Intersectional Bias in Japanese Large Language Models from a Contextualized Perspective

Hitomi Yanaka; Jie Lu; Katsuhiko Watabe; Namgi Han; Ryoma Kumon; Sunjin Oh; Xinqi He; Yuko Itatsu; Yuma Matsuoka

REVIEW 4 major objections 6 minor 20 references

Intersectional Bias in Japanese Large Language Models from a Contextualized Perspective

T0 review · 4 major / 6 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 Context, not just identity, decides which group an LLM biases: the same mix of gender, nationality, and income is judged differently by GPT-4o and Swallow depending on whether the scenario is a marriage market or a PhD program.

desk verdict Useful Japanese benchmark for intersectional bias, but the headline claim about equal attribute combinations across contexts isn't tested cleanly; still deserves a serious referee. read the letter →

arxiv 2506.12327 v2 pith:OOWEXKLN submitted 2025-06-14 cs.CL cs.AI

Hitomi Yanaka , Xinqi He , Jie Lu , Namgi Han , Sunjin Oh , Ryoma Kumon , Yuma Matsuoka , Katsuhiko Watabe

show 1 more author

Yuko Itatsu

This is my paper

classification cs.CLcs.AI

keywords intersectionalbiasJapaneselanguagemodelsbenchmarkquestionansweringcontextualizedsocialattributesmicroaggressionsconstruction

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

The paper sets out to establish that the social bias of large language models is intersectional and contextual: the same combination of social attributes can be judged neutrally, positively, or negatively depending on the scenario in which it appears. To make this measurable, the authors build inter-JBBQ, a Japanese multiple-choice benchmark whose ambiguous questions pair profiles that differ by gender, nationality, age, income, and other attributes, and ask both stereotype-inviting and neutral questions. Using this benchmark on GPT-4o and the Swallow family, they report that models consistently favor a female profile over male and transgender profiles in marriage-market scenarios, but show no such preference, and even a reversed pattern, in academic-research scenarios. The upshot is that a model's prejudice is not a fixed property of a demographic group, so bias benchmarks that test attributes in isolation or in a single context will misdescribe how models discriminate.

What carries the argument

The load-bearing object is inter-JBBQ, a template-based Japanese question-answering benchmark. Each item has an ambiguous context for which the correct answer is unknowable, a disambiguated context that supplies the answer, a negative question that invites a stereotype-based answer, and a non-negative question asking the opposite; profile sentences for A and B vary the combination of social attributes while the question and template remain fixed. A model that answers A or B on an ambiguous item instead of the unknown label is treated as revealing the stereotype it attaches to that attribute combination. To separate stereotype from artifact, the authors randomize option order and balance five distinct unknown labels across items, and they compare the same gender contrast across topics and across one to five intersecting attributes.

What would settle it

Reverse the order of the two profiles in each template (putting the female profile in slot B instead of slot A) and rerun the benchmark; if the model's preference follows the slot rather than the gender category, the apparent stereotype-driven bias is an order artifact and the paper's central attribution does not hold. A second check: show human raters the ambiguous items and ask whether a correct answer is inferable; if raters can infer an answer, the template's 'unknown' premise fails.

Watch

Extended reading notes

Core claim

The central claim is that the influence of a particular gender category emerges in interaction with other social attributes and is further shaped by the broader social context. On inter-JBBQ's ambiguous marriage-market items, Swallow models almost always select the female candidate (A) over male or transgender candidates (B), which the authors interpret as the commodification of women in marriage-related activities. On the research PhD (humanities) items, the same female-versus-male comparison shows no consistent preference, and the direction of the model's choice shifts from female to male and then back to female as nationality and income attributes are added. The paper takes this as evidence that intersectional bias is constitutive and contextualized, not a stable group-level stereotype.

Load-bearing premise

The load-bearing premise is that inter-JBBQ's ambiguous questions are genuinely unanswerable and that a model choosing A or B instead of the unknown label reveals its social stereotypes, rather than a response to option order, wording quirks, or test-construction noise.

Editorial extensions

If this is right

If the paper is right, single-attribute bias benchmarks understate the problem: a model can appear neutral on gender alone while showing consistent bias on gender-plus-nationality or gender-plus-income.
Debiasing methods that work on single-attribute tests may not transfer, because the bias is not a stable property of an attribute but of an attribute-in-context.
Bias evaluations should report the topic or scenario as part of the score, and aggregated accuracy across topics can hide opposite biases that cancel each other out.
The observed reversal in research settings implies that intersectional bias is not always a simple stacking of stereotypes; it can change direction as attributes are added, so per-combination reporting is needed.
Culturally localized benchmarks matter: the same model shows a specifically Japanese social valuation (women as objects in the marriage market), which a generic English benchmark would not surface.

Reading between the lines

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

A consequence the authors leave implicit: rankings of models by overall bias score are not portable across domains; a model that looks best on a general benchmark could be the most biased in a high-stakes context like hiring or housing decisions.
A direct extension would translate inter-JBBQ into English and run it on English-centric models to separate the contribution of Japanese cultural priors in the training data from model-general stereotyping.
A testable refinement is to record the model's token probabilities rather than the final choice; if the bias is stereotype-driven, the probability gap between A and B should track the social stereotype rather than being a threshold effect.
A topic-ablation probe (neutral filler words instead of 'matchmaking event' or 'PhD program') would tell whether the context effect comes from the social scenario or from lexical associations with the topic words.

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.

Desk Editor's Note

read the letter

The real contribution is inter-JBBQ, a Japanese QA benchmark for intersectional bias built with sociologists and grounded in official Japanese statistics. That resource is genuinely useful, and the ambiguous/disambiguated context design follows the BBQ/JBBQ lineage sensibly. The evaluation across five Swallow variants and GPT-4o is careful: option order randomized, separate debias prompt condition, and results broken down by number of attributes.

The load-bearing claim—that biased output varies by context even with an equal combination of social attributes—is not cleanly supported. Table 4 compares female vs transgender in the marriage market with female vs male in research, so the contrast category and other attributes differ across contexts. The stress-test note is right about that. The appendix contains more detailed per-profile response tables, and some of them do hold the contrast category fixed across topics, but the paper never reports a matched comparison that actually controls the attribute set. Given the Limitations admission that there are too few items per attribute combination for statistical analysis, the Section 5 patterns are anecdotal, not established.

Two minor concerns: GPT-4o proofread the templates it is later evaluated on, which could introduce artifacts, and the dataset isn't released yet, so the resource isn't independently usable at this point.

To the paper's credit, the limitations are stated plainly, the construction process is thoughtful, and the benchmark concept is sound. The weakness is in the empirical claim, not the resource design. A serious referee should push for one of two fixes: release the data and add statistical validation, or soften the 'equal combination' claim and frame the qualitative patterns as hypotheses for future work.

Bottom line: this deserves peer review. It's a preliminary resource paper that needs revision, but the benchmark idea is solid and worth engaging with.

Referee Report

4 major / 6 minor

Summary. The paper introduces inter-JBBQ, a Japanese-language multiple-choice QA benchmark for evaluating intersectional bias in LLMs, with templates covering four topics (house renting, marriage market, research, social etiquette) and multiple social attributes such as gender, nationality, age, and sexuality. The dataset uses ambiguous and disambiguated contexts, negative and non-negative questions, and answer choices A, B, and an unknown label. The authors evaluate GPT-4o and four Swallow models (two sizes, with and without instruction tuning), reporting accuracy by topic and analyzing responses to ambiguous questions. They claim that biased output varies with context even when the combination of social attributes is the same, suggesting that LLM bias is context-dependent rather than a fixed property of a demographic group.

Significance. If the central claim is well supported, this paper would make an important contribution by demonstrating that single-attribute bias benchmarks are insufficient and that intersectional bias is context-dependent—a viewpoint grounded in intersectionality theory. The dataset itself, built with templates based on sociological literature and official statistics, is a potentially useful resource for the Japanese NLP community, and the paper's experimental design (option randomization, multiple unknown labels, both basic and debias prompts) is thoughtful. The authors are also candid about limitations. However, the empirical evidence for the headline claim is currently not robust: the key comparison in Table 4 confounds context with differences in attribute sets, and the qualitative analysis lacks statistical validation. The resource is valuable even if the empirical claim is subsequently weakened.

major comments (4)

[§5 and Table 4] The abstract's claim that biased output varies 'even with the equal combination of social attributes' is not supported by the evidence presented. The rows compared across the two topics do not share the same combination: in marriage market: matching rate the contrast is A: Female versus B: Transgender person, while in research: PhD (humanities) it is A: Female versus B: Male, and the accompanying attributes also differ (e.g., 'Female from the Philippines' versus 'Female from the US'; 10 million yen versus 5 million yen). Consequently, the observed cross-topic variation could be driven by the difference in the contrast category or the other attributes rather than by the broader context. To support the claim, the authors need either a matched comparison in which the identical attribute combination (including the contrast pair) appears in multiple contexts, or a careful control analysis that separates context effects from attribute-set effects.
[Limitations section and §5] The paper admits that 'our dataset does not contain enough numbers of questions corresponding to each attribute combination to allow statistical analysis to be conducted.' Yet the central qualitative conclusion—that the influence of a gender category 'emerges in interaction with other social attributes and is further shaped by the broader social context'—is drawn from single responses per profile in Tables 12–19, with no significance tests, confidence intervals, or repeated evaluations across random orderings. For example, Table 4 shows only three rows per topic, and the pattern in the research topic is based on three cases. At a minimum, the authors should report variance across repeated runs or option permutations and ideally provide a statistical test (e.g., a proportion test or bootstrap) for the claim that response patterns differ across contexts.
[§3.3] The dataset creation pipeline uses GPT-4o to proofread the problem text, and GPT-4o is also one of the models evaluated in §4 and §5. This creates a potential circularity: the benchmark may inadvertently be aligned with GPT-4o's own preferences, which could affect the very response patterns that Section 5 interprets as bias. The authors should either exclude GPT-4o from the qualitative bias analysis or demonstrate that the proofreading step did not systematically change the content of the templates (e.g., by comparing model outputs on pre-proofread and post-proofread versions). Without this, the GPT-4o results are hard to interpret as independent evidence.
[§3.2 and §5] The inference that choosing A or B on an ambiguous question reveals stereotype-based bias assumes that all other factors, especially position bias, are controlled. The paper randomizes option order across test instances, which is good, but the qualitative analysis in Section 5 reports a single deterministic response per profile (temperature 0 for Swallow models). A model could have a systematic first-option or second-option preference that is independent of the profile content, and with one response per profile there is no way to distinguish this from content-driven bias. The authors should report responses across multiple random option orders or provide per-profile distributions.

minor comments (6)

[Table 7] The topic label 'Rescarch: PhD (sciences)' contains a typo; it should read 'Research: PhD (sciences).'
[Table 7] In the house renting: security deposit row, the non-negative question is listed as 'Who was the one that paid the 5,000-yen participation fee?', which appears to be a copy-paste error from the marriage market template; it should ask about the security deposit or additional costs.
[§3.2] The example of profile sentences lists only four combinations for a two-attribute cross (gender x age); the full cross would include additional ordered pairs such as (30s female, 20s male). The example should be completed or labeled as a partial list.
[§3.2] The paper mentions 'five distinct unknown options' but does not specify what these are; providing them in an appendix would improve reproducibility.
[Appendix D] The caption says these are results with basic prompts, and the text says 'We checked for the same tendency with debias prompts,' but the debias-prompt tables are not shown; including them or clearly stating that they are omitted would help.
[§3.3] The manual checks by three researchers are described, but no inter-annotator agreement measure is reported; a simple agreement figure would help readers assess reliability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the empirical findings are self-contained; the main concerns are validity confounds, not circular derivation.

full rationale

This paper is an empirical benchmark study rather than a derivation. It constructs inter-JBBQ from templates and vocabulary lists, measures LLM accuracy and answer choices, and interprets cross-topic patterns. No parameter is fitted to the target claim, and no equation defines the conclusion in terms of the template inputs. The closest definitional moment is Section 3.2: 'By observing how model predictions change depending on the difference in intersectional attributes of the profile sentences in the same question, we can analyze the intersectional bias inherent in the model.' That is an operational definition of a measurement, not a circular derivation of the empirical finding. The central observation that models choose A or B in ambiguous contexts differently across topics is an independent model output. The abstract's phrase 'even with the equal combination of social attributes' is not fully controlled in Table 4, since the marriage-market rows compare Female vs Transgender while the research rows compare Female vs Male, and accompanying attributes also differ; this weakens the contextual-dependence interpretation, but it is a confound and external-validity concern, not a circularity. The paper's own Limitations section states that the dataset 'does not contain enough numbers of questions corresponding to each attribute combination to allow statistical analysis to be conducted,' which further supports treating the finding as preliminary rather than derived. Self-citations such as JBBQ (Yanaka et al., 2024) establish provenance but are not load-bearing for the empirical result. The use of GPT-4o to proofread templates that are later used to evaluate GPT-4o is a contamination risk, but it does not make the measured predictions logically equivalent to the template inputs. No circular step can be exhibited, so the appropriate finding is no significant circularity.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The central claim rests on three assumptions: the microaggression and intersectionality frameworks define what counts as bias, the correctness scheme (unknown for ambiguous items) is a valid bias probe, and the vocabulary lists faithfully represent Japanese social categories. None of these is derived in the paper. There are no numeric free parameters fitted to data.

assumptions (3)

domain assumption Intersectionality and microaggression frameworks (Collins and Bilge 2020; Sue and Spanierman 2020) define what counts as bias in the benchmark.
The benchmark operationalizes bias through a QA accuracy proxy based on these frameworks; if the frameworks do not map onto QA choice behavior, the metric loses validity. Section 3.1.
ad hoc to paper The unknown label is the correct answer for ambiguous questions, and A or B is correct for disambiguated questions, regardless of attribute combination.
This correctness scheme is imposed by construction (Section 3.2); the paper calls it 'the correct answer' without independent grounding. Models choosing A or B in ambiguous contexts are thereby classified as biased.
domain assumption Vocabulary lists derived from official statistics and literature are representative of Japanese social attribute distributions.
Appendix B; the relevance and realism of the profile sentences depend on this representativeness.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Intersectional Bias in Japanese Large Language Models from a Contextualized Perspective." pith.science (2026). https://pith.science/paper/OOWEXKLN

@misc{pith2026250612327,
  author       = {Pith},
  title        = {Pith review of: Intersectional Bias in Japanese Large Language Models from a Contextualized Perspective},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OOWEXKLN}},
  note         = {Machine review of arXiv:2506.12327}
}

read the original abstract

An increasing number of studies have examined the social bias of rapidly developed large language models (LLMs). Although most of these studies have focused on bias occurring in a single social attribute, research in social science has shown that social bias often occurs in the form of intersectionality -- the constitutive and contextualized perspective on bias aroused by social attributes. In this study, we construct the Japanese benchmark inter-JBBQ, designed to evaluate the intersectional bias in LLMs on the question-answering setting. Using inter-JBBQ to analyze GPT-4o and Swallow, we find that biased output varies according to its contexts even with the equal combination of social attributes.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 9 canonical work pages

[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]

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]

Nishant Balepur, Abhilasha Ravichander, and Rachel Rudinger. 2024. https://doi.org/10.18653/v1/2024.acl-long.555 Artifacts or abduction: How do LLM s answer multiple-choice questions without the question? In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10308--10330, Bangkok, Thailan...

doi:10.18653/v1/2024.acl-long.555 2024
[4]

Su Lin Blodgett, Solon Barocas, Hal Daum \'e III, and Hanna Wallach. 2020. https://doi.org/10.18653/v1/2020.acl-main.485 Language (technology) is power: A critical survey of bias in NLP . In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5454--5476, Online. Association for Computational Linguistics

doi:10.18653/v1/2020.acl-main.485 2020
[5]

Patricia Hill Collins and Sirma Bilge. 2020. Intersectionality. Polity Press

work page 2020
[6]

e Stat. https://www.e-stat.go.jp/stat-search/files?page=1&layout=datalist&toukei=00250011&tstat=000001012480&cycle=7&year=20230&month=0&tclass1=000001012481&tclass2val=0 Immigration control statistics 23-00-02: Number of foreign entrants by nationality/region and port

work page
[7]

Kazuki Fujii, Taishi Nakamura, Mengsay Loem, Hiroki Iida, Masanari Ohi, Kakeru Hattori, Hirai Shota, Sakae Mizuki, Rio Yokota, and Naoaki Okazaki. 2024. http://arxiv.org/abs/2404.17790 Continual pre-training for cross-lingual llm adaptation: Enhancing japanese language capabilities . arXiv:2404.17790

arXiv 2024
[8]

Gallegos, Ryan A

Isabel O. Gallegos, Ryan A. Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K. Ahmed. 2024. https://doi.org/10.1162/coli_a_00524 Bias and fairness in large language models: A survey . Computational Linguistics, 50(3):1097--1179

doi:10.1162/coli_a_00524 2024

Show all 20 references

[9]

Hannah Rose Kirk, Yennie Jun, Filippo Volpin, Haider Iqbal, Elias Benussi, Frederic Dreyer, Aleksandar Shtedritski, and Yuki Asano. 2021. https://proceedings.neurips.cc/paper_files/paper/2021/file/1531beb762df4029513ebf9295e0d34f-Paper.pdf Bias out-of-the-box: An empirical ana...

2021
[10]

John Lalor, Yi Yang, Kendall Smith, Nicole Forsgren, and Ahmed Abbasi. 2022. https://doi.org/10.18653/v1/2022.naacl-main.263 Benchmarking intersectional biases in NLP . In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Lin...

2022 doi
[11]

Brigitte Lhomond, Marie-Jos^^c3^^a8phe Saurel-Cubizolles, and Stuart Michaels. 2014. https://doi.org/10.1007/s10508-013-0124-y A multidimensional measure of sexual orientation, use of psychoactive substances, and depression: Results of a national survey on sexual behavior in f...

2014 doi
[12]

Weicheng Ma, Brian Chiang, Tong Wu, Lili Wang, and Soroush Vosoughi. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.575 Intersectional stereotypes in large language models: Dataset and analysis . In Findings of the Association for Computational Linguistics: EMNLP 2023, ...

2023 doi
[13]

Ministry of Health, Labour and Welfare . a . https://www.mhlw.go.jp/toukei/saikin/hw/k-tyosa/k-tyosa23/index.html 2023 national livelihood survey: Summary of results

2023
[14]

Ministry of Health, Labour and Welfare . b . https://www.mhlw.go.jp/toukei/itiran/roudou/chingin/kouzou/z2023/index.html Summary of the 2023 basic survey on wage structure

2023
[15]

https://www.soumu.go.jp/toukei_toukatsu/index/seido/shokgyou/kou_h21.htm Japan standard occupational classification: Classification item names

Ministry of Internal Affairs and Communications . https://www.soumu.go.jp/toukei_toukatsu/index/seido/shokgyou/kou_h21.htm Japan standard occupational classification: Classification item names
[16]

Alicia Parrish, Angelica Chen, Nikita Nangia, Vishakh Padmakumar, Jason Phang, Jana Thompson, Phu Mon Htut, and Samuel Bowman. 2022. https://doi.org/10.18653/v1/2022.findings-acl.165 BBQ : A hand-built bias benchmark for question answering . In Findings of the Association for ...

2022 doi
[17]

Audrey Smedley and Brian D Smedley. 2005. Race as biology is fiction, racism as a social problem is real: Anthropological and historical perspectives on the social construction of race. The American psychologist, 60(1):16--26

2005
[18]

Derald Wing Sue and Lisa Spanierman. 2020. Microaggressions in Everyday Life: Race, Gender, and Sexual Orientation. Wiley

2020
[19]

Toyotaro Suzumura, Akiyoshi Sugiki, Hiroyuki Takizawa, Akira Imakura, Hiroshi Nakamura, Kenjiro Taura, Tomohiro Kudoh, Toshihiro Hanawa, Yuji Sekiya, Hiroki Kobayashi, Yohei Kuga, Ryo Nakamura, Renhe Jiang, Junya Kawase, Masatoshi Hanai, Hiroshi Miyazaki, Tsutomu Ishizaki, Dai...

2022
[20]

Hitomi Yanaka, Namgi Han, Ryoma Kumon, Jie Lu, Masashi Takeshita, Ryo Sekizawa, Taisei Kato, and Hiromi Arai. 2024. https://arxiv.org/abs/2406.02050 Analyzing social biases in japanese large language models . arxiv:2406.02050

2024 arXiv

Pith tools

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

[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]

Nishant Balepur, Abhilasha Ravichander, and Rachel Rudinger. 2024. https://doi.org/10.18653/v1/2024.acl-long.555 Artifacts or abduction: How do LLM s answer multiple-choice questions without the question? In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10308--10330, Bangkok, Thailan...

doi:10.18653/v1/2024.acl-long.555 2024

[4] [4]

Su Lin Blodgett, Solon Barocas, Hal Daum \'e III, and Hanna Wallach. 2020. https://doi.org/10.18653/v1/2020.acl-main.485 Language (technology) is power: A critical survey of bias in NLP . In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5454--5476, Online. Association for Computational Linguistics

doi:10.18653/v1/2020.acl-main.485 2020

[5] [5]

Patricia Hill Collins and Sirma Bilge. 2020. Intersectionality. Polity Press

work page 2020

[6] [6]

e Stat. https://www.e-stat.go.jp/stat-search/files?page=1&layout=datalist&toukei=00250011&tstat=000001012480&cycle=7&year=20230&month=0&tclass1=000001012481&tclass2val=0 Immigration control statistics 23-00-02: Number of foreign entrants by nationality/region and port

work page

[7] [7]

Kazuki Fujii, Taishi Nakamura, Mengsay Loem, Hiroki Iida, Masanari Ohi, Kakeru Hattori, Hirai Shota, Sakae Mizuki, Rio Yokota, and Naoaki Okazaki. 2024. http://arxiv.org/abs/2404.17790 Continual pre-training for cross-lingual llm adaptation: Enhancing japanese language capabilities . arXiv:2404.17790

arXiv 2024

[8] [8]

Gallegos, Ryan A

Isabel O. Gallegos, Ryan A. Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K. Ahmed. 2024. https://doi.org/10.1162/coli_a_00524 Bias and fairness in large language models: A survey . Computational Linguistics, 50(3):1097--1179

doi:10.1162/coli_a_00524 2024