs and rules is robust to this noise.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the main claims are established by direct held-out experiments and external MTOB validation.

full rationale

The paper's central claims—that rule retrieval is a bottleneck, that code-format rules improve retrieval and application, and that the combined pipeline improves translation—are supported by direct experiments on test instances with gold rule annotations, not by fitting a parameter and then predicting the same quantity. ZhuangRules is a dataset collected from published Zhuang grammar books with 608 parallel examples; gold rules and lexicons are test labels, not fitted outputs. Code rules are generated by GPT-4o from the textual rules and evaluated on the same held-out test sentences; no equation or fitted parameter reduces to the reported chrF++/BLEU values. The MTOB benchmark provides external Kalamang validation, making the result independent of the authors' own dataset. Self-citations (e.g., Zhang et al. 2024a for Zhuang's low-resource status, Liu et al. 2023 for code-reasoning advantages) are background support, but the paper's own baselines (e.g., No Rule with lexicon near zero BLEU in Table 2) and the code-vs-text comparisons in Tables 2 and 3 carry the argument; these citations are not load-bearing. The abstract's '13.1% BLEU improvement' is inconsistent with Section 4.4, which reports 'up to 13.1% increase in chrF++', and the code-rule gains may be confounded with added procedural detail and longer prompts (Table 8), but these are attribution/correctness concerns, not circularity.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

No fitted parameters; only hand-chosen hyperparameters. The central claims rest on dataset annotation correctness and faithful rule conversion, both only lightly validated.

free parameters (2)
  • ICL exemplars for code conversion = 5
    Chosen by hand; no sweep reported.
  • Parallel examples in application prompts = 2
    Randomly sampled; fixed at two per prompt.
assumptions (4)
  • domain assumption Each ZhuangRules test instance requires exactly one grammar rule and is correctly labeled with it
    Section 2 asserts this without an inter-annotator agreement or validation protocol; retrieval recall and the main pipeline depend on label correctness.
  • domain assumption GPT-4o-generated code rules faithfully represent the original textual rules
    Only 10 of 109 rules were manually checked (Section 3.2); the rest are assumed accurate.
  • domain assumption The provided lexicons cover all lexical items in test sentences
    Section 2.1 says the lexicon covers all relevant words, which isolates grammar from vocabulary knowledge.
  • domain assumption Zhuang and Kalamang are unknown to the evaluated LLMs
    Appendix A and prior work claim near-zero model proficiency; if models had some knowledge, the controlled comparison would be confounded.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Read it in Two Steps: Translating Extremely Low-Resource Languages with Code-Augmented Grammar Books." pith.science (2026). https://pith.science/paper/24I54TBJ

@misc{pith2026250601796,
  author       = {Pith},
  title        = {Pith review of: Read it in Two Steps: Translating Extremely Low-Resource Languages with Code-Augmented Grammar Books},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/24I54TBJ}},
  note         = {Machine review of arXiv:2506.01796}
}
read the original abstract

While large language models (LLMs) have shown promise in translating extremely low-resource languages using resources like dictionaries, the effectiveness of grammar books remains debated. This paper investigates the role of grammar books in translating extremely low-resource languages by decomposing it into two key steps: grammar rule retrieval and application. To facilitate the study, we introduce ZhuangRules, a modularized dataset of grammar rules and their corresponding test sentences. Our analysis reveals that rule retrieval constitutes a primary bottleneck in grammar-based translation. Moreover, although LLMs can apply simple rules for translation when explicitly provided, they encounter difficulties in handling more complex rules. To address these challenges, we propose representing grammar rules as code functions, considering their similarities in structure and the benefit of code in facilitating LLM reasoning. Our experiments show that using code rules significantly boosts both rule retrieval and application, ultimately resulting in a 13.1% BLEU improvement in translation.

Figures

Figures reproduced from arXiv: 2506.01796 by the authors.

Figure 1
Figure 1. An illustration of LLMs using grammar rules [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Change in translation performance of pro [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. An illustration of three strategies for combin [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: A simplified extract from the Zhuang grammar book. The watko ‘here’ (left), wane ‘this’ (middle) and wane [PITH_FULL_IMAGE:figures/full_fig_p019_4.png]
Figure 6
Figure 6. Figure 6: An testing instance that requires multiple grammar rules. To construct this, we modify sentences governed [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: Performance of BLEU by rule number (Qwen-2.5-72B-Instruct, Zhuang [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 12 canonical work pages

  1. [1]

    Seth Aycock, David Stap, Di Wu, Christof Monz, and Khalil Sima'an. 2024. Can llms really learn to translate a low-resource language from one grammar book? arXiv preprint arXiv:2409.19151

    arXiv 2024

  2. [2]

    Yushi Bai, Shangqing Tu, Jiajie Zhang, Hao Peng, Xiaozhi Wang, Xin Lv, Shulin Cao, Jiazheng Xu, Lei Hou, Yuxiao Dong, et al. 2024. Longbench v2: Towards deeper understanding and reasoning on realistic long-context multitasks. arXiv preprint arXiv:2412.15204

    arXiv 2024

  3. [3]

    Ben Bogin, Shivanshu Gupta, Peter Clark, and Ashish Sabharwal. 2024. https://doi.org/10.18653/v1/2024.naacl-long.279 Leveraging code to improve in-context learning for semantic parsing . In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers),...

    doi:10.18653/v1/2024.naacl-long.279 2024

  4. [4]

    Hyungjoo Chae, Yeonghyeon Kim, Seungone Kim, Kai Tzu-iunn Ong, Beong-woo Kwak, Moohyeon Kim, Sunghwan Kim, Taeyoon Kwon, Jiwan Chung, Youngjae Yu, and Jinyoung Yeo. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.1253 Language models as compilers: Simulating pseudocode execution improves algorithmic reasoning in language models . In Proceedings of the 2...

    work page doi:10.18653/v1/2024.emnlp-main.1253 2024

  5. [5]

    Wenhu Chen, Xueguang Ma, Xinyi Wang, and William W. Cohen. 2023. https://openreview.net/forum?id=YfZ4ZPt8zd Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks . Transactions on Machine Learning Research

    2023

  6. [6]

    Sara Court and Micha Elsner. 2024. https://doi.org/10.18653/v1/2024.wmt-1.125 Shortcomings of LLM s for low-resource translation: Retrieval and understanding are both the problem . In Proceedings of the Ninth Conference on Machine Translation, pages 1332--1354, Miami, Florida, USA. Association for Computational Linguistics

    doi:10.18653/v1/2024.wmt-1.125 2024

  7. [7]

    Antonios Dimakis, Stella Markantonatou, and Antonios Anastasopoulos. 2024. https://doi.org/10.18653/v1/2024.findings-acl.152 Dictionary-aided translation for handling multi-word expressions in low-resource languages . In Findings of the Association for Computational Linguistics: ACL 2024, pages 2588--2595, Bangkok, Thailand. Association for Computational ...

    doi:10.18653/v1/2024.findings-acl.152 2024

  8. [8]

    Dryer and Martin Haspelmath, editors

    Matthew S. Dryer and Martin Haspelmath, editors. 2013. https://doi.org/10.5281/zenodo.13950591 WALS Online (v2020.4) . Zenodo

    doi:10.5281/zenodo.13950591 2013

Show all 47 references
  1. [9]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    2024 arXiv

  2. [10]

    Micha Elsner and Jordan Needle. 2023. https://doi.org/10.18653/v1/2023.sigmorphon-1.2 Translating a low-resource language using GPT -3 and a human-readable dictionary . In Proceedings of the 20th SIGMORPHON workshop on Computational Research in Phonetics, Phonology, and Morpho...

    2023 doi

  3. [11]

    Julen Etxaniz, Gorka Azkune, Aitor Soroa, Oier Lacalle, and Mikel Artetxe. 2024. https://doi.org/10.18653/v1/2024.naacl-short.46 Do multilingual language models think better in E nglish? In Proceedings of the 2024 Conference of the North American Chapter of the Association for...

    2024 doi

  4. [12]

    Mikel L. Forcada, Mireia Ginest \'i -Rosell, Jacob Nordfalk, Jimmy O'Regan, Sergio Ortiz Rojas, Juan Antonio P \'e rez-Ortiz, Felipe S \'a nchez-Mart \'i nez, Gema Ram \'i rez-S \'a nchez, and Francis M. Tyers. 2011. https://api.semanticscholar.org/CorpusID:5698842 Apertium: a...

    2011

  5. [13]

    Gemini Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. 2024. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530

    2024 arXiv

  6. [14]

    Marjan Ghazvininejad, Hila Gonen, and Luke Zettlemoyer. 2023. https://arxiv.org/abs/2302.07856 Dictionary-based phrase-level prompting of large language models for machine translation . Preprint, arXiv:2302.07856

    2023 arXiv

  7. [15]

    Michael Ginn, Mans Hulden, and Alexis Palmer. 2024 a . https://doi.org/10.18653/v1/2024.findings-emnlp.337 Can we teach language models to gloss endangered languages? In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 5861--5876, Miami, Florida, US...

    2024 doi

  8. [16]

    Michael Ginn, Lindia Tjuatja, Taiqi He, Enora Rice, Graham Neubig, Alexis Palmer, and Lori Levin. 2024 b . https://doi.org/10.18653/v1/2024.emnlp-main.683 G loss LM : A massively multilingual corpus and pretrained model for interlinear glossed text . In Proceedings of the 2024...

    2024 doi

  9. [17]

    Ping Guo, Yubing Ren, Yue Hu, Yunpeng Li, Jiarui Zhang, Xingsheng Zhang, and Heyan Huang. 2024. https://aclanthology.org/2024.lrec-main.1362/ Teaching large language models to translate on low-resource languages with textbook prompting . In Proceedings of the 2024 Joint Intern...

    2024

  10. [18]

    Jonathan Hus and Antonios Anastasopoulos. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.1127 Back to school: Translation using grammar books . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 20207--20219, Miami, Florida, USA...

    2024 doi

  11. [19]

    Pratik Joshi, Sebastin Santy, Amar Budhiraja, Kalika Bali, and Monojit Choudhury. 2020. https://doi.org/10.18653/v1/2020.acl-main.560 The state and fate of linguistic diversity and inclusion in the NLP world . In Proceedings of the 58th Annual Meeting of the Association for Co...

    2020 doi

  12. [20]

    Liguo Lan. 2016. Annotated Zhuang Grammar Text. Social Sciences Academic Press (China)

    2016

  13. [21]

    Bryan Li, Tamer Alkhouli, Daniele Bonadiman, Nikolaos Pappas, and Saab Mansour. 2024. https://doi.org/10.18653/v1/2024.acl-long.281 Eliciting better multilingual structured reasoning from LLM s through code . In Proceedings of the 62nd Annual Meeting of the Association for Com...

    2024 doi

  14. [22]

    Emmy Liu, Graham Neubig, and Jacob Andreas. 2024. https://openreview.net/forum?id=nUNbjMDBWC An incomplete loop: Instruction inference, instruction following, and in-context learning in language models . In First Conference on Language Modeling

    2024

  15. [23]

    Xiao Liu, Da Yin, Chen Zhang, Yansong Feng, and Dongyan Zhao. 2023. https://doi.org/10.18653/v1/2023.findings-acl.574 The magic of IF : Investigating causal reasoning abilities in large language models of code . In Findings of the Association for Computational Linguistics: ACL...

    2023 doi

  16. [24]

    Xiao Liu, Da Yin, Chen Zhang, Dongyan Zhao, and Yansong Feng. 2025. Eliciting and improving the causal reasoning abilities of large language models with conditional statements. Computational Linguistics, pages 1--38

    2025

  17. [25]

    Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. 2020. https://doi.org/10.1162/tacl_a_00343 Multilingual denoising pre-training for neural machine translation . Transactions of the Association for Computational...

    2020 doi

  18. [26]

    Hongyuan Lu, Haoran Yang, Haoyang Huang, Dongdong Zhang, Wai Lam, and Furu Wei. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.55 Chain-of-dictionary prompting elicits translation in large language models . In Proceedings of the 2024 Conference on Empirical Methods in Natur...

    2024 doi

  19. [27]

    Aman Madaan, Shuyan Zhou, Uri Alon, Yiming Yang, and Graham Neubig. 2022. https://doi.org/10.18653/v1/2022.emnlp-main.90 Language models of code are few-shot commonsense learners . In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages...

    2022 doi

  20. [28]

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

    2024 arXiv

  21. [29]

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. https://doi.org/10.3115/1073083.1073135 B leu: a method for automatic evaluation of machine translation . pages 311--318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics

    2002

  22. [30]

    Maja Popovi \'c . 2017. https://doi.org/10.18653/v1/W17-4770 chr F ++: words helping character n-grams . In Proceedings of the Second Conference on Machine Translation, pages 612--618, Copenhagen, Denmark. Association for Computational Linguistics

    2017 doi

  23. [31]

    Matt Post. 2018. https://doi.org/10.18653/v1/W18-6319 A call for clarity in reporting BLEU scores . In Proceedings of the Third Conference on Machine Translation: Research Papers, pages 186--191, Brussels, Belgium. Association for Computational Linguistics

    2018 doi

  24. [32]

    Haritz Puerto, Martin Tutek, Somak Aditya, Xiaodan Zhu, and Iryna Gurevych. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.629 Code prompting elicits conditional reasoning abilities in T ext+ C ode LLM s . In Proceedings of the 2024 Conference on Empirical Methods in Natura...

    2024 doi

  25. [33]

    Rita Ramos, Everlyn Asiko Chimoto, Maartje ter Hoeve, and Natalie Schluter. 2024. Grammamt: Improving machine translation with grammar-informed in-context learning. arXiv preprint arXiv:2410.18702

    2024 arXiv

  26. [34]

    Stephen Robertson, Hugo Zaragoza, et al. 2009. The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends in Information Retrieval , 3(4):333--389

    2009

  27. [35]

    Leipzig Glossing Rules. 2008. Conventions for interlinear morpheme-by-morpheme glosses. URL: https://www. eva. mpg. de/lingua/resources/glossing-rules. php

    2008

  28. [36]

    Freda Shi, Mirac Suzgun, Markus Freitag, Xuezhi Wang, Suraj Srivats, Soroush Vosoughi, Hyung Won Chung, Yi Tay, Sebastian Ruder, Denny Zhou, Dipanjan Das, and Jason Wei. 2023. https://openreview.net/forum?id=fR3wGCk-IXp Language models are multilingual chain-of-thought reasone...

    2023

  29. [37]

    Garrett Tanzer, Mirac Suzgun, Eline Visser, Dan Jurafsky, and Luke Melas-Kyriazi. 2024. https://openreview.net/forum?id=tbVWug9f2h A benchmark for learning to translate a new language from one grammar book . In The Twelfth International Conference on Learning Representations

    2024

  30. [38]

    Eline Visser. 2022. https://doi.org/10.5281/zenodo.6499927 A grammar of Kalamang . Comprehensive Grammar Library. Language Science Press, Germany

    2022 doi

  31. [39]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2023. https://arxiv.org/abs/2201.11903 Chain-of-thought prompting elicits reasoning in large language models . Preprint, arXiv:2201.11903

    2023 arXiv

  32. [40]

    Jingyun Wei and Xiangzhou Qin. 2008. Basic Course of Zhuang Language. China Minzu University Press

    2008

  33. [41]

    Jingyun Wei and Xiaohang Qin. 2006. General Introduction to Zhuang Language. China Minzu University Press

    2006

  34. [42]

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

    2024 arXiv

  35. [43]

    Zheng Xin Yong, Hailey Schoelkopf, Niklas Muennighoff, Alham Fikri Aji, David Ifeoluwa Adelani, Khalid Almubarak, M Saiful Bari, Lintang Sutawika, Jungo Kasai, Ahmed Baruwa, Genta Winata, Stella Biderman, Edward Raff, Dragomir Radev, and Vassilina Nikoulina. 2023. https://doi....

    2023 doi

  36. [44]

    Chen Zhang, Xiao Liu, Jiuheng Lin, and Yansong Feng. 2024 a . https://doi.org/10.18653/v1/2024.findings-acl.519 Teaching large language models an unseen language on the fly . In Findings of the Association for Computational Linguistics: ACL 2024, pages 8783--8800, Bangkok, Tha...

    2024 doi

  37. [45]

    Kexun Zhang, Yee Choi, Zhenqiao Song, Taiqi He, William Yang Wang, and Lei Li. 2024 b . https://doi.org/10.18653/v1/2024.findings-acl.925 Hire a linguist!: Learning endangered languages in LLM s with in-context linguistic descriptions . In Findings of the Association for Compu...

    2024 doi

  38. [46]

    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...

  39. [47]

    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 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.