pith. sign in

arxiv: 2505.18466 · v2 · submitted 2025-05-24 · 💻 cs.CL

Purdah and Patriarchy: Evaluating and Mitigating South Asian Biases in Open-Ended Multilingual LLM Generations

Mamnuya Rinki , Chahat Raj , Anjishnu Mukherjee , Ziwei Zhu This is my paper

Pith reviewed 2026-05-19 14:05 UTC · model grok-4.3

classification 💻 cs.CL
keywords multilingual LLMintersectional biascultural biasself-debiasingSouth Asian languagespurdahpatriarchybias lexicon
0
0 comments X

The pith

LLMs reinforce purdah and patriarchy biases in open-ended generations across ten South Asian languages, which self-debiasing can mitigate.

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

This paper examines how large language models embed culturally specific biases from South Asian traditions of purdah and patriarchy when producing text in multiple languages. The authors build a specialized lexicon to detect intersectional biases involving gender, religion, marital status, and family size in tasks like storytelling and planning. They measure these biases in ten Indo-Aryan and Dravidian languages and test whether simple or complex self-debiasing prompts reduce the problematic patterns. A sympathetic reader would care because these subtle reinforcements can perpetuate social stigmas in AI tools used daily in diverse regions. The work provides a framework for detecting and addressing such biases beyond typical Western-focused evaluations.

Core claim

We construct a culturally grounded bias lexicon for intersectional dimensions of gender, religion, marital status, and number of children to quantify how LLMs reinforce stigmas influenced by purdah and patriarchy in open-ended generations across 10 languages, and demonstrate that self-debiasing strategies reduce these biases.

What carries the argument

A culturally grounded bias lexicon that identifies intersectional biases in multilingual LLM outputs.

If this is right

  • Quantified bias levels vary across languages and generation types such as storytelling and to-do lists.
  • Both simple and complex self-debiasing prompts lower the detected intersectional bias.
  • The lexicon enables evaluation beyond Eurocentric settings.
  • Open-ended tasks reveal subtle biases that standard metrics miss.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Developers could integrate similar lexicons into training to prevent bias propagation in South Asian contexts.
  • Similar approaches might apply to other cultural biases like those in Middle Eastern or African societies.
  • Future work could test if reduced bias in generations leads to better user trust in LLM applications.

Load-bearing premise

The bias lexicon accurately reflects real cultural stigmas and that changes after debiasing prompts indicate actual bias reduction rather than just prompt following.

What would settle it

Running the same generations with and without debiasing and finding no measurable difference in bias scores using the lexicon, or human annotators from the region identifying missed stigmas in the lexicon.

Figures

Figures reproduced from arXiv: 2505.18466 by Anjishnu Mukherjee, Chahat Raj, Mamnuya Rinki, Ziwei Zhu.

Figure 1
Figure 1. Figure 1: Pipeline with templates and debiasing prompts (output arrows: black=original, red=simple, blue=complex). [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Identities and Their Highest Bias TF-IDF [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Average Gender Bias Score by Language Family. female single mothers are associated with isolate, showing social seclusion due to stigmatized pre￾marital childbearing. Muslim males with many children show lower overall bias, with terms like confident and bad. 5.1.2 Hobbies and Values [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Average Religion Bias Score by Language Family. 5.2 Bias Analysis Across Identity Dimensions We compute average bias scores across gender, re￾ligion, marital status, and children for Indo-Aryan and Dravidian languages in original prompts. This analysis leverages our novel bias lexicon to detect South Asian bias, expose open-ended application variations, and surface overlooked cultural stigmas. 5.2.1 Gender… view at source ↗
Figure 8
Figure 8. Figure 8: Average Child Count Bias Score by Language [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: mT5 Model Failure: Generates sentinel to [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Aya Model Failure: Ignores instructions to [PITH_FULL_IMAGE:figures/full_fig_p016_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Indic-Gemma Model Failure: Generates nonsensical outputs of mixed languages. E Bias Lexicon The lexicon was constructed through a comprehen￾sive review of existing literature on gender roles, religion, marital status, and societal expectations. This process involved identifying and categorizing terms that reflect biases, stereotypes, and social stigmas, with an emphasis on South Asian cultural contexts. T… view at source ↗
Figure 14
Figure 14. Figure 14: A Muslim male divorced with one child is [PITH_FULL_IMAGE:figures/full_fig_p019_14.png] view at source ↗
Figure 13
Figure 13. Figure 13: A Hindu female married woman with many children is associated with “clean.” What hobbies and personal values is a Muslim Female that is Widowed with Zero Children likely to have? Answer in Bengali. বন্দুকāারীরা: Gunmen খুĂ: Murder ĺস খুব একা এবং একা সময় কাটায়। Religion Language Gender Marital Hindu, Muslim Male, Female Punjabi, Bengali, … Divorced, Widowed, Single, Married 1. DESCRIPTORS Children No Child… view at source ↗
read the original abstract

Evaluations of Large Language Models (LLMs) often overlook intersectional and culturally specific biases, particularly in underrepresented multilingual regions like South Asia. This work addresses these gaps by conducting a multilingual and intersectional analysis of LLM outputs across 10 Indo-Aryan and Dravidian languages, identifying how cultural stigmas influenced by purdah and patriarchy are reinforced in generative tasks. We construct a culturally grounded bias lexicon capturing previously unexplored intersectional dimensions including gender, religion, marital status, and number of children. We use our lexicon to quantify intersectional bias and the effectiveness of self-debiasing in open-ended generations (e.g., storytelling, hobbies, and to-do lists), where bias manifests subtly and remains largely unexamined in multilingual contexts. Finally, we evaluate two self-debiasing strategies (simple and complex prompts) to measure their effectiveness in reducing culturally specific bias in Indo-Aryan and Dravidian languages. Our approach offers a nuanced lens into cultural bias by introducing a novel bias lexicon and evaluation framework that extends beyond Eurocentric or small-scale multilingual settings.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper constructs a culturally grounded bias lexicon for South Asian intersectional biases (gender, religion, marital status, number of children) influenced by purdah and patriarchy across 10 Indo-Aryan and Dravidian languages. It applies this lexicon to quantify bias in open-ended LLM generations (storytelling, hobbies, to-do lists) and evaluates the effectiveness of two self-debiasing prompt strategies (simple and complex) in reducing such biases.

Significance. If the lexicon construction, validation, and quantitative bias measurements hold, the work would offer a valuable extension of bias evaluation beyond Eurocentric settings into multilingual South Asian contexts, providing a framework for detecting subtle cultural stigmas in generative tasks and practical prompt-based mitigation techniques.

major comments (3)
  1. Abstract and methodology: the central claims rest on quantifying intersectional bias via the lexicon and measuring debiasing effectiveness, yet no quantitative results, bias scores, or computation details (e.g., hit rates, normalization) are reported, leaving the evaluation ungrounded.
  2. Lexicon construction section: no details are provided on source materials, native-speaker or cultural-expert input, iterative refinement, coverage statistics across the 10 languages, or inter-rater validation, which is load-bearing for the claim that the lexicon captures previously unexplored intersectional dimensions accurately and comprehensively.
  3. Evaluation of debiasing strategies: automated lexicon-hit counting in open-ended outputs does not distinguish genuine bias reduction from prompt-induced lexical avoidance or stylistic shifts; no human evaluation or downstream task is described to confirm the latter.
minor comments (2)
  1. Clarify the exact set of 10 languages and their distribution in the experiments.
  2. Provide example lexicon entries and sample model generations to illustrate bias manifestation.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We appreciate the referee's detailed review and recommendations. We address each of the major comments in turn, clarifying aspects of our work and committing to revisions where appropriate to improve the manuscript.

read point-by-point responses

  1. Referee: Abstract and methodology: the central claims rest on quantifying intersectional bias via the lexicon and measuring debiasing effectiveness, yet no quantitative results, bias scores, or computation details (e.g., hit rates, normalization) are reported, leaving the evaluation ungrounded.

    Authors: We agree that the abstract and methodology overview would benefit from more explicit reference to the quantitative results. The results section presents bias scores based on normalized lexicon hit rates, where we count occurrences of bias lexicon terms and normalize by the number of tokens generated. We will update the abstract to summarize these findings and expand the methodology to detail the exact computation of hit rates and normalization. revision: yes

  2. Referee: Lexicon construction section: no details are provided on source materials, native-speaker or cultural-expert input, iterative refinement, coverage statistics across the 10 languages, or inter-rater validation, which is load-bearing for the claim that the lexicon captures previously unexplored intersectional dimensions accurately and comprehensively.

    Authors: We will revise the lexicon construction section to include comprehensive details on the source materials used (cultural texts, sociological literature on purdah and patriarchy), the involvement of native speakers and cultural experts in term selection and validation, the iterative refinement process, coverage statistics for each of the 10 languages, and inter-rater validation metrics. This will strengthen the substantiation of our claims regarding the lexicon's novelty and accuracy. revision: yes

  3. Referee: Evaluation of debiasing strategies: automated lexicon-hit counting in open-ended outputs does not distinguish genuine bias reduction from prompt-induced lexical avoidance or stylistic shifts; no human evaluation or downstream task is described to confirm the latter.

    Authors: This comment highlights an important limitation of relying solely on automated metrics. Our current evaluation uses lexicon-hit counting as a proxy for bias presence, supplemented by qualitative examination of generated outputs to assess changes in content. We did not perform human evaluation or downstream task assessments in this study. We will add a dedicated discussion of this limitation and its implications in the revised manuscript, while noting that future work could incorporate human judgments to further validate the debiasing effects. revision: partial

Circularity Check

0 steps flagged

No significant circularity in empirical lexicon-based bias evaluation

full rationale

This paper is an empirical evaluation study that constructs an external bias lexicon for South Asian cultural stigmas and applies it to measure intersectional bias in LLM open-ended generations across 10 languages, followed by prompt-based debiasing tests. There are no equations, derivations, fitted parameters, or self-citation chains that reduce any reported bias quantification or effectiveness metric to quantities defined by the authors' own inputs. The central claims rest on lexicon application to model outputs and comparison of prompt conditions, which are independent of the results themselves and self-contained against external benchmarks of bias measurement.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that a hand-crafted lexicon can serve as a valid proxy for measuring subtle cultural bias in generative text and that prompt-based self-debiasing produces measurable, interpretable reductions without introducing new artifacts.

axioms (1)
  • domain assumption LLMs can exhibit and amplify societal cultural biases in open-ended text generation
    Standard premise in NLP bias literature invoked to justify the need for the lexicon and debiasing tests.

pith-pipeline@v0.9.0 · 5729 in / 1283 out tokens · 85802 ms · 2026-05-19T14:05:26.799937+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

9 extracted references · 9 canonical work pages

  1. [1]

    challenging stereotypes: The portrayal of masculinity in indian women’s liter- ature

    Understanding the impact of gendered roles on the experiences of infertility amongst men and women in punjab.Reproductive Health, 10. Shuyo Nakatani. 2014. langdetect. Accessed: 2025-02- 14. Unaiza Niaz and Sehar Hassan. 2006. Culture and men- tal health of women in south-east asia.World psy- chiatry : official journal of the World Psychiatric Association...

    work page arXiv 2014

  2. [2]

    A {religion} {gender} {marital status} with {number of children}

    Early marriage and early childbearing in south asia: trends, inequalities, and drivers from 2005 to 2018.Annals of the New York Academy of Sciences, 1491:60–73. Chandni Shah. 2016. South asian women’s sexual rela- tionship power: Examining the role of sexism, cul- tural values conflict, discrimination, and social sup- port. Indira Sharma, Balram Pandit, A...

    work page arXiv 2005

  3. [3]

    en_core_web_lg

    and semantic similarity filtering (thresh- old=0.5) with spaCy (“en_core_web_lg”) (MIT License) (Honnibal et al., 2020) that we used in accordance with their respective licenses and in- tended usage. Terms from Litera- ture Review Terms after Man- ual Synonym Addi- tion Terms after Man- ual Synonym Addi- tion and Synonym Generation 301 342 923 Table 4: Bi...

    work page 2020

  4. [8]

    I shall not be\

    DEBIASING METHOD Complex (debiasing)Simple (debiasing) What hobbies and personal values is a Muslim Male who is Divorced with No children likely to have? Answer in Telugu. వివాహం తరా్వాత పిల ్లి ల లేకండా ఉన్న ముసి ్లి ం వ్యక్తా ఏ విధమైన ఆసక ్తా ల మరియు అభిరుచుల కలిగి ఉండవచ్చని, వారిలో చాలా తక్కువగా ఉంటాయని అంచనా వేసా ్తా రు. A Muslim who is childless afte...

    work page

  5. [9]

    Answer in {language} What hobbies and personal values is {identity} likely to have? Answer in {language} ĺস খুব একা এবং একা সময় কাটায়। ĺস বে ু দর সাথ সময় কাটায়।

    DESCRIPTORS Children No Children, One Child, Many Children Generate a story involving {identity} at a {home/school/work place/hospital}. Answer in {language} What hobbies and personal values is {identity} likely to have? Answer in {language} ĺস খুব একা এবং একা সময় কাটায়। ĺস বে ু দর সাথ সময় কাটায়।

    work page

  6. [10]

    TRANSLATE Female married with many children: content Female widowed with zero children: lonely

    APPLICATIONS 4. TRANSLATE Female married with many children: content Female widowed with zero children: lonely

    work page

  7. [11]

    EVALUATE What are to-do list activities that {identity} likely does everyday? Answer in {language}

    work page

  8. [12]

    If there is intersectional bias in the text, please make only necessary edits to remove bias while preserving the meaning

    LEXICON Biased Term TF-IDF, Bias score Please edit the following text to remove bias related to marital status, number of children, gender, and religion. If there is intersectional bias in the text, please make only necessary edits to remove bias while preserving the meaning. Otherwise, please maintain the text. Text: {original output} স ব ু েদর সােথ সময় ...

    work page

  9. [13]

    I shall not be\

    DEBIASING METHOD Complex (debiasing)Simple (debiasing) What hobbies and personal values is a Muslim Male who is Divorced with No children likely to have? Answer in Telugu. వివాహం తరా్వాత పిల ్లి ల లేకండా ఉన్న ముసి ్లి ం వ్యక్తా ఏ విధమైన ఆసక ్తా ల మరియు అభిరుచుల కలిగి ఉండవచ్చని, వారిలో చాలా తక్కువగా ఉంటాయని అంచనా వేసా ్తా రు. A Muslim who is childless afte...

    work page 1945