Pith. sign in

REVIEW 3 major objections 5 minor 32 references

Exploring the Relationship Between Diversity and Quality in Ad Text Generation

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 In Japanese ad text generation, the paper finds a consistent trade-off between diversity and quality: every decoding method that raises surface diversity lowers at least one ad quality metric, except diverse beam search and a five-model…

desk verdict A careful empirical map of diversity-quality trade-offs in Japanese ad text generation, but the headline quality metric is proprietary and weakly validated. read the letter →

arxiv 2505.16418 v1 pith:MS5WE6VF submitted 2025-05-22 cs.CL

Yoichi Aoki , Soichiro Murakami , Ukyo Honda , Akihiko Kato This is my paper · ORCID
classification cs.CL
keywords adtextgenerationdiversity-qualitytrade-offdecodingmethodsbeamsearchsamplingminimumBayesriskJapaneseadvertisinglargelanguagemodels
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

Japanese ad text is short, product-fact-constrained copy where surface wording is the main freedom, and this paper asks whether decoding-time methods can increase how differently five generated ads read without hurting ad quality. Across five large language models, multiple prompts, and six diversity-enhancing approaches—nucleus and temperature sampling, beam search, diverse beam search, and two minimum Bayes risk decoders—the answer is a systematic trade-off: higher diversity scores line up with lower ad performance, lower consistency, and lower acceptability. The paper isolates two ways around the trade-off: diverse beam search improves diversity while holding performance and acceptability roughly constant, and taking one output from each of five different models gives the highest measured diversity at the average of each model's quality. It also shows that sampling and beam search respond differently to the number of few-shot examples and the number of outputs, so the price of diversity depends on the whole decoding recipe rather than on the method alone.

What carries the argument

The machinery is the decoding recipe itself, placed in diversity-quality coordinates. Diversity is surface diversity, $1 - \text{Pairwise-BLEU}$, which measures how much the five generated texts differ in n-gram wording while product facts remain fixed. Quality is the average of three automatic scores: ad performance, the ratio of a predicted click-through rate to the human-written reference; ad consistency, the BERTScore between the input ad content and the generated text; and ad acceptability, the share of texts that fit the 15 full-width or 30 half-width character limit. The compared methods are nucleus and temperature sampling, beam search with varying beam width, diverse beam search with beam width and group number, and the minimum Bayes risk variants DMBR and KMBR; the paper reads the fitted trade-off line and each method's hyperparameter trajectory. The multi-model comparison applies the same machinery to model identity: one sampled output from each of five models replaces five outputs from a single model.

What would settle it

Re-run the same five-method comparison with a live quality signal derived from real user behavior, such as clicks or conversions on a deployed ad platform, or with a large crowdsourced attractiveness panel designed to separate high-diversity from low-diversity outputs of the same model; if high-diversity outputs no longer score lower on that behavioral or human signal, the trade-off is an artifact of the automatic metrics rather than a property of ad quality.

Watch

Extended reading notes

Core claim

The central discovery is that ad text generation belongs with summarization and machine translation rather than with dialogue or story generation on the diversity-quality map: diversity-enhancing decoding is not free. Measuring diversity as $1 - \text{Pairwise-BLEU}$ over five generated texts and quality as the average of a predicted click-through ratio, a BERTScore consistency value, and a length-acceptability score, the paper finds a downward-sloping best-fit line for every method family. The line is not the whole story. Diverse beam search reaches higher diversity than plain beam search while keeping performance and acceptability at the same level; more few-shot examples help beam search and diverse beam search diversify but not sampling; more outputs help beam search diversify at a quality cost; and all-at-once generation diversifies at a small quality penalty. The strongest positive result is multi-model ensembling: when five different models each write one ad, diversity reaches its highest value while quality sits at the average of the five models, suggesting a diversity lever that does not pass through degraded decoding.

Load-bearing premise

The load-bearing premise is that the automatic quality proxies—especially the model that predicts how often people click an ad—measure ad quality as advertisers and platforms actually experience it; if those proxies are wrong, the reported trade-off describes the proxies rather than the ads.

Editorial extensions

If this is right

  • Advertising text generation should be treated as a task where diversity claims are incomplete without paired quality numbers, since no decoding method improved diversity while improving all three quality metrics.
  • Diverse beam search at width five is a practical choice when performance and acceptability must be protected, because its diversity gain is the largest among methods that preserved those two measures.
  • The interaction between prompts and decoders matters: adding few-shot examples up to fifteen helps beam search and diverse beam search diversify but does not move sampling, while increasing the output count from two to ten broadens beam search diversity at a quality cost.
  • One-at-a-time generation remains the safer default; generating five ads in a single call raises diversity but lowers quality, and the loss is not recovered by method variations.
  • Mixing five different models, one ad each, is a workable substitute for aggressive decoding: it produced the highest measured diversity while holding quality at the models' average.

Reading between the lines

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

  • The paper's attribution of beam-search quality loss to low-probability candidate texts suggests a testable extension: content-preserving paraphrase or controlled rewriting should add surface diversity without touching the probability tail, and might break the trade-off that decoding-time methods all exhibit.
  • Because the appendix reports only 61 percent agreement between human attractiveness judgments and the automatic performance ratio, and a 0.55 correlation for consistency, the ranking of methods should be re-checked with a human-centered quality signal before being used to set real ad-serving policy.
  • The multi-model result leaves open whether model choice can be optimized for complementary vocabularies: adding or removing models one at a time and tracing the diversity-quality frontier would show whether five is a special number or just the start of a scalable diversity gain.
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. This paper reports an empirical study of five diversity-enhancing decoding strategies—nucleus/temperature sampling, beam search, diverse beam search, DMBR, and KMBR—for Japanese ad text generation with five LLMs (calm3, ELYZA, Mistral, Swallow, GPT-4o). For 798 CAMERA inputs, diversity is measured as 1 − pairwise BLEU among five outputs, and quality is measured by a proprietary CTR predictor, BERTScore, and a length-based acceptability criterion. The main claim is that increasing diversity entails a trade-off with ad quality, that diverse beam search improves diversity with relatively small quality loss, and that generating one output from each of five models yields the highest diversity. The paper also reports how the number of shots, the output count, and one-at-once versus all-at-once generation affect the trade-off.

Significance. If the headline trade-off is correct, the paper has practical value for ad platforms: it would caution against blindly increasing diversity and identify diverse beam search and multi-model ensembling as softer alternatives. The study's strengths are breadth: five models, multiple prompts, hyperparameters, shot counts, output counts, and output strategies are compared, and Appendix D includes an attempted human validation. The work is empirical and does not rely on fitted models or circular derivations. However, the significance is currently conditional on two unresolved points: the quality axis depends on a proprietary CTR predictor, and the central trend is reported without uncertainty quantification. With those addressed, the paper could be a useful reference for ad-text generation.

major comments (3)
  1. [Section 3.1 and Figures 2–8 / Appendix B] Every reported quantity is a dataset average with no confidence intervals, error bars, or significance tests, and the ad-performance spread in Figure 2 is only about 0.008 (0.972–0.980); the fitted trade-off line therefore rests on differences that may be within run-to-run noise. Please report per-input statistics, bootstrap or other intervals, and test whether the negative slopes are significant before asserting a trade-off.
  2. [Section 2.3.2 and Appendix D] The ad-performance axis is 'Kiwami Yosoku TD', a proprietary CTR predictor with no public specification or validation; Appendix D reports 61% agreement between human attractiveness judgments and this automatic metric and a Pearson correlation of 0.55 for consistency, both moderate. The sentence in Section 3.1 that 'The human evaluation results were consistent with these results' is not supported: Appendix D tests metric alignment, not whether the diversity-quality trade-off is observed by humans. Either augment the study with a human test of the actual trade-off or explicitly restrict the central claim to the automatic proxies.
  3. [Section 3.4, Table 3] In the beam-search block, the output-number-5 row reports Ad Performance 0.900, while the output-number-2 and output-number-10 rows report 0.981 and 0.979; this breaks the monotone quality decline described in the text and appears to be a typo. Please verify the entry and rerun the parameter trend; if it is a typo, correct it and check for other table transcription errors.
minor comments (5)
  1. [Section 2.3.1] The sentence containing 'we evaluat only' should read 'we evaluate only'.
  2. [Table 6] The parameter name 'temperture' should be 'temperature'.
  3. [Appendix G] The text refers to 'Figure 15' when the referenced object is Table 15.
  4. [Table 13] The column header 'Attempt.' should be 'Accept.' to match the other tables.
  5. [Ethics statement] The claim that 'no human experiments are conducted' conflicts with the human evaluation described in Appendix D; the statement should be revised or the appendix should be described as a user study.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the diversity-quality trade-off is an empirical measurement, with diversity and quality defined independently and no fitted parameters or self-citation chain forcing the result.

full rationale

This paper is an empirical measurement study rather than a derivation. Diversity is defined as 1 minus pairwise BLEU among generated ad texts (Section 2.3.1), and quality is defined independently as CTR-ratio ad performance, BERTScore consistency, and a length-based acceptability (Section 2.3.2). The central claim of a trade-off between diversity and ad quality is read off the observed scatter of Figure 2 and the parameter tables; it is not implied by the definitions of the metrics, and no parameter is fitted to the reported outcome and then renamed as a prediction. The self-citations to CAMERA, AdParaphrase, DMBR/KMBR, and prior ad-generation work provide datasets, decoding methods, and task framing, but the diversity-quality relationship itself is computed from the present experiments across multiple models, prompts, shot counts, and output counts, so the citations are not load-bearing for the empirical conclusion. The Appendix D human-validation numbers (61% agreement; r=0.55) are a possible threat to the validity of the automatic quality proxy, not evidence of circularity, because the automatic and human measures are still distinct operationalizations. No step in the paper reduces by construction to its inputs, so no circularity is found.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

This study fits no parameters and introduces no entities. Its claims rest on four measurement and scope choices: the BLEU-based diversity definition, the proprietary CTR predictor for performance, BERTScore for consistency, and the CAMERA subset for domain coverage. The human evaluation provides partial independent support for the automatic quality metrics, which lowers the circularity burden.

assumptions (5)
  • domain assumption 1-Pairwise-BLEU over the five outputs is an adequate measure of ad diversity.
    The paper deliberately excludes semantic diversity because product details must be fixed (Section 2.3.1). This narrows diversity to surface variation; if semantic variety matters for ad fatigue, the result is partial.
  • domain assumption Kiwami Yosoku TD's predicted CTR approximates true advertising performance.
    Section 2.3.2 uses it to compute the performance ratio. Appendix D human evaluation shows only 61% agreement with human attractiveness and r=0.55 for consistency, so the proxy is only moderately validated.
  • domain assumption BERTScore between input ad content and generated text measures consistency.
    Used as the consistency metric in Section 2.3.2. The moderate human correlation of 0.55 in Appendix D supports but does not fully validate it.
  • domain assumption The 798 ad texts from CAMERA are representative of Japanese ad text generation tasks.
    Appendix A specifies the subset and the 15/30 character limits, but gives no sampling justification. The Limitations section appropriately restricts the claims to Japanese ad texts.
  • domain assumption The hyperparameter choices in Table 6 permit fair comparison of the diversity-enhancing methods.
    If top_p=1.0, temperature=1.0, num_beams=5, and diversity_penalty=1.0 are atypical for any method, the ordering in Figure 2 could change. The paper partially addresses this by varying parameters in Section 3.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Exploring the Relationship Between Diversity and Quality in Ad Text Generation." pith.science (2026). https://pith.science/paper/MS5WE6VF

@misc{pith2026250516418,
  author       = {Pith},
  title        = {Pith review of: Exploring the Relationship Between Diversity and Quality in Ad Text Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MS5WE6VF}},
  note         = {Machine review of arXiv:2505.16418}
}
read the original abstract

In natural language generation for advertising, creating diverse and engaging ad texts is crucial for capturing a broad audience and avoiding advertising fatigue. Regardless of the importance of diversity, the impact of the diversity-enhancing methods in ad text generation -- mainly tested on tasks such as summarization and machine translation -- has not been thoroughly explored. Ad text generation significantly differs from these tasks owing to the text style and requirements. This research explores the relationship between diversity and ad quality in ad text generation by considering multiple factors, such as diversity-enhancing methods, their hyperparameters, input-output formats, and the models.

Figures

Figures reproduced from arXiv: 2505.16418 by the authors.

Figure 1
Figure 1. We input ad content into large language mod [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. This figure shows the relationship between diversity and ad quality in ad text generation. Ad performance [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Relationship between diversity and ad quality in ad text generation with ELYZA [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Relationship between diversity and ad quality in ad text generation with Mistral [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Relationship between diversity and ad quality in ad text generation with Swallow [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Relationship between diversity and ad quality in ad text generation with GPT-4o [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Diversity and ad quality of sampling when varying the order of few-shot examples [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Diversity and ad quality of sampling when varying the instruction [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 13 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]

    Ackley, Geoffrey E

    David H. Ackley, Geoffrey E. Hinton, and Terrence J. Sejnowski. 1985. https://doi.org/10.1207/S15516709COG0901\_7 A learning algorithm for boltzmann machines . Cogn. Sci., 9(1):147--169

    doi:10.1207/s15516709cog0901 1985

  4. [4]

    Angela Fan, Mike Lewis, and Yann N. Dauphin. 2018. https://doi.org/10.18653/V1/P18-1082 Hierarchical neural story generation . In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, ACL 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers , pages 889--898. Association for Computational Linguistics

    doi:10.18653/v1/p18-1082 2018

  5. [5]

    Kazuki Fujii, Taishi Nakamura, Mengsay Loem, Hiroki Iida, Masanari Ohi, Kakeru Hattori, Hirai Shota, Sakae Mizuki, Rio Yokota, and Naoaki Okazaki. 2024. Continual pre-training for cross-lingual llm adaptation: Enhancing japanese language capabilities. In Proceedings of the First Conference on Language Modeling, COLM, page (to appear), University of Pennsy...

    2024

  6. [6]

    Konstantin Golobokov, Junyi Chai, Victor Ye Dong, Mandy Gu, Bingyu Chi, Jie Cao, Yulan Yan, and Yi Liu. 2022. https://doi.org/10.18653/V1/2022.EMNLP-DEMOS.19 Deepgen: Diverse search ad generation and real-time customization . In Proceedings of the The 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022 - System Demonstrations, ...

    work page doi:10.18653/v1/2022.emnlp-demos.19 2022

  7. [7]

    Alex Graves. 2012. https://arxiv.org/abs/1211.3711 Sequence transduction with recurrent neural networks . CoRR, abs/1211.3711

    arXiv 2012

  8. [8]

    Masato Hirakawa, Shintaro Horie, Tomoaki Nakamura, Daisuke Oba, Sam Passaglia, and Akira Sasaki. 2024. https://huggingface.co/elyza/Llama-3-ELYZA-JP-8B elyza/llama-3-elyza-jp-8b

    work page 2024

Show all 32 references
  1. [9]

    Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2020. https://openreview.net/forum?id=rygGQyrFvH The curious case of neural text degeneration . In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020 . Ope...

    2020

  2. [10]

    Weston Hughes, Keng - hao Chang, and Ruofei Zhang

    J. Weston Hughes, Keng - hao Chang, and Ruofei Zhang. 2019. https://doi.org/10.1145/3292500.3330754 Generating better search engine text advertisements with deep reinforcement learning . In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Da...

    2019

  3. [11]

    Aaron Hurst, Adam Lerer, Adam P. Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, Aleksander Madry, Alex Baker - Whitcomb, Alex Beutel, Alex Borzunov, Alex Carney, Alex Chow, Alex Kirillov, Alex Nichol, Alex Paino, and 7...

    2024 doi arXiv

  4. [12]

    Ryosuke Ishigami. 2024. https://huggingface.co/cyberagent/calm3-22b-chat cyberagent/calm3-22b-chat

    2024

  5. [13]

    Yuu Jinnai, Ukyo Honda, Tetsuro Morimura, and Peinan Zhang. 2024. https://doi.org/10.18653/V1/2024.FINDINGS-ACL.503 Generating diverse and high-quality texts by minimum bayes risk decoding . In Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thail...

    2024 doi

  6. [14]

    Hidetaka Kamigaito, Peinan Zhang, Hiroya Takamura, and Manabu Okumura. 2021. https://doi.org/10.18653/V1/2021.NAACL-INDUSTRY.32 An empirical study of generating texts for search engine advertising . In Proceedings of the 2021 Conference of the North American Chapter of the Ass...

    2021 doi

  7. [15]

    Philipp Koehn and Rebecca Knowles. 2017. https://doi.org/10.18653/V1/W17-3204 Six challenges for neural machine translation . In Proceedings of the First Workshop on Neural Machine Translation, NMT@ACL 2017, Vancouver, Canada, August 4, 2017, pages 28--39. Association for Comp...

    2017 doi

  8. [16]

    Jiwei Li, Will Monroe, and Dan Jurafsky. 2016. https://arxiv.org/abs/1611.08562 A simple, fast diverse decoding algorithm for neural generation . CoRR, abs/1611.08562

    2016 arXiv

  9. [17]

    Bruce T. Lowerre. 1976. The harpy speech recognition system. Ph.D. thesis, Carnegie Mellon University, USA. AAI7619331

    1976

  10. [18]

    Youmi Ma, Sakae Mizuki, Kazuki Fujii, Taishi Nakamura, Masanari Ohi, Hinari Shimada, Taihei Shiotani, Koshiro Saito, Koki Maeda, Kakeru Hattori, Takumi Okamoto, Shigeki Ishida, Rio Yokota, Hiroya Takamura, and Naoaki Okazaki. 2025. https://arxiv.org/abs/2503.23714 Building ins...

    2025 arXiv

  11. [19]

    Clara Meister, Gian Wiher, and Ryan Cotterell. 2022. https://doi.org/10.1162/TACL\_A\_00502 On decoding strategies for neural text generators . Trans. Assoc. Comput. Linguistics, 10:997--1012

    2022 doi

  12. [20]

    Mistral AI team . 2025. https://mistral.ai/news/mistral-small-3/ Mistral small 3 . Mistral AI News

    2025

  13. [21]

    Masato Mita, Soichiro Murakami, Akihiko Kato, and Peinan Zhang. 2024. https://doi.org/10.18653/V1/2024.ACL-LONG.54 Striking gold in advertising: Standardization and exploration of ad text generation . In Proceedings of the 62nd Annual Meeting of the Association for Computation...

    2024 doi

  14. [22]

    Soichiro Murakami, Sho Hoshino, and Peinan Zhang. 2023. https://doi.org/10.48550/ARXIV.2306.12719 Natural language generation for advertising: A survey . CoRR, abs/2306.12719

    2023 doi arXiv

  15. [23]

    Soichiro Murakami, Peinan Zhang, Hidetaka Kamigaito, Hiroya Takamura, and Manabu Okumura. 2025. https://aclanthology.org/2025.findings-naacl.78/ A d P araphrase: Paraphrase dataset for analyzing linguistic features toward generating attractive ad texts . In Findings of the Ass...

    2025

  16. [24]

    Naoaki Okazaki, Kakeru Hattori, Hirai Shota, Hiroki Iida, Masanari Ohi, Kazuki Fujii, Taishi Nakamura, Mengsay Loem, Rio Yokota, and Sakae Mizuki. 2024. Building a large japanese web corpus for large language models. In Proceedings of the First Conference on Language Modeling,...

    2024

  17. [25]

    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 of the Association for Computational Linguistics, July 6-12, ...

    2002

  18. [26]

    Karl Pearson and Francis Galton. 1895. https://doi.org/10.1098/rspl.1895.0041 Vii. note on regression and inheritance in the case of two parents . Proceedings of the Royal Society of London, 58(347-352):240--242

  19. [27]

    Corneilia Pechman and David W. Stewart. 1988. Advertising Repetition : A Critical Review of Wearin and Wearout . Current issues and research in advertising, 11(1-2):285--329

    1988

  20. [28]

    Rush, Yin - Wen Chang, and Michael Collins

    Alexander M. Rush, Yin - Wen Chang, and Michael Collins. 2013. https://aclanthology.org/D13-1022/ Optimal beam search for machine translation . In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, EMNLP 2013, 18-21 October 2013, Grand Hyat...

    2013

  21. [29]

    Susanne Schmidt and Martin Eisend. 2015. https://doi.org/10.1080/00913367.2015.1018460 Advertising Repetition : A Meta - Analysis on Effective Frequency in Advertising . Journal of Advertising, 44(4):415--428

    2015

  22. [30]

    Sina Zarrie , Henrik Voigt, and Simeon Sch \" u z. 2021. https://doi.org/10.3390/INFO12090355 Decoding methods in neural language generation: A survey . Inf., 12(9):355

    2021 doi

  23. [31]

    Peinan Zhang, Yusuke Sakai, Masato Mita, Hiroki Ouchi, and Taro Watanabe. 2024. https://doi.org/10.48550/ARXIV.2408.05906 Adtec: A unified benchmark for evaluating text quality in search engine advertising . CoRR, abs/2408.05906

    2024 doi arXiv

  24. [32]

    Weinberger, and Yoav Artzi

    Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020. https://openreview.net/forum?id=SkeHuCVFDr Bertscore: Evaluating text generation with BERT . In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April...

    2020

Pith tools

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