Pith. sign in

REVIEW 5 major objections 4 minor 1 cited by

AdParaphrase v2.0: Generating Attractive Ad Texts Using a Preference-Annotated Paraphrase Dataset

T0 review · 5 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 By releasing a 20-fold larger preference-annotated ad-paraphrase dataset of 16,460 Japanese pairs, this paper claims that attractive ad text is tied to specific linguistic features — length, noun density, low perplexity, specificity, and…

desk verdict The dataset is the real contribution; the analytic claims on top need tightening before you trust the feature findings. read the letter →

arxiv 2505.20826 v1 pith:DGURRXWZ submitted 2025-05-27 cs.CL

Soichiro Murakami , Peinan Zhang , Hidetaka Kamigaito , Hiroya Takamura , Manabu Okumura This is my paper
classification cs.CL
keywords adtextgenerationparaphrasedatasethumanpreferenceannotationdirectoptimizationlinguisticfeatureanalysisJapaneseadvertisingclick-throughrateLLM-as-judge
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper introduces AdParaphrase v2.0, a Japanese ad-text paraphrase dataset of 16,460 pairs, each judged for attractiveness by ten crowd workers — roughly 20 times larger than its predecessor. The authors use the scale to isolate how-to-say effects: holding content fixed, they identify linguistic features (text length, noun count, low perplexity, specificity, brackets) that shift human preference, several unseen in v1.0. They then compare generation routes — in-context learning, instruction tuning, and direct preference optimization — and find DPO yields the most attractive texts while instruction-tuned models best preserve paraphrase validity and length constraints. They also report that majority human preference aligns with predicted click-through rate and, in a live A/B test, paraphrased headlines improved conversion metrics. The contribution is a reusable public benchmark plus evidence that preference data can guide both the generation and evaluation of attractive ad copy.

What carries the argument

The load-bearing object is a paraphrase pair (source ad text, rewritten ad text) that passes a crowd paraphrase-identification filter and carries a preference score from ten evaluators, with "skip" allowed when the two texts feel equally attractive. Because both texts share the same meaning, any preference difference is attributed to how-to-say rather than what-to-say. The dataset's strong-preference subset (3,570 pairs where at least eight of ten judges agree) is the engine of the feature analysis; the 8,721 triplets formed from source, preferred, and dispreferred paraphrases provide the DPO training signal. The pCTR alignment and reference-free LLM-as-judge correlations are supporting mechanisms linking the preference signal to ad performance and to cheaper evaluation.

What would settle it

Re-annotate a random sample of the 3,570 high-agreement pairs with a fresh panel of ten judges: if the majority preference flips or the kappa on that subset drops toward zero, the feature rankings and the DPO training signal rest on unstable labels. A second check is to confirm that the source ad text of no training triplet appears in the test split of the 8,721 triplets; same-source leakage would inflate the DPO evaluation.

Watch

Extended reading notes

Core claim

AdParaphrase v2.0 is large enough (16,460 pairs, 10 judgments each; 3,570 pairs with at least eight of ten judges agreeing) to support reliable linguistic analysis and model training. Using chi-square tests on the high-agreement subset, the paper claims attractive ad texts are longer, contain more nouns and noun phrases, more kanji and symbols, lower perplexity and shallower dependency trees, higher specificity, and brackets — whereas features such as adjective counts and dependency link length show no significant effect. Preference-tuned models (DPO) produce the most attractive outputs, but often violate the 30-character ad limit; feeding the feature findings into prompts raises attractiveness while respecting length. Human preference correlates with predicted CTR (Pearson coefficient 0.946) and, in a Google Ads A/B test, paraphrased headlines from two companies improve conversion metrics relative to baseline. Reference-free GPT-4o judgment correlates positively with human attractiveness evaluations, while BLEU and BERTScore correlate negatively, so the paper argues reference-free LLM metrics are better suited to evaluating ad appeal.

Load-bearing premise

The analysis depends on the assumption that crowd preference is a stable signal: although agreement is low overall (kappa = 0.167), the paper treats the 3,570 pairs with at least eight of ten judges on one side as capturing genuine attractiveness rather than noise or selection bias.

Editorial extensions

If this is right

  • Researchers can now train ad-text rewriting models on 8,721 preference triplets; the paper's DPO results promise gains in attractiveness, with instruction tuning remaining the safer choice when length constraints matter.
  • Prompt writers can apply the discovered features directly: longer text, more nouns, lower perplexity, higher specificity, and brackets are the paper's recipe for preferred ad copy.
  • Human preference is a proxy for performance: majority-preferred texts align with predicted CTR at r = 0.946 and with improved conversion metrics in live ads, so preference labels can serve as a development target when CTR data is unavailable.
  • Reference-free LLM evaluation can replace human attractiveness judgments at the system level; BLEU and BERTScore should not be used to rank ad appeal because they correlate negatively with human preference.

Reading between the lines

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

  • If the at-least-eight-of-ten agreement threshold is what makes the analysis work, an immediate test is whether DPO trained only on those 3,570 high-agreement pairs outperforms DPO trained on all 8,721 triplets; the paper does not report this comparison, and it would separate genuine signal from label noise.
  • The finding that human preference predicts pCTR only about 60 percent of the time even when all ten evaluators agree suggests an upper bound on how much copy style alone can move ad performance; content and placement presumably account for the rest, a consequence the paper states only implicitly.
  • The Japanese-specific features (hiragana, kanji, katakana, brackets) invite a cross-lingual version: building an English or Chinese dataset from scratch and testing whether length, specificity, and noun density replicate would show whether these are general copywriting principles or language-specific conventions.
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

5 major / 4 minor

Summary. The paper introduces AdParaphrase v2.0, a Japanese ad-text paraphrase dataset containing 16,460 paraphrase pairs, each with preference judgments from ten crowd workers. The dataset is constructed from CAMERA source ads via LLM and crowdworker paraphrasing, followed by paraphrase identification and preference annotation. The authors use the dataset to analyze linguistic features associated with human preference, to train and evaluate ICL, instruction-tuned, and DPO-based ad-text generators, and to relate human preferences to predicted and real ad performance. They also evaluate reference-free LLM-based metrics against human judgments. The central claims are that the dataset enables a more comprehensive and reliable analysis than v1.0, that several new linguistic features of attractive ad texts are identified, and that preference signals correlate with ad performance.

Significance. If the claims hold, AdParaphrase v2.0 is a valuable public resource: it is roughly 20 times larger than v1.0, is released under a permissive research license, and is constructed with a documented pipeline. The paper also contributes an interesting comparison of ICL, instruction tuning, and DPO for ad-text rewriting, and the finding that reference-free LLM judgments track human attractiveness ratings better than reference-based metrics is useful. However, the strength of the linguistic-feature findings is undercut by the low inter-annotator agreement on preference judgments and by the outcome-dependent manner in which the analysis subset is selected. The dataset release itself is the most solid contribution; the analytical conclusions require substantial additional validation.

major comments (5)
  1. [§4.2, §4.4, §5.1] The core feature analysis is built on an outcome-dependent subset. Overall preference judgments have a Fleiss' kappa of only 0.167 (§4.2), and the analysis then restricts to 3,570 pairs with at least 8 of 10 evaluators agreeing (§4.4). Because this agreement threshold is defined by the same preference votes that serve as the dependent variable in the chi-square tests of Table 4, the selection induces a truncation-on-outcome bias: features that make agreement easier, such as text length or bracket use, will appear associated with preference even if they have no effect on attractiveness. The paper does not address this bias, and Section 8 (Limitations) does not mention it. The authors should analyze the full 16,460 pairs with a model that accounts for rater disagreement, or validate the high-agreement subset on held-out preference labels.
  2. [Table 4 / Appendix F] The significance claims rest on 26 separate chi-square tests with a fixed p<0.01 threshold and no multiple-testing correction. Several tests also use very small samples, such as sadness with N=17 and question marks with N=78. With 26 comparisons, multiple 'significant' results would be expected by chance even if no feature truly mattered. Report corrected p-values (for example, Benjamini-Hochberg) or use a hierarchical model that pools information across features, and provide confidence intervals for the phi effect sizes.
  3. [§5.2] The finding-condition prompts (zeroshot-findings and fewshot-findings) are constructed directly from the outcome-selected analysis in §5.1. The improved attractiveness of these conditions over zeroshot in Table 5 can therefore reflect the selection mechanism rather than a genuine improvement in preference, because the prompts instruct the model to use brackets and longer text, which are exactly the features that drive agreement-based selection. The authors should evaluate the finding-condition generations against held-out human preference judgments from the full dataset, and compare against a control prompt that injects features selected identically but from a random split of the data.
  4. [§5.2.1 (Dataset)] The paper does not state whether the source ad text x is disjoint across the 9:0.5:0.5 train/dev/test split of the 8,721 DPO triplets. If the same source text appears in more than one partition, the DPO evaluation is contaminated by data leakage. The authors should clarify the splitting criterion; if the split is triplet-level only, they should re-split at the source-ad level.
  5. [§6.1–§6.2, Table 7] The online A/B evidence is mixed: for fitness ads, CTR decreased to 91.5 and CPC increased to 110.7 relative to baseline while CVR and CTVR improved; for education, the two-week run shows CTR 77.7 and the one-month run shows CTR 93.4, and only a subset of metrics is marked significant. Additionally, the pCTR analysis (§6.1) uses a model prediction rather than observed CTR, so the wording 'achieved higher CTRs' overstates what is measured. Please report confidence intervals and impression/click counts for all online metrics, and temper the conclusion to 'suggest' rather than 'demonstrate' a relationship between human preference and ad performance.
minor comments (4)
  1. [Figure 1] The x-axis labels and tick labels in Figure 1 appear garbled in the submitted PDF; please regenerate the figure so that the preference-vote values are readable.
  2. [§5.2.1] The split ratio '9 : 0.5 : 0.5' is non-standard; please state the actual numbers of triplets in the training, development, and test sets, or express the ratio as percentages.
  3. [Table 5] Human evaluation results are reported as point estimates without confidence intervals or significance tests, making comparisons such as CALM3-22B dpo-zeroshot 84.4 versus Swallow70B dpo-zeroshot 71.2 difficult to interpret; add error bars or significance tests, or state explicitly that the comparisons are descriptive only.
  4. [Table 8] The negative correlations of BLEU and BERTScore with attractiveness deserve a sentence of interpretation; as written, readers may not understand that reference-based metrics reward similarity to a reference paraphrase rather than attractiveness.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the dataset construction, feature analyses, and ATG evaluations are self-contained and externally evaluated.

full rationale

The paper makes no derivation claim in which a predicted quantity equals an input by construction. The linguistic-feature analysis in Section 5.1 is a descriptive chi-square association computed on the released preference data itself, not a prediction from a fitted parameter to a held-out quantity, so it cannot reduce to its inputs. The ATG experiment is an intervention: Section 5.2.2 compares zeroshot with zeroshot-findings, where the latter prompt incorporates findings from Section 5.1; the generated texts are then judged by independent human evaluators on PI and attractiveness. This comparison is not forced because the prompted model could fail to produce preferred outputs, and therefore it is not circular. The online A/B test in Section 6.2 uses actual platform metrics, and the pCTR analysis in Section 6.1 uses an independent in-house CTR prediction model. Self-citations to AdParaphrase v1.0 (Murakami et al., 2025) are used for annotation guidelines and few-shot paraphrase examples during construction and as a comparison baseline; they are not invoked to justify the new findings, so this same-author citation is not load-bearing. The paper itself discloses relevant limitations, including low preference IAA (kappa = 0.167 in Section 4.2), LLM-generated texts, and limited participant diversity. Methodological concerns such as restricting the feature analysis to the 3,570 high-agreement pairs selected on the preference outcome, and using GPT-4o both as a generator and as an automatic judge, are validity risks rather than definitional circularity: there is no equation-level identity between the selection rule and the chi-square results, nor between GPT-4o's generated outputs and the human preference labels used for evaluation. Under the stated standard of exhibiting a specific reduction, no circular step is present.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The core empirical conclusions rest on design choices and domain assumptions rather than on free parameters in a derivation. The most consequential choices are the high-agreement threshold, the 30-character rule, the p<0.01 significance level, and the assumption that crowd preference is a stable signal despite slight agreement.

free parameters (4)
  • High-agreement threshold for feature analysis = at least 8 of 10 evaluators
    Chosen by hand to define reliable preference. It selects 3,570 pairs and shapes all chi-square results in Section 5.1.
  • Preference pass threshold for judging generated texts = at least 8 of 10
    Used in Table 2 pass rates and to define attractive outputs in ATG human evaluation. Different thresholds would change reported rates.
  • Significance level for chi-square tests = p < 0.01
    No multiple-comparison correction across 26 features, so some flagged features may be false discoveries.
  • Train/dev/test split ratio = 9:0.5:0.5
    Unusual 90/5/5 split chosen for DPO triplets; no source-level deduplication is described.
assumptions (5)
  • domain assumption Crowd preferences are meaningful signals of ad attractiveness
    The whole dataset and DPO training rely on this. Section 4.2 reports Fleiss kappa 0.167, so the signal is weak.
  • domain assumption Paraphrase identification by majority of five workers preserves meaning
    If paraphrase pairs differ in content, linguistic feature effects are confounded. PI kappa is 0.442, moderate.
  • domain assumption External classifiers provide valid emotion and specificity labels
    LUKE emotion classifier accuracy 68.6% and GPT-4 specificity classifier 88% on 100 samples; label noise is not propagated.
  • domain assumption Chi-square tests on selected high-agreement pairs identify causal factors
    Correlational analysis with no multiple-testing correction is interpreted as factors influencing preference.
  • domain assumption In-house pCTR model is a valid proxy for ad performance
    Section 6.1 uses pCTR to link preference to clicks; the model is not described or released.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AdParaphrase v2.0: Generating Attractive Ad Texts Using a Preference-Annotated Paraphrase Dataset." pith.science (2026). https://pith.science/paper/DGURRXWZ

@misc{pith2026250520826,
  author       = {Pith},
  title        = {Pith review of: AdParaphrase v2.0: Generating Attractive Ad Texts Using a Preference-Annotated Paraphrase Dataset},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DGURRXWZ}},
  note         = {Machine review of arXiv:2505.20826}
}
read the original abstract

Identifying factors that make ad text attractive is essential for advertising success. This study proposes AdParaphrase v2.0, a dataset for ad text paraphrasing, containing human preference data, to enable the analysis of the linguistic factors and to support the development of methods for generating attractive ad texts. Compared with v1.0, this dataset is 20 times larger, comprising 16,460 ad text paraphrase pairs, each annotated with preference data from ten evaluators, thereby enabling a more comprehensive and reliable analysis. Through the experiments, we identified multiple linguistic features of engaging ad texts that were not observed in v1.0 and explored various methods for generating attractive ad texts. Furthermore, our analysis demonstrated the relationships between human preference and ad performance, and highlighted the potential of reference-free metrics based on large language models for evaluating ad text attractiveness. The dataset is publicly available at: https://github.com/CyberAgentAILab/AdParaphrase-v2.0.

Figures

Figures reproduced from arXiv: 2505.20826 by the authors.

Figure 1
Figure 1. Distribution of maximum number of votes between ad text pair in preference judgments. 4.2 Inter-Annotator Agreement Inter-annotator agreement (IAA) for PI (§3.3) and preference judgment (§3.4) was measured using Fleiss’ kappa (Fleiss et al., 1971). The kappa value for PI was 0.442, indicating moderate agreement, whereas that for preference judgment was 0.167, in￾dicating slight agreement (Landis and Koch, 1977). The… view at source ↗
Figure 2
Figure 2. Alignment between human preferences and predicted Click-Through Rate (pCTR). The x-axis in￾dicates human agreement level (number of evaluators with same preference). The y-axis shows the alignment ratio with pCTR. Higher human agreement correlates with increased alignment ratio, suggesting stronger con￾sensus means better alignment. tures incorporated into the prompt. The results in￾dicate that models with higher at… view at source ↗
Figure 3
Figure 3. Prompt for paraphrase candidate generation [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Guidelines for paraphrase candidate creation [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Guidelines for paraphrase identification pre [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Guidelines for preference judgments pre [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Prompt for fewshot-findings in ATG ex￾periment. (in characters) for ad text pairs, where Ad 1 and Ad 2 refer to the source and paraphrased ad text, re￾spectively. In [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. AI-Generated Content in Cross-Domain Applications: Research Trends, Challenges and Propositions

    cs.AI 2025-09 conditional novelty 2.0 of 10

    A cross-domain vision paper that surveys AI-generated content and proposes research directions, without introducing new empirical results.

Reference graph

Works this paper leans on

48 extracted references · 23 canonical work pages · cited by 1 Pith paper

  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]

    Kevin Bartz, Cory Barr, and Adil Aijaz. 2008. https://doi.org/10.1145/1386790.1386792 Natural language generation for sponsored-search advertisements . In Proceedings of the 9th ACM Conference on Electronic Commerce, pages 1--9

    work page arXiv 2008

  4. [4]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gr...

    work page 2020

  5. [5]

    Jan Cegin, Jakub Simko, and Peter Brusilovsky. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.117 C hat GPT to replace crowdsourcing of paraphrases for intent classification: Higher diversity and comparable model robustness . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 1889--1905

    doi:10.18653/v1/2023.emnlp-main.117 2023

  6. [6]

    Jacob Cohen. 1988. Statistical Power Analysis for the Behavioral Sciences, 2nd edition. Lawrence Erlbaum Associates, Hillsdale, NJ

    1988

  7. [7]

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. Qlora: efficient finetuning of quantized llms. In Advances in Neural Information Processing Systems 36

    work page 2023

  8. [8]

    Fleiss et al

    J.L. Fleiss et al. 1971. Measuring nominal scale agreement among many raters . Psychological Bulletin, 76(5):378--382

    work page 1971

Show all 48 references
  1. [9]

    Kazuki Fujii, Taishi Nakamura, Mengsay Loem, Hiroki Iida, Masanari Ohi, Kakeru Hattori, Hirai Shota, Sakae Mizuki, Rio Yokota, and Naoaki Okazaki. 2024. https://openreview.net/forum?id=TQdd1VhWbe Continual pre-training for cross-lingual LLM adaptation: Enhancing japanese langu...

    2024

  2. [10]

    Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Yinghan Shen, Shengjie Ma, Honghao Liu, Yuanzhuo Wang, and Jian Guo. 2025. https://arxiv.org/abs/2411.15594 A survey on llm-as-a-judge . Preprint, arXiv:2411.15594

    2025 arXiv

  3. [11]

    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 and Da...

    2019

  4. [12]

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

    2024

  5. [13]

    Aditi Jha, Sam Havens, Jeremy Dohmann, Alex Trott, and Jacob Portes. 2023. Limit: Less is more for instruction tuning across evaluation paradigms. arXiv preprint arXiv:2311.13133

    2023 arXiv

  6. [14]

    Tomoyuki Kajiwara, Chenhui Chu, Noriko Takemura, Yuta Nakashima, and Hajime Nagahara. 2021. https://doi.org/10.18653/v1/2021.naacl-main.169 WRIME : A new dataset for emotional intensity estimation with subjective and objective annotations . In Proceedings of the 2021 Conferenc...

    2021 doi

  7. [15]

    Hidetaka Kamigaito, Peinan Zhang, Hiroya Takamura, and Manabu Okumura. 2021. https://aclanthology.org/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 Assoc...

    2021

  8. [16]

    Jan-Christoph Klie, Richard Eckart de Castilho, and Iryna Gurevych. 2024. https://doi.org/10.1162/coli_a_00516 Analyzing dataset annotation quality management in the wild . Computational Linguistics, 50(3):817--866

    2024 doi

  9. [17]

    Wuwei Lan, Siyu Qiu, Hua He, and Wei Xu. 2017. https://doi.org/10.18653/v1/D17-1126 A continuously growing dataset of sentential paraphrases . In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1224--1234, Copenhagen, Denmark. Asso...

    2017 doi

  10. [18]

    Richard Landis and Gary G

    J. Richard Landis and Gary G. Koch. 1977. http://www.jstor.org/stable/2529310 The measurement of observer agreement for categorical data . Biometrics, 33(1):159--174

    1977

  11. [19]

    Vladimir I Levenshtein. 1966. Binary codes capable of correcting deletions, insertions and reversals. Soviet Physics Doklady, 10:707

    1966

  12. [20]

    Chin-Yew Lin. 2004. https://aclanthology.org/W04-1013 ROUGE : A package for automatic evaluation of summaries . In Proceedings of the ACL Workshop: Text Summarization Branches Out , pages 74--81

    2004

  13. [21]

    Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.153 G -eval: NLG evaluation using gpt-4 with better human alignment . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language ...

    2023 doi

  14. [22]

    Kikuo Maekawa, Makoto Yamazaki, Takehiko Maruyama, Masaya Yamaguchi, Hideki Ogura, Wakako Kashino, Toshinobu Ogiso, Hanae Koiso, and Yasuharu Den. 2010. http://www.lrec-conf.org/proceedings/lrec2010/pdf/99_Paper.pdf Design, compilation, and preliminary analyses of B alanced C ...

    2010

  15. [23]

    Shaunak Mishra, Manisha Verma, Yichao Zhou, Kapil Thadani, and Wei Wang. 2020. https://doi.org/10.1145/3340531.3412720 Learning to create better ads: Generation and ranking approaches for ad creative refinement . In Proceedings of the 29th ACM International Conference on Infor...

    2020

  16. [24]

    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 62th Annual Meeting of the Association for Computation...

    2024 doi

  17. [25]

    Soichiro Murakami, Sho Hoshino, and Peinan Zhang. 2023. https://arxiv.org/abs/2306.12719 Natural language generation for advertising: A survey . Preprint, arXiv:2306.12719

    2023 arXiv

  18. [26]

    Soichiro Murakami, Peinan Zhang, Sho Hoshino, Hidetaka Kamigaito, Hiroya Takamura, and Manabu Okumura. 2022. https://doi.org/10.18653/v1/2022.naacl-industry.9 Aspect-based analysis of advertising appeals for search engine advertising . In Proceedings of the 2022 Conference of ...

    2022 doi

  19. [27]

    Soichiro Murakami, Peinan Zhang, Hidetaka Kamigaito, Hiroya Takamura, and Manabu Okumura. 2025. https://arxiv.org/abs/2502.04674 A d P araphrase: Paraphrase dataset for analyzing linguistic features toward generating attractive ad texts . Preprint, arXiv:2502.04674

    2025 arXiv

  20. [28]

    OpenAI. 2024. https://openai.com/index/hello-gpt-4o/ Hello gpt-4o . Accessed: 2025-01-03

    2024

  21. [29]

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. https://aclanthology.org/P02-1040 BLEU : A method for automatic evaluation of machine translation . In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311--318

    2002

  22. [30]

    Ryan Park, Rafael Rafailov, Stefano Ermon, and Chelsea Finn. 2024. https://doi.org/10.18653/v1/2024.findings-acl.297 Disentangling length from quality in direct preference optimization . In Findings of the Association for Computational Linguistics: ACL 2024, pages 4998--5017

    2024 doi

  23. [31]

    Reid Pryzant, Sugato Basu, and Kazoo Sone. 2018. https://doi.org/10.18653/v1/W18-5415 Interpretable neural architectures for attributing an ad ' s performance to its writing style . In Proceedings of the 2018 EMNLP Workshop B lackbox NLP : Analyzing and Interpreting Neural Net...

    2018 doi

  24. [32]

    Manning, and Chelsea Finn

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. 2023. Direct preference optimization: your language model is secretly a reward model. In Advances in Neural Information Processing Systems 36

    2023

  25. [33]

    Satoshi Sato, Suguru Matsuyoshi, and Yohsuke Kondoh. 2008. http://www.lrec-conf.org/proceedings/lrec2008/pdf/165_paper.pdf Automatic assessment of J apanese text readability based on a textbook corpus . In Proceedings of the Sixth International Conference on Language Resources...

    2008

  26. [34]

    Victor O. Schwab. 2013. How to Write a Good Advertisement: A Short Course in Copywriting, illustrated edition edition. Echo Point Books & Media

    2013

  27. [35]

    Chantal Shaib, Yanai Elazar, Junyi Jessy Li, and Byron C Wallace. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.368 Detection and measurement of syntactic templates in generated text . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processin...

    2024 doi

  28. [36]

    Kazuma Takaoka, Sorami Hisamoto, Noriko Kawahara, Miho Sakamoto, Yoshitaka Uchida, and Yuji Matsumoto. 2018. Sudachi: a japanese tokenizer for business. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation

    2018

  29. [37]

    Hsueh-Cheng Wang and Marc Pomplun. 2012. https://doi.org/10.1167/12.6.26 The attraction of visual attention to texts in real-world scenes . Journal of Vision, 12(6):26--26

    2012 doi

  30. [38]

    Peiyi Wang, Lei Li, Liang Chen, Zefan Cai, Dawei Zhu, Binghuai Lin, Yunbo Cao, Lingpeng Kong, Qi Liu, Tianyu Liu, and Zhifang Sui. 2024. https://doi.org/10.18653/v1/2024.acl-long.511 Large language models are not fair evaluators . In Proceedings of the 62nd Annual Meeting of t...

    2024 doi

  31. [39]

    Xiting Wang, Xinwei Gu, Jie Cao, Zihua Zhao, Yulan Yan, Bhuvan Middha, and Xing Xie. 2021. https://doi.org/10.1145/3447548.3467105 Reinforcing pretrained models for generating attractive text advertisements . In Proceedings of the 27th ACM SIGKDD International Conference on Kn...

    2021

  32. [40]

    Jason Wei, Maarten Paul Bosma, Vincent Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew Mingbo Dai, and Quoc V. Le. 2022. https://openreview.net/forum?id=gEZrGCozdqR Finetuned language models are zero-shot learners . In The Tenth International Conference on Learnin...

    2022

  33. [41]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mari...

    2020 doi

  34. [42]

    Ikuya Yamada, Akari Asai, Hiroyuki Shindo, Hideaki Takeda, and Yuji Matsumoto. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.523 LUKE : Deep contextualized entity representations with entity-aware self-attention . In Proceedings of the 2020 Conference on Empirical Methods ...

    2020 doi

  35. [43]

    Brit Youngmann, Elad Yom-Tov, Ran Gilad-Bachrach, and Danny Karmon. 2020. https://doi.org/10.1145/3366423.3380211 The automated copywriter: Algorithmic rephrasing of health-related advertisements to improve their performance . In Proceedings of The Web Conference 2020, pages 1...

    2020

  36. [44]

    Yuan Yuan, Fengli Xu, Hancheng Cao, Guozhen Zhang, Pan Hui, Yong Li, and Depeng Jin. 2023. https://doi.org/10.1109/TKDE.2021.3110724 Persuade to click: Context-aware persuasion model for online textual advertisement . IEEE Transactions on Knowledge and Data Engineering, 35(2):...

    2023

  37. [45]

    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 The Eighth International Conference on Learning Representations

    2020

  38. [46]

    Yuan Zhang, Jason Baldridge, and Luheng He. 2019. https://doi.org/10.18653/v1/N19-1131 PAWS : Paraphrase adversaries from word scrambling . In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech...

    2019 doi

  39. [47]

    Jianing Zhou and Suma Bhat. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.414 Paraphrase generation: A survey of the state of the art . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5075--5086. Association for Computationa...

    2021 doi

  40. [48]

    Yaoming Zhu, Sidi Lu, Lei Zheng, Jiaxian Guo, Weinan Zhang, Jun Wang, and Yong Yu. 2018. https://doi.org/10.1145/3209978.3210080 Texygen: A benchmarking platform for text generation models . In Proceedings of the 41st International ACM SIGIR Conference on Research and Developm...

    2018

Pith tools

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