Pith. sign in

REVIEW 4 major objections 4 minor 37 references

FormosanBench: Benchmarking Low-Resource Austronesian Languages in the Era of Large Language Models

T0 review · 4 major / 4 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 Current LLMs score near zero on three endangered Formosan languages.

desk verdict A useful first benchmark for Formosan languages with credible negative results, but the presentation needs a cleanup before I'd trust the finer claims. read the letter →

arxiv 2506.21563 v1 pith:UMDCI4PE submitted 2025-06-12 cs.CL

Kaiying Kevin Lin , Hsiyu Chen , Haopeng Zhang This is my paper
classification cs.CL
keywords Formosanlanguageslow-resourceNLPbenchmarkmachinetranslationautomaticspeechrecognitiontextsummarizationendangeredlargelanguagemodels
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

FormosanBench is the first multi-task benchmark for three endangered Formosan languages—Atayal, Amis, and Paiwan—spanning machine translation, automatic speech recognition, and text summarization. The paper evaluates state-of-the-art large language models in zero-shot, 10-shot, and fine-tuned settings and finds that they consistently fail: translation BLEU scores are near zero, ASR word error rates remain high, and summarization ROUGE scores stay below 20 for most models. The paper argues that this failure indicates a lack of genuine understanding of Formosan linguistic structure, and that standard adaptation techniques such as few-shot learning and small-scale fine-tuning provide only limited, inconsistent gains.

What carries the argument

The load-bearing mechanism is the benchmark itself: FormosanBench couples three tasks with standard automatic metrics—BLEU for translation, word error rate for ASR, ROUGE for summarization—so that a near-total failure on any task is visible as an extreme score. The evaluation protocol—zero-shot prompting, 10-shot in-context learning, and task-specific fine-tuning—lets the paper separate the effect of data scarcity from the effect of model scale and pretraining.

What would settle it

If a human-annotated sample of model outputs on FormosanBench showed that many translations and summaries are actually understandable and content-correct despite low automatic scores, the central claim of near-total failure would be falsified; recomputing metrics with morpheme-aware tokenization or using human ratings would settle the question.

Watch

Extended reading notes

Core claim

The central discovery is that current large language models have no meaningful capability in Atayal, Amis, or Paiwan. Across all non-fine-tuned models, Formosan-to-Mandarin translation produces BLEU scores of essentially zero, and Mandarin-to-Formosan scores barely exceed 0.003. Fine-tuning NLLB raises Mandarin-to-Formosan BLEU to roughly 0.15 at best, while fine-tuned Mistral stays at or below 0.02, and Formosan-to-Mandarin translation remains at zero for every model. ASR is the one task where adaptation helps substantially: fine-tuned Whisper lowers word error rate to about 0.3, whereas zero-shot models range from about 0.5 to 1.5. Summarization ROUGE-2 scores are below 20 for most model-language combinations; the paper interprets GPT-4o's few-shot gains on Atayal and Amis as surface pattern replication rather than genuine understanding.

Load-bearing premise

The evaluation assumes that BLEU, ROUGE, and word error rate, computed with standard tokenizers and reference alignments, accurately measure model capability on these morphologically rich languages; if tokenization or alignment mismatches inflate error, the near-zero scores could overstate the failure.

Editorial extensions

If this is right

  • No current LLM can serve as a usable translator or summarizer for Formosan languages without substantial additional work.
  • Few-shot learning and small-scale fine-tuning are insufficient adaptation strategies for these languages, so future progress hinges on language-specific data or morphologically aware models.
  • Speech-based adaptation is more promising than text adaptation; the fine-tuned Whisper reaches a word error rate near 0.3, the only result approaching practical use.
  • The uniform near-zero translation scores across model families indicate that pretraining corpora effectively lack Formosan data, so scaling alone will not close the gap.
  • FormosanBench provides a fixed, public benchmark against which future Formosan-language methods can be measured.

Reading between the lines

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

  • The near-zero BLEU scores might partly reflect tokenizer and reference-alignment artifacts on morphologically rich Austronesian words; a human rating study of the same outputs would separate genuine failure from metric failure.
  • If the voice-system complexity is a core barrier, then models that take morphemes or grammatical roles as input units may outperform existing tokenizers on Formosan languages.
  • The benchmark could be extended to other early-diverging Austronesian languages with available resources, or to tasks such as morphological analysis, to test whether the observed failure is a data problem or a typological distance problem.
  • A useful next test would be to apply continued pretraining on Formosan text alone and measure whether translation and summarization improve, since the paper's fine-tuning results suggest that current adaptation budgets are too small.
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

4 major / 4 minor

Summary. The manuscript introduces FormosanBench, a new multi-task benchmark covering three endangered Formosan languages (Amis, Atayal, Paiwan) with datasets for machine translation, automatic speech recognition, and text summarization. The authors evaluate a range of open and proprietary models in zero-shot, 10-shot, and fine-tuned settings, reporting very low BLEU and ROUGE scores and high WER. They conclude that current LLMs underperform on these languages and that standard adaptation techniques yield limited gains, and they release datasets and code. The paper is primarily an empirical resource paper.

Significance. If the results are reliable, FormosanBench is a valuable addition to low-resource and endangered-language NLP. It is, to the authors' knowledge, the first multi-task benchmark for Formosan languages, and the public release of datasets and code is a concrete contribution. The broad finding that large multilingual models perform poorly on these typologically distinct languages is plausible and useful for motivating further work. However, the strength of the conclusions is currently limited by evaluation-metric questions and internal inconsistencies in the reported results; these need to be resolved before the benchmark can serve as a reliable reference.

major comments (4)
  1. [Abstract; §6.2; §7] The central claim that 10-shot learning and fine-tuning offer 'only limited improvements' is contradicted by the ASR and summarization results. Table 3 shows fine-tuned Whisper reducing WER from 1.026/1.057/1.098 to 0.296/0.315/0.373 across Amis/Atayal/Paiwan, a roughly 70% relative reduction that §6.2 describes as 'notable' and §7 as 'substantial.' In §6.3, 10-shot GPT-4o raises ROUGE-L for Atayal from 13.03 to 50.41 and for Amis from 14.52 to 30.02. The abstract's blanket statement about limited improvements should be qualified to match these results.
  2. [§5.3; Table 2; §6.1] Near-zero BLEU scores are interpreted as 'a near-total failure to understand and translate' (§6.1), but the paper does not report the tokenizer, normalization, or BLEU implementation used, and no qualitative examples or human evaluation of MT outputs are provided. For morphologically rich verb-initial Austronesian languages, exact n-gram matching can yield near-zero scores for outputs that are reasonable but non-identical translations. The authors should either provide tokenization details plus human or reference-based qualitative checks, or soften the 'near-total failure' claim.
  3. [Table 1; §4] The stated split is '70:10:20' in §4, but the numbers in Table 1 do not follow that ratio. For example, MT Amis has 3,259/543/1,631 datapoints, which is 60.0%/10.0%/30.0% of 5,433 total; the same pattern appears in every row of Table 1. The split must be corrected or the text changed to reflect the actual ratio, since the released benchmark splits are part of the paper's contribution.
  4. [§5.2; Table 3] The paper states that Atayal and Paiwan were not included in MMS's pretraining corpus and that Amis was selected as the target language, yet Table 3 reports MMS WER for Atayal and Paiwan. It is unclear how MMS produced transcriptions for languages it does not support, and whether forcing the Amis language ID distorts the comparison. Please clarify the decoding procedure and, if these languages are unsupported, remove or explicitly mark the MMS rows for Atayal and Paiwan.
minor comments (4)
  1. [Table 1] MT Total Words for Amis and Atayal are both reported as 38,946; since the languages and dataset sizes differ, this appears to be a copy error and should be corrected.
  2. [Appendix A] The word 'Transaltion' in the MT prompt templates is misspelled; this also appears in the example prompts.
  3. [§6.3] The paper acknowledges that ROUGE scores below 20 are unreliable and fluctuations likely reflect noise, but then draws conclusions about improvements or declines without statistical significance tests or confidence intervals. A brief statement of variance or a paired test would strengthen these claims.
  4. [§5.3] The human evaluation of GPT-4o summarization outputs is described only qualitatively; reporting the number of annotators, evaluation criteria, and agreement would make the claims in Section 6.3 more transparent.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: FormosanBench is an empirical benchmark whose central underperformance claim is measured against external models and external data.

full rationale

This paper is an empirical benchmarking study, not a derivation. It assembles datasets from external sources (Taiwan Indigenous Languages E-Dictionary, FormosanBank/Klokah, Wikipedia) and evaluates independently developed models (LLaMA, Gemma, Mistral, NLLB, GPT-4o, Whisper, Seamless, MMS) on those datasets. The central claim that existing LLMs consistently underperform on Formosan languages is an experimental measurement, reported as BLEU, WER, and ROUGE scores against held-out test splits. No parameter is fitted to the result it is then used to predict; no evaluation quantity is defined in terms of the paper's own outputs; and no conclusion is justified by a self-citation chain. The self-citations that appear (Yuan and Zhang 2024; Zhang et al. 2022, 2024) are background references for summarization construction and survey context, not load-bearing justification for the empirical findings. A possible concern that near-zero BLEU scores could be inflated or deflated by tokenization or reference alignment is a validity concern about the metric, not a circularity: the paper never equates its conclusion with its input assumptions, and its released data and code make the result externally checkable. The benchmark conclusions are therefore self-contained and falsifiable against independent model predictions.

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

No free parameters or invented entities are introduced. The benchmark relies on standard NLP metrics and the quality of the underlying data sources, which are reasonable but unvalidated assumptions.

assumptions (2)
  • domain assumption BLEU, ROUGE, and WER scores computed with standard tokenizers appropriately measure model performance on Formosan languages.
    The paper uses these metrics without adaptation to the morphological and orthographic properties of Formosan languages; if tokenization mismatches occur, scores may be misleading.
  • domain assumption The introductory paragraph of each Formosan Wikipedia article is a valid gold-standard summary of the full article.
    The summarization dataset is built on this assumption; the paper filters low-length articles but does not validate summary quality independently.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FormosanBench: Benchmarking Low-Resource Austronesian Languages in the Era of Large Language Models." pith.science (2026). https://pith.science/paper/UMDCI4PE

@misc{pith2026250621563,
  author       = {Pith},
  title        = {Pith review of: FormosanBench: Benchmarking Low-Resource Austronesian Languages in the Era of Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UMDCI4PE}},
  note         = {Machine review of arXiv:2506.21563}
}
read the original abstract

While large language models (LLMs) have demonstrated impressive performance across a wide range of natural language processing (NLP) tasks in high-resource languages, their capabilities in low-resource and minority languages remain significantly underexplored. Formosan languages -- a subgroup of Austronesian languages spoken in Taiwan -- are both linguistically rich and endangered, largely due to the sociolinguistic dominance of Mandarin. In this work, we introduce FORMOSANBENCH, the first benchmark for evaluating LLMs on low-resource Austronesian languages. It covers three endangered Formosan languages: Atayal, Amis, and Paiwan, across three core NLP tasks: machine translation, automatic speech recognition (ASR), and text summarization. We assess model performance in zero-shot, 10-shot, and fine-tuned settings using FORMOSANBENCH. Our results reveal a substantial performance gap between high-resource and Formosan languages. Existing LLMs consistently underperform across all tasks, with 10-shot learning and fine-tuning offering only limited improvements. These findings underscore the urgent need for more inclusive NLP technologies that can effectively support endangered and underrepresented languages. We release our datasets and code to facilitate future research in this direction.

Figures

Figures reproduced from arXiv: 2506.21563 by the authors.

Figure 1
Figure 1. Four patterns of a sentence ’Zepulj eats sweet [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. 16 Formosan languages in Taiwan Early Austronesian Malayo-Polynesian Paiwan Amis Atayal Tagalog Hawaiian etc [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Austronesian family tree [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Task Description for FORMOSANBENCH datasets. All the prompts were given in Mandarin, but we provide them in English for clarity. In the case of summarization, text appearing in the parentheses indicates content written in the Formosan language. book section contains wr…
Figure 5
Figure 5. Figure 5: BLEU scores ↑ compared with Zheng et al. (2024) these low-resource languages. In the reverse di￾rection, Mandarin-to-Formosan, performance was only marginally higher: for example, the high￾est non-fine-tuned BLEU scores were just 0.0019 for Paiwan (Gemma-1.1-7B-it), 0.…
Figure 6
Figure 6. Figure 6: , the fine-tuned Whisper model shows a notable reduction in WER, achieving scores around 0.3. This improvement may be attributed to Whis￾per’s extensive pretraining on hundreds of thou￾Amis Atayal Paiwan Zero-shot Models Whisper Base 1.026 1.057 1.098 Seamless 1.225 1.…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 18 canonical work pages

  1. [1]

    David Ifeoluwa Adelani, Jessica Ojo, Israel Abebe Azime, Jian Yun Zhuang, Jesujoba O. Alabi, Xuanli He, Millicent Ochieng, Sara Hooker, Andiswa Bukula, En-Shiun Annie Lee, Chiamaka Chukwuneke, Happy Buzaaba, Blessing Sibanda, Godson Kalipe, Jonathan Mukiibi, Salomon Kabongo, Foutse Yuehgoh, Mmasibidi Setaka, Lolwethu Ndolela, Nkiruka Odu, Rooweither Mabuy...

    arXiv 2025

  2. [2]

    Roee Aharoni, Melvin Johnson, and Orhan Firat. 2019. https://doi.org/10.18653/v1/N19-1388 Massively multilingual neural machine translation . In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) , pages 3874--3884, Minneapolis, M...

    doi:10.18653/v1/n19-1388 2019

  3. [3]

    Voices Unheard: NLP Resources and Models for Yor\`ub\'a Regional Dialects

    Orevaoghene Ahia, Anuoluwapo Aremu, Diana Abagyan, Hila Gonen, David Ifeoluwa Adelani, Daud Abolade, Noah A. Smith, and Yulia Tsvetkov. 2024. https://arxiv.org/abs/2406.19564 Voices unheard: Nlp resources and models for yor\`ub\'a regional dialects . Preprint, arXiv:2406.19564

    work page Pith review arXiv 2024

  4. [4]

    Meta AI. 2024. https://ai.meta.com/blog/meta-llama-3-1/ Introducing llama 3.1: Our most capable models to date . Accessed: 2025-04-09

    work page 2024

  5. [5]

    Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. https://arxiv.org/abs/2006.11477 wav2vec 2.0: A framework for self-supervised learning of speech representations . Preprint, arXiv:2006.11477

    arXiv 2020

  6. [7]

    Robert Blust. 1984. http://www.jstor.org/stable/42928105 The austronesian homeland: A linguistic perspective . Asian Perspectives, 26(1):45--67

    work page arXiv 1984

  7. [8]

    Robert Blust. 2019. https://doi.org/10.1146/annurev-linguistics-011718-012440 The austronesian homeland and dispersal . Annual Review of Linguistics, 5(Volume 5, 2019):417--434

    work page doi:10.1146/annurev-linguistics-011718-012440 2019

  8. [9]

    Shuanfan Chang. 2018. Paiwan yufa gailun [A concise grammar of Paiwan]. Council of Indigenous Peoples, Taiwan

    work page 2018

Show all 37 references
  1. [10]

    Seamless Communication, Loïc Barrault, Yu-An Chung, Mariano Cora Meglioli, David Dale, Ning Dong, Paul-Ambroise Duquenne, Hady Elsahar, Hongyu Gong, Kevin Heffernan, John Hoffman, Christopher Klaiber, Pengwei Li, Daniel Licht, Jean Maillard, Alice Rakotoarison, Kaushik Ram Sad...

    2023 arXiv

  2. [11]

    Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzm \'a n, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. https://doi.org/10.18653/v1/2020.acl-main.747 Unsupervised cross-lingual representation learning ...

    2020 doi

  3. [12]

    FormosanBank contributors . 2024. Formosanbank: A repository of formosan language resources. https://github.com/FormosanBank/FormosanBank. Accessed: 2024-05-08

    2024

  4. [13]

    Hedderich, Lukas Lange, Heike Adel, Jannik Strötgen, and Dietrich Klakow

    Michael A. Hedderich, Lukas Lange, Heike Adel, Jannik Strötgen, and Dietrich Klakow. 2021. https://arxiv.org/abs/2010.12309 A survey on recent approaches for natural language processing in low-resource scenarios . Preprint, arXiv:2010.12309

    2021 arXiv

  5. [14]

    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

  6. [15]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas ...

    2023 arXiv

  7. [16]

    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

  8. [17]

    Raviraj Joshi, Kanishk Singla, Anusha Kamath, Raunak Kalani, Rakesh Paul, Utkarsh Vaidya, Sanjay Singh Chauhan, Niranjan Wartikar, and Eileen Long. 2025. https://arxiv.org/abs/2410.14815 Adapting multilingual llms to low-resource languages using continued pre-training and synt...

    2025 arXiv

  9. [18]

    Eric Liao. 2023. https://hdl.handle.net/11296/37xrcb Multilingual machine translation between atayalic languages and chinese . Master's thesis, National Taiwan Ocean University. Taiwan Thesis and Dissertation System

    2023

  10. [19]

    Chin-Yew Lin. 2004. https://aclanthology.org/W04-1013/ ROUGE : A package for automatic evaluation of summaries . In Text Summarization Branches Out, pages 74--81, Barcelona, Spain. Association for Computational Linguistics

    2004

  11. [20]

    Liu, Mohammad Saleh, Etienne Pot, Ben Goodrich, Ryan Sepassi, Lukasz Kaiser, and Noam Shazeer

    Peter J. Liu, Mohammad Saleh, Etienne Pot, Ben Goodrich, Ryan Sepassi, Lukasz Kaiser, and Noam Shazeer. 2018. https://arxiv.org/abs/1801.10198 Generating wikipedia by summarizing long sequences . Preprint, arXiv:1801.10198

    2018 arXiv

  12. [21]

    Alexandre Magueresse, Vincent Carles, and Evan Heetderks. 2020. Low-resource languages: A review of past work and future challenges. arXiv preprint arXiv:2006.07264

    2020 arXiv

  13. [22]

    Atharva Mutsaddi and Aditya Prashant Choudhary. 2025. https://aclanthology.org/2025.loreslm-1.6/ Enhancing plagiarism detection in M arathi with a weighted ensemble of TF - IDF and BERT embeddings for low-resource language processing . In Proceedings of the First Workshop on L...

    2025

  14. [23]

    OpenAI. 2024. Gpt-4o technical report. https://openai.com/index/gpt-4o. Accessed: 2025-05-05

    2024

  15. [24]

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. https://doi.org/10.3115/1073083.1073135 Bleu: a method for automatic evaluation of machine translation . In Proceedings of the 40th Annual Meeting on Association for Computational Linguistics, ACL '02, page 311...

    2002

  16. [25]

    Vineel Pratap, Andros Tjandra, Bowen Shi, Paden Tomasello, Arun Babu, Sayani Kundu, Ali Elkahky, Zhaoheng Ni, Apoorv Vyas, Maryam Fazel-Zarandi, Alexei Baevski, Yossi Adi, Xiaohui Zhang, Wei-Ning Hsu, Alexis Conneau, and Michael Auli. 2023. Scaling speech technology to 1,000+ ...

    2023

  17. [26]

    Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2022. https://arxiv.org/abs/2212.04356 Robust speech recognition via large-scale weak supervision . Preprint, arXiv:2212.04356

    2022 arXiv

  18. [27]

    Sebastian Ruder, Ivan Vuli \'c , and Anders S gaard. 2022. Square one bias in nlp: Towards a multi-dimensional exploration of the research manifold. arXiv preprint arXiv:2206.09755

    2022 arXiv

  19. [28]

    Guokan Shang, Hadi Abdine, Yousef Khoubrane, Amr Mohamed, Yassine Abbahaddou, Sofiane Ennadir, Imane Momayiz, Xuguang Ren, Eric Moulines, Preslav Nakov, Michalis Vazirgiannis, and Eric Xing. 2024. https://arxiv.org/abs/2409.17912 Atlas-chat: Adapting large language models for ...

    2024 arXiv

  20. [29]

    Hoang Thang Ta, Abu Bakar Siddiqur Rahman, Navonil Majumder, Amir Hussain, Lotfollah Najjar, Newton Howard, Soujanya Poria, and Alexander Gelbukh. 2022. https://arxiv.org/abs/2209.13101 Wikides: A wikipedia-based dataset for generating short descriptions from paragraphs . Prep...

    2022 arXiv

  21. [30]

    Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, Pouya Tafti, Léonard Hussenot, Pier Giuseppe Sessa, Aakanksha Chowdhery, Adam Roberts, Aditya Barua, Alex Botev, Alex...

    2024 arXiv

  22. [31]

    NLLB Team, Marta R. Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, Anna Sun, Skyler Wang, Guillaume Wenzek, Al Youngblood, Bapi Akula, Loic Barrault, Gabriel Mejia Gonzalez, Prang...

    2022 arXiv

  23. [32]

    Guillaume Wenzek, Marie-Anne Lachaux, Alexis Conneau, Vishrav Chaudhary, Francisco Guzm \'a n, Armand Joulin, and Edouard Grave. 2020. https://aclanthology.org/2020.lrec-1.494/ CCN et: Extracting high quality monolingual datasets from web crawl data . In Proceedings of the Twe...

    2020

  24. [33]

    Haohan Yuan and Haopeng Zhang. 2024. Domainsum: A hierarchical benchmark for fine-grained domain shift in abstractive text summarization. arXiv preprint arXiv:2410.15687

    2024 arXiv

  25. [34]

    Haopeng Zhang, Semih Yavuz, Wojciech Kryscinski, Kazuma Hashimoto, and Yingbo Zhou. 2022. Improving the faithfulness of abstractive summarization via entity coverage control. arXiv preprint arXiv:2207.02263

    2022 arXiv

  26. [35]

    Haopeng Zhang, Philip S Yu, and Jiawei Zhang. 2024. A systematic survey of text summarization: From statistical methods to large language models. ACM Computing Surveys

    2024

  27. [36]

    Francis Zheng, Edison Marrese-Taylor, and Yutaka Matsuo. 2024. https://doi.org/10.18653/v1/2024.findings-acl.670 Improving low-resource machine translation for formosan languages using bilingual lexical resources . In Findings of the Association for Computational Linguistics: ...

    2024 doi

  28. [37]

    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...

  29. [38]

    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.