Pith. sign in

REVIEW 4 major objections 3 minor 51 references

Bias Attribution in Filipino Language Models: Extending a Bias Interpretability Metric for Application on Agglutinative Languages

T0 review · 4 major / 3 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 Filipino model bias is driven by people and objects, not actions

desk verdict First application of bias attribution to Filipino, but the unvalidated mean-of-subwords aggregation carries the whole English-vs-Filipino contrast, so the headline finding is not yet established. read the letter →

arxiv 2506.07249 v1 pith:JJYI2D4K submitted 2025-06-08 cs.CL

Lance Calvin Lim Gamboa , Yue Feng , Mark Lee This is my paper
classification cs.CL
keywords biasattributioninterpretabilityFilipinoagglutinativemorphologyCrowS-Pairstokenlanguagemodelmultilingualmodels
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 adapts an information-theoretic bias attribution score to work on agglutinative languages like Filipino by averaging the bias scores of a word's subword tokens. Applying this to four language models, it finds that biased behavior in Filipino-processing models is triggered most by words for people, objects, and relationships, contrasting with English models whose bias is triggered by action words. If correct, this shows that token-level bias interpretability must account for morphology and that the semantic triggers of model bias are not language-neutral.

What carries the argument

The central object is the bias attribution score $b(u) = \mathrm{pJSD}(P_{u,\text{more}} \| G_u) - \mathrm{pJSD}(P_{u,\text{less}} \| G_u)$, where $P_{u,\text{more}}$ and $P_{u,\text{less}}$ are the model's masked-token probability distributions in the stereotypical and less stereotypical versions of a CrowS-Pairs sentence, and $G_u$ is the one-hot ground truth. To adapt this to agglutinative Filipino, the paper defines the word-level score as the unweighted mean of its subword scores, $b(u) = \frac{1}{n}\sum_{i=1}^{n} b(t_i)$, where $t_1,\dots,t_n$ are the tokens produced by the model tokenizer for word $u$. This aggregation step is what lets the method attribute bias to whole words in a language where one word can contain multiple morphemes.

What would settle it

Compare the paper's word-level aggregation against a direct full-word masking experiment on the same Filipino CrowS-Pairs prompts: instead of masking each subword and averaging, mask the entire word and compute the bias score. If the full-word scores diverge in sign from the mean score, or if the top semantic categories change, the mean aggregation is not a faithful attribution.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that when the bias attribution score is aggregated from tokenizer subwords by a simple mean, the semantic categories that most drive biased choices in Filipino models are entity-based—people, concrete objects, and relationship words—whereas prior English results found action-based categories like criminal, sexual, and prosocial behaviors. This contrast suggests models handle sociodemographic associations differently depending on the language's morphology and the cultural content of the training data.

Load-bearing premise

The paper assumes that averaging subword bias scores without weighting gives a faithful word-level bias attribution; if subword contributions combine nonlinearly, the word-level scores and the semantic-category rankings derived from them would be unreliable.

Editorial extensions

If this is right

  • For agglutinative languages, token-level bias attribution scores should be aggregated over subwords rather than read per token.
  • Evaluating a multilingual model only on English bias benchmarks can miss language-specific bias triggers.
  • The method can be transferred to other agglutinative languages without retraining, provided a CrowS-Pairs-style benchmark exists.
  • Bias mitigation for Filipino models might target entity-related vocabulary rather than action verbs.

Reading between the lines

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

  • The unweighted mean may understate or overstate the contribution of rare, high-bias subwords; alternative aggregations (max, length-weighted) could be tested on the same data to see if the semantic ranking is stable.
  • If the semantic contrast is robust, it suggests that bias attribution findings from one language or culture should not be naively transferred to another; a similar adaptation in another agglutinative language would test whether the entity-versus-action difference comes from morphology or from the cultural content of the Filipino benchmark.
  • The method's reliance on translated semantic tags (Filipino to English via googletrans) could be replaced by a native Filipino tagger when one exists, which would test whether translation noise affects the category proportions.
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 / 3 minor

Summary. The paper adapts the bias attribution score of Gamboa and Lee (2024) to Filipino, an agglutinative language, by defining the attribution score of a multi-token word as the unweighted mean of its subword scores (Section 4.4). The adapted method is applied to four models—a Filipino masked LM, a causal LM described as trained on worldwide languages, and two Southeast Asian causal LMs—evaluated on Filipino CrowS-Pairs (Gamboa and Lee, 2025). Tokens are machine-translated to English and semantically tagged with pymusas, and the semantic categories with the largest proportions of bias-increasing tokens are reported in Table 5. The authors conclude that Filipino models are driven toward bias by entity-based words (people, objects, relationships), in contrast to the action-based categories (criminal, intimate, prosocial behaviors) reported for English in prior work.

Significance. If the central result were established, this would be a valuable first step toward morphology-aware bias interpretability for low-resource agglutinative languages, with testable predictions about which word classes contribute to stereotypical associations. The paper also provides code, uses a culturally adapted benchmark, and makes a concrete cross-linguistic comparison. However, the adaptation is currently asserted rather than derived or validated, and several pipeline choices—machine-translation-based semantic tagging, an unspecified causal-LM procedure, and a cross-paper English comparison—are load-bearing for the headline claim. The empirical contribution is therefore promising but not yet convincing.

major comments (4)
  1. [Section 4.4] The aggregation b(u) = (1/n) Σ b(t_i) is asserted without derivation or validation. The quantity being averaged is the model's conditional next-token distribution for a masked subword; within a longer agglutinative word, that distribution can be dominated by morphological predictability (e.g., after a prefix, the root subword is highly predictable regardless of stereotype), so per-subword scores may measure word-internal predictability rather than bias-driven contextual preference. The unweighted mean also dilutes semantically loaded roots with near-zero function morphemes. No alternative aggregation (sum, max, root-only) is compared, and no human or behavioral validation is reported; the Limitations section does not acknowledge this threat. Because every word-level score in Tables 3–5 and hence the abstract's English-vs-Filipino contrast is built on this aggregation, this is a load-bearing gap.
  2. [Section 4.3 and Table 2] The method description assumes a masked language model: P_u,more is described as the probability distribution computed for <MASK> when the token is masked. Three of the four models (gpt2, sea-lion-3b, SeaLLMs-v3-7B-Chat) are causal language models without a <MASK> token, and the paper never specifies how masked-token distributions or the ground-truth one-hot distributions are obtained for these models. This makes the bias scores in Table 2 and the attribution scores in Tables 4 and 5 for these models irreproducible. The authors should either restrict the analysis to masked LMs or provide the exact causal-LM protocol.
  3. [Section 4.5 and Table 5] The semantic categories are obtained by translating Filipino words to English with googletrans and then tagging with pymusas. No evaluation of translation accuracy on the 204 prompt pairs is provided, and no human verification of the assigned semantic fields is reported; the Limitations section acknowledges possible inaccuracies but does not connect them to the load-bearing semantic-category claim. In addition, Table 5 reports only proportions, without counts, confidence intervals, or significance tests; several entries (e.g., 'Weapons' 66.67%, 'People: female' 80.00%) are likely based on very few words, making the category rankings unstable. The 1% frequency threshold is also adopted without justification and directly determines which categories enter Table 5.
  4. [Section 5.2] The cross-linguistic contrast with English is based entirely on prior work (Gamboa and Lee, 2024), but the paper does not report whether that English analysis used the same subword aggregation, the same semantic tagger, or the same frequency threshold. Since the adapted method is new and the Filipino pipeline includes machine translation, the observed 'entity-based vs. action-based' difference could be an artefact of the different pipelines rather than a genuine linguistic difference. The authors should re-run the English analysis with the identical pipeline or clearly quantify the pipeline sensitivity.
minor comments (3)
  1. [Table 2] The gpt2 model is described as trained on 'languages worldwide,' but GPT-2 is an English-only model trained on WebText; this should be corrected, and the appropriateness of using GPT-2 for Filipino prompts should be justified.
  2. [Table 5] There is a typo in the roberta-tagalog-base column: 'Languauge' should be 'Language'; please proofread the manuscript for similar errors.
  3. [Section 6] The final sentence of the conclusion contains an incomplete parenthetical citation: 'and working to reduce their toxic and harmful effects (e.g., Liu et al., 2024; Ermis et al., 2024; Gupta et al., 2025' is missing a closing parenthesis and a period.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found; the Filipino bias-attribution analysis is computed directly from model probabilities and the semantic-category proportions are measured outputs, not fitted inputs.

full rationale

The paper's central claim is that Filipino models are driven toward bias by entity-based words (people, objects, relationships) rather than the action-heavy categories found in English. This claim is supported by a concrete computation chain: b(t_i) is obtained from the model's next-token or masked-token distributions compared with a one-hot ground truth via Jensen-Shannon distance (Eq. 1), then word-level scores are aggregated as b(u) = (1/n) * sum b(t_i) in Section 4.4. The semantic proportions in Table 5 are tabulated from these scores. No parameter is fitted to the semantic labels, and the semantic categories are not used to define or post-hoc adjust the scores. The b(u) = mean(subword scores) rule is an unvalidated and potentially fragile modeling choice, and the English comparison relies on the authors' prior Gamboa and Lee (2024) findings, but neither constitutes circularity: the rule is an explicit operationalization rather than a hidden restatement of the conclusion, and the prior English result is used as an external contrast rather than as the mechanism producing the Filipino scores. Self-citations for the metric and the Filipino CrowS-Pairs benchmark are present, but the underlying measurements in this paper are new computations on four models and would stand or fall independently of the citations' rhetorical framing. The Limitations section acknowledges translation and model-selection limitations but does not discuss the aggregation rule; that is a completeness gap, not a circularity step. Because the headline result is not equivalent by construction to any fitted parameter or self-cited theorem, the circularity score is 0.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The paper's empirical claims rest on five assumptions: two inherited from prior work (bias attribution score validity, CrowS-Pairs cultural validity), the unvalidated mean aggregation, the machine translation semantic pipeline, and a post-hoc frequency threshold. The only fitted or selected numeric parameter is the 1% word frequency cutoff.

free parameters (1)
  • minimum word frequency threshold = 1% of dataset total word count (words occurring fewer than 10 times excluded)
    Section 4.5 removes all words occurring less than 10 times from the semantic analysis. This post-hoc cutoff affects every reported semantic category proportion in Table 5 and is not statistically justified.
assumptions (5)
  • domain assumption The bias attribution score b(u) from Gamboa and Lee (2024) correctly measures a token's contribution to model bias.
    Eq. (1) treats JSD distance between mask-filling distributions and a one-hot ground truth as a measure of token-level bias contribution; this is an inherited theoretical assumption, not proven in this paper.
  • ad hoc to paper The unweighted mean of subword token scores equals the word-level bias attribution score.
    Section 4.4 defines b(u) = (1/n) * sum b(t_i) with no derivation or validation; the central results depend on this assumption.
  • domain assumption English translations of Filipino words preserve their semantic fields for pymusas tagging.
    Section 4.5 translates all words with googletrans before tagging with pymusas; machine translation can distort meaning, and the authors acknowledge this limitation.
  • ad hoc to paper The 1% word frequency threshold yields a stable and representative set of words for semantic analysis.
    Section 4.5 excludes words occurring fewer than 10 times; the cutoff is arbitrary and affects all reported proportions.
  • domain assumption Filipino CrowS-Pairs provides culturally valid bias measurements.
    Section 4.1 assumes the benchmark from Gamboa and Lee (2025) is contextually appropriate; this is a self-cited dataset with no external validation in this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bias Attribution in Filipino Language Models: Extending a Bias Interpretability Metric for Application on Agglutinative Languages." pith.science (2026). https://pith.science/paper/JJYI2D4K

@misc{pith2026250607249,
  author       = {Pith},
  title        = {Pith review of: Bias Attribution in Filipino Language Models: Extending a Bias Interpretability Metric for Application on Agglutinative Languages},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JJYI2D4K}},
  note         = {Machine review of arXiv:2506.07249}
}
read the original abstract

Emerging research on bias attribution and interpretability have revealed how tokens contribute to biased behavior in language models processing English texts. We build on this line of inquiry by adapting the information-theoretic bias attribution score metric for implementation on models handling agglutinative languages, particularly Filipino. We then demonstrate the effectiveness of our adapted method by using it on a purely Filipino model and on three multilingual models: one trained on languages worldwide and two on Southeast Asian data. Our results show that Filipino models are driven towards bias by words pertaining to people, objects, and relationships, entity-based themes that stand in contrast to the action-heavy nature of bias-contributing themes in English (i.e., criminal, sexual, and prosocial behaviors). These findings point to differences in how English and non-English models process inputs linked to sociodemographic groups and bias.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 28 canonical work pages

  1. [1]

    AI Singapore . 2023. https://github.com/aisingapore/sealion SEA-LION ( Southeast Asian Languages In One Network ): A family of large language models for Southeast Asia

    work page 2023

  2. [2]

    Giuseppe Attanasio, Debora Nozza, Eliana Pastor, and Dirk Hovy. 2022. https://doi.org/10.18653/v1/2022.nlppower-1.11 Benchmarking post-hoc interpretability approaches for transformer-based misogyny detection . In Proceedings of NLP Power! The First Workshop on Efficient Benchmarking in NLP, pages 100--112, Dublin, Ireland. Association for Computational Li...

    work page doi:10.18653/v1/2022.nlppower-1.11 2022

  3. [3]

    o rn Gamb \

    Selma Bergstrand and Bj \"o rn Gamb \"a ck. 2024. https://doi.org/10.18653/v1/2024.gebnlp-1.22 Detecting and mitigating LGBTQIA + bias in large N orwegian language models . In Proceedings of the 5th Workshop on Gender Bias in Natural Language Processing (GeBNLP), pages 351--364, Bangkok, Thailand. Association for Computational Linguistics

    work page doi:10.18653/v1/2024.gebnlp-1.22 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]

    coming out

    Bradley Cardozo. 2014. A "coming out" party in Congress ? LGBT advocacy and party-list politics in the Philippines . Master's thesis, University of California, Los Angeles

    work page 2014

  6. [6]

    Hanjie Chen, Guangtao Zheng, and Yangfeng Ji. 2020. https://doi.org/10.18653/v1/2020.acl-main.494 Generating hierarchical explanations on text classification via feature interaction detection . In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5578--5593, Online. Association for Computational Linguistics

    doi:10.18653/v1/2020.acl-main.494 2020

  7. [7]

    Kate Crawford. 2017. The trouble with bias. In Conference on Neural Information Processing Systems, invited speaker

    work page 2017

  8. [8]

    Jan Christian Blaise Cruz and Charibeth Cheng. 2022. https://aclanthology.org/2022.lrec-1.703 Improving large-scale language models and resources for F ilipino . In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 6548--6555, Marseille, France. European Language Resources Association

    work page 2022

Show all 51 references
  1. [9]

    Endres and J.E

    D.M. Endres and J.E. Schindelin. 2003. https://doi.org/10.1109/TIT.2003.813506 A new metric for probability distributions . IEEE Transactions on Information Theory, 49(7):1858--1860

    2003

  2. [10]

    Beyza Ermis, Luiza Pozzobon, Sara Hooker, and Patrick Lewis. 2024. https://doi.org/10.18653/v1/2024.findings-acl.893 From one to many: Expanding the scope of toxicity mitigation in language models . In Findings of the Association for Computational Linguistics: ACL 2024, pages ...

    2024 doi

  3. [11]

    Schmeisser-Nieto, Emma Raimundo Schulz, Thiziri Saci, Sarah Saidi, and 4 others

    Karen Fort, Laura Alonso Alemany, Luciana Benotti, Julien Bezan c on, Claudia Borg, Marthese Borg, Yongjian Chen, Fanny Ducel, Yoann Dupont, Guido Ivetta, Zhijian Li, Margot Mieskes, Marco Naguib, Yuyan Qian, Matteo Radaelli, Wolfgang S. Schmeisser-Nieto, Emma Raimundo Schulz,...

    2024

  4. [12]

    Steinunn Rut Fri riksd \'o ttir and Hafsteinn Einarsson. 2024. https://aclanthology.org/2024.lrec-main.671/ Gendered grammar or ingrained bias? exploring gender bias in I celandic language models . In Proceedings of the 2024 Joint International Conference on Computational Ling...

    2024

  5. [13]

    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

    2024 doi

  6. [14]

    Lance Gamboa and Mark Lee. 2024. https://arxiv.org/abs/2410.15464 A novel interpretability metric for explaining bias in language models: Applications on multilingual models from southeast asia . In Proceedings of the 38th Pacific Asia Conference on Language, Information and C...

    2024 arXiv

  7. [15]

    Lance Calvin Lim Gamboa and Mark Lee. 2025. https://aclanthology.org/2025.loreslm-1.9/ F ilipino benchmarks for measuring sexist and homophobic bias in multilingual language models from S outheast A sia . In Proceedings of the First Workshop on Language Models for Low-Resource...

    2025

  8. [16]

    J. Neil C. Garcia. 1996. Philippine Gay Culture: Binabae to Bakla, Silahis to MSM. Hong Kong University Press

    1996

  9. [17]

    Jonathan Gerona, Dörte de Kok, Christos Salis, Janet Webster, and Roel Jonkers and. 2025. https://doi.org/10.1080/02687038.2024.2353164 Characterization of agrammatism in tagalog: Evidence from narrative spontaneous speech . Aphasiology, 39(3):385--417

    2025

  10. [18]

    Daniela Godoy and Antonela Tommasel. 2021. Is my model biased? Exploring unintended bias in misogyny detection tasks. In AIofAI 2021: 1st Workshop on Adverse Impacts and Collateral Effects of Artificial Intelligence Technologies, volume 2942 of CEUR Workshop Proceedings, pages...

    2021

  11. [19]

    Riccardo Guidotti, Anna Monreale, Salvatore Ruggieri, Franco Turini, Fosca Giannotti, and Dino Pedreschi. 2018. https://doi.org/10.1145/3236009 A survey of methods for explaining black box models . ACM Comput. Surv., 51(5)

    2018 doi

  12. [20]

    Soumyajit Gupta, Venelin Kovatchev, Anubrata Das, Maria De-Arteaga, and Matthew Lease. 2025. https://arxiv.org/abs/2204.07661 Finding pareto trade-offs in fair and accurate detection of toxic speech . Preprint, arXiv:2204.07661

    2025 arXiv

  13. [21]

    Vipul Gupta, Pranav Narayanan Venkit, Shomir Wilson, and Rebecca Passonneau. 2024. https://doi.org/10.18653/v1/2024.gebnlp-1.19 Sociodemographic bias in language models: A survey and forward path . In Proceedings of the 5th Workshop on Gender Bias in Natural Language Processin...

    2024 doi

  14. [22]

    Haoyang Huang, Tianyi Tang, Dongdong Zhang, Xin Zhao, Ting Song, Yan Xia, and Furu Wei. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.826 Not all languages are created equal in LLM s: Improving multilingual capability by cross-lingual-thought prompting . In Findings of...

    2023 doi

  15. [23]

    Yufei Huang and Deyi Xiong. 2024. https://aclanthology.org/2024.lrec-main.260/ CBBQ : A C hinese bias benchmark dataset curated with human- AI collaboration for large language models . In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Lang...

    2024

  16. [24]

    Katsumi Ibaraki, Winston Wu, Lu Wang, and Rada Mihalcea. 2024. https://aclanthology.org/2024.lrec-main.86/ Analyzing occupational distribution representation in J apanese language models . In Proceedings of the 2024 Joint International Conference on Computational Linguistics, ...

    2024

  17. [25]

    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

  18. [26]

    Hwaran Lee, Seokhee Hong, Joonsuk Park, Takyoung Kim, Meeyoung Cha, Yejin Choi, Byoungpil Kim, Gunhee Kim, Eun-Ju Lee, Yong Lim, Alice Oh, Sangchul Park, and Jung-Woo Ha. 2023. https://doi.org/10.18653/v1/2023.acl-long.370 SQ u AR e: A large-scale dataset of sensitive question...

    2023 doi

  19. [27]

    J. Lin. 1991. https://doi.org/10.1109/18.61115 Divergence measures based on the Shannon entropy . IEEE Transactions on Information Theory, 37(1):145--151

    1991 doi

  20. [28]

    Zachary C. Lipton. 2018. https://doi.org/10.1145/3236386.3241340 The mythos of model interpretability: In machine learning, the concept of interpretability is both important and slippery. Queue, 16(3):31–57

    2018

  21. [29]

    Yan Liu, Yu Liu, Xiaokang Chen, Pin-Yu Chen, Daoguang Zan, Min-Yen Kan, and Tsung-Yi Ho. 2024. The devil is in the neurons: Interpreting and mitigating social biases in language models. In The Twelfth International Conference on Learning Representations

    2024

  22. [30]

    Nikita Nangia, Clara Vania, Rasika Bhalerao, and Samuel R. Bowman. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.154 C row S -pairs: A challenge dataset for measuring social biases in masked language models . In Proceedings of the 2020 Conference on Empirical Methods in Na...

    2020 doi

  23. [31]

    Rodrigo Navarro. 2024. https://www.electronicshub.org/generative-ai-global-interest-report-2023/ Generative AI global interest report

    2024

  24. [32]

    Thomas E. Payne. 2017. Morphological typology. In Alexandra Y. Aikhenvald and R. M. W.Editors Dixon, editors, The Cambridge Handbook of Linguistic Typology, Cambridge Handbooks in Language and Linguistics, page 78–94. Cambridge University Press

    2017

  25. [33]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, and 1 others. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9

    2019

  26. [34]

    Paul Rayson, Dawn E Archer, Scott L Piao, and Tony McEnery. 2004. The UCREL semantic analysis system. In Proceedings of the workshop on Beyond Named Entity Recognition Semantic labelling for NLP tasks, in association with LREC-04, pages 7--12. European Language Resources Association

    2004

  27. [35]

    Manon Reusens, Philipp Borchert, Margot Mieskes, Jochen De Weerdt, and Bart Baesens. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.175 Investigating bias in multilingual language models: Cross-lingual transfer of debiasing techniques . In Proceedings of the 2023 Conference...

    2023 doi

  28. [36]

    Why should i trust you?

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. https://doi.org/10.1145/2939672.2939778 " Why should i trust you?": Explaining the predictions of any classifier . In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Minin...

    2016

  29. [37]

    Julian Risch, Robin Ruff, and Ralf Krestel. 2020. https://aclanthology.org/2020.trac-1.22 Offensive language detection explained . In Proceedings of the Second Workshop on Trolling, Aggression and Cyberbullying, pages 137--143, Marseille, France. European Language Resources As...

    2020

  30. [38]

    Santiago and Norma G

    Alfonso O. Santiago and Norma G. Tiangco. 2003. Modern Filpino Grammar: New Edition 2003 . REX book Store

    2003

  31. [39]

    Lilia Quindoza Santiago. 1996. Patriarchal discourse in language and literature. In Pamela C. Constantino and Monico M. Atienza, editors, Selected Discoruses on Language and Society. University of the Philippines Press, Quezon City

    1996

  32. [40]

    Sujan Sarkar. 2023. https://writerbuddy.ai/blog/ai-industry-analysis AI industry analysis: 50 most visited AI tools and their 24B+ traffic behavior

    2023

  33. [41]

    Paul Schachter and Lawrence Reid. 2008. https://scholarspace.manoa.hawaii.edu/server/api/core/bitstreams/8e7f819f-68be-4cb3-95a3-18a057f436fe/content Tagalog . In Bernard Comrie, editor, The World's Major Languages, pages 833--855. Routledge

    2008

  34. [42]

    Victor Steinborn, Philipp Dufter, Haris Jabbar, and Hinrich Schuetze. 2022. https://doi.org/10.18653/v1/2022.findings-naacl.69 An information-theoretic approach and dataset for probing gender stereotypes in multilingual masked language models . In Findings of the Association f...

    2022 doi

  35. [43]

    Ahmet \"U st \"u n, Viraat Aryabumi, Zheng Yong, Wei-Yin Ko, Daniel D ' souza, Gbemileke Onilude, Neel Bhandari, Shivalika Singh, Hui-Lee Ooi, Amr Kayid, Freddie Vargus, Phil Blunsom, Shayne Longpre, Niklas Muennighoff, Marzieh Fadaee, Julia Kreutzer, and Sara Hooker. 2024. ht...

    2024 doi

  36. [44]

    Elly van Gelderen . 2006. A History of the English Language. John Benjamins Publishing Company, Amsterdam, Netherlands

    2006

  37. [45]

    That’s My Tomboy

    Gina Velasco. 2022. https://www.jstor.org/stable/48656853 “That’s My Tomboy” : Queer Filipinx diasporic transmasculinities . Alon: Journal for Filipinx American and Diasporic Studies, 2(1):67--73

    2022

  38. [46]

    Tong Xiang, Sean MacAvaney, Eugene Yang, and Nazli Goharian. 2021. https://aclanthology.org/2021.wassa-1.1 T ox CCI n: Toxic content classification with interpretability . In Proceedings of the Eleventh Workshop on Computational Approaches to Subjectivity, Sentiment and Social...

    2021

  39. [47]

    Sean Xie, Soroush Vosoughi, and Saeed Hassanpour. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.261 Proto-lm: A prototypical network-based framework for built-in interpretability in large language models . In Findings of the Association for Computational Linguistics: E...

    2023 doi

  40. [48]

    Wenxuan Zhang, Hou Pong Chan, Yiran Zhao, Mahani Aljunied, Jianyu Wang, Chaoqun Liu, Yue Deng, Zhiqiang Hu, Weiwen Xu, Yew Ken Chia, Xin Li, and Lidong Bing. 2024. https://arxiv.org/abs/2407.19672 SeaLLMs 3: Open foundation and chat multilingual large language models for South...

    2024 arXiv

  41. [49]

    Jun Zhao, Zhihao Zhang, Luhui Gao, Qi Zhang, Tao Gui, and Xuanjing Huang. 2024. https://arxiv.org/abs/2401.01055 Llama beyond english: An empirical study on language capability transfer . Preprint, arXiv:2401.01055

    2024 arXiv

  42. [50]

    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...

  43. [51]

    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.