Datasets:
Tasks:
Text Classification
Modalities:
Text
Formats:
csv
Sub-tasks:
sentiment-classification
Size:
10K - 100K
License:
| language: | |
| - mai | |
| - en | |
| - bh | |
| license: apache-2.0 | |
| tags: | |
| - sentiment-analysis | |
| - text-classification | |
| - maithili | |
| - devanagari | |
| - low-resource-nlp | |
| - indic-nlp | |
| - cross-lingual-transfer | |
| - bihari-languages | |
| task_categories: | |
| - text-classification | |
| task_ids: | |
| - sentiment-classification | |
| size_categories: | |
| - 10K<n<100K | |
| # Maithili 64K Dataset | |
| # 🌾 Maithili Multi-Dimensional Sentiment Corpus | |
|  | |
|  | |
|  | |
|  | |
| ## 📌 Executive Summary | |
| Standard sentiment analysis in Indian vernaculars relies on flat, one-dimensional classification. The **Maithili Multi-Dimensional Sentiment Corpus** (64,215 rows) introduces a high-resolution, socio-linguistically grounded architecture. It utilizes a dual-axis classification system—predicting both primary sentiment and emotional intensity—mapped across highly specific demographic speaker profiles and rural domains. | |
| ## 🧠 Architectural Innovations | |
| ### 1. Dual-Axis Taxonomy (Sentiment + Intensity) | |
| Models often fail to distinguish between minor inconveniences and severe grievances in low-resource settings. This dataset introduces the `sentiment_intensity` vector: | |
| * **Mild** (22,782 samples) | |
| * **Moderate** (23,224 samples) | |
| * **Strong** (18,209 samples) | |
| This allows researchers to train models with a nuanced loss function that scales based on the severity of the utterance. | |
| ### 2. Socio-Linguistic Anchoring (`speaker_type`) | |
| Maithili morphology shifts drastically based on profession and generation. Utterances are explicitly categorized by `speaker_type` (e.g., *Farmer, Teacher, Elder, Student*). This metadata enables the measurement and mitigation of demographic and generational bias in LLM generations. | |
| ### 3. Template-Anchored Rural Elicitation | |
| To combat the urban/social-media bias prevalent in web-scraped data, this corpus utilizes a template-anchored elicitation protocol. Translations are anchored around specific geographic entities (e.g., *Khagaria, Muzaffarpur, Jamui*) and underrepresented rural domains (Agriculture, Livestock, Health), ensuring the LLM learns to map sentiment to tangible local realities. | |
| ## 📊 Dataset Schema | |
| * `id`: Unique identifier. | |
| * `text`: The Maithili utterance. | |
| * `english_translation`: English semantic equivalent. | |
| * `label`: Primary sentiment (positive, negative, neutral). | |
| * `sentiment_intensity`: Intensity of the sentiment (mild, moderate, strong). | |
| * `domain`: e.g., agriculture, health, economy. | |
| * `speaker_type`: Demographic origin of the syntax (e.g., farmer, housewife). | |
| ## ⚙️ Intended Use & Limitations | |
| * **Best For:** Intensity regression tasks, socio-linguistic bias evaluation, and sentiment classification in rural/agricultural domains. | |
| * **Limitations:** Due to the template-anchored elicitation used to guarantee geographical coverage, the syntactic variance is lower than highly spontaneous conversational data. It is optimal for representation learning rather than open-ended generative chat. | |
| ## 📝 Citation | |
| If you use this dataset in your research, please cite the accompanying paper: | |
| ```bibtex | |
| @article{prasad2026maithili, | |
| title={abhiprd20/Maithili_Sentiment_8K}, | |
| author={Prasad, Abhimanyu}, | |
| year={2026}, | |
| } |