necrosyth/indic-codeswitched-social-bias

Dataset Description

This dataset contains Indian code-switched sentences — informal text where speakers fluidly mix English with Indian regional languages — annotated for social bias across multiple dimensions.

Languages covered include (but may not be limited to):

• Hindi-English (Hinglish) — hi-en
• Tamil-English (Tanglish) — ta-en
• Punjabi-English — pa-en
• Bengali-English — bn-en
• Gujarati-English — gu-en
• Telugu-English — te-en
• Kannada-English — kn-en
• Marathi-English — mr-en

Code-switching is the dominant informal communication mode for 1B+ Indians across WhatsApp, Instagram, Twitter/X, YouTube comments, and regional forums. Despite this scale, virtually no bias detection dataset exists that captures this multilingual reality. Most existing work assumes monolingual English or formal Hindi — failing entirely on how Indians actually write online.

Supported Tasks

• Bias Classification — Multi-class classification of social bias type
• Bias Detection — Binary detection (biased vs. neutral)
• Text Debiasing / Style Transfer — Given a biased sentence, generate its neutral counterpart
• Explainability Research — The explanation column supports rationale-based or chain-of-thought bias detection models
• Confidence Calibration — Confidence scores enable threshold tuning and uncertainty modeling

---

Dataset Structure

Columns

Column	Type	Description
sentence	string	Original Hinglish sentence (may contain bias)
languages	string	Language mix tag (e.g., hi-en)
bias_type	string	Category of bias: gender, religion, caste, color, regional, or none
confidence_score	float	Model/annotator confidence in the bias label (0–1)
explanation	string	Natural language rationale explaining the bias
unbiased_sentence	string	A neutralized rewrite of the sentence

Bias Categories

• Gender — Stereotypes or unequal treatment based on gender roles
• Religion — Prejudice or assumptions based on religious identity
• Caste — Discrimination rooted in caste-based social hierarchy
• Color — Colorism and appearance-based bias (skin tone)
• Regional — Stereotypes based on state, region, or linguistic community
• None — Neutral sentences included for balance

---

Use Cases

1. Training Bias Classifiers for Indian Social Media

Platforms like Instagram, Twitter/X, and ShareChat need moderation tools that understand Hinglish. This dataset directly enables training classifiers for that context.

2. Debiasing / Counterfactual Generation

The unbiased_sentence column makes this ideal for seq2seq debiasing models (T5, mBART, IndicBART) that learn to neutralize biased text.

3. Evaluating LLMs on Hindi-English Code-Switch Bias

Use this as a benchmark to probe whether multilingual LLMs (GPT-5, Gemini, LLaMA) correctly identify and avoid social bias in Hinglish contexts.

4. Fairness Research in South Asian NLP

Caste and colorism are dimensions largely absent from Western bias datasets. This fills a critical gap for researchers working on inclusive AI.

5. Explainability & Rationale Datasets

The explanation column can train or evaluate models that must justify their bias predictions — useful for RLHF, rationale extraction, and chain-of-thought benchmarking.

6. Cross-Lingual Bias Transfer Research

Do bias patterns in Hinglish transfer to Tanglish? Does regional bias manifest differently in Punjabi-English vs Bengali-English? This dataset enables systematic cross-lingual analysis across Indian language families (Indo-Aryan vs Dravidian).

7. Indic NLP Benchmarking

Complements existing Indic NLP benchmarks (IndicGLUE, Sangraha, AI4Bharat corpora) with a bias-specific evaluation layer across code-switched registers that those benchmarks miss entirely.

8. Regional Stereotype Analysis

The regional bias category, combined with the languages tag, lets researchers study whether regional stereotypes are more prevalent in text from/about specific communities — a unique sociolinguistic signal.

Languages

• Hinglish (hi-en): Hindi-English code-switched informal text

---

Dataset Creation

Source & Annotation

[Describe your annotation process here — was it human-annotated, LLM-assisted, or a combination? What guidelines did annotators follow?]

Curation Rationale

Existing bias datasets (WinoBias, StereoSet, CrowS-Pairs) are English-only and culturally Western. This dataset addresses the need for bias resources rooted in South Asian social and linguistic contexts.

---

Limitations

• Sentences are informal/colloquial; may not generalize to formal Hindi or standard English

• Confidence scores reflect annotator/model agreement and should not be treated as ground truth probability

• Caste and regional bias categories reflect Indian social hierarchies that may require domain knowledge to interpret correctly

• Dataset may reflect annotator positionality in edge cases

• Language coverage is uneven — some regional languages may have significantly fewer samples than Hindi-English

• Annotators fluent in all covered languages simultaneously are rare; annotation quality may vary by language pair

• Dravidian languages (Tamil, Kannada, Telugu) have structurally different code-switching patterns than Indo-Aryan ones — models should not be assumed to generalize across language families without evaluation

---

Citation

If you use this dataset, please cite: [Add your citation here once published]

---

License

Creative Commons Attribution 4.0 (CC BY 4.0)

tags:

• code-switching
• multilingual
• indic-languages
• hinglish
• tanglish
• social-bias
• bias-detection
• debiasing
• fairness
• hindi
• tamil
• punjabi
• bengali
• gujarati
• low-resource
• ai4bharat