Sabihaa12/sindhi-sentiment-dataset

Sindhi Sentiment Analysis Dataset

Dataset Description

This dataset is a Sindhi-language sentiment analysis dataset consisting of short text samples labeled as positive, negative, or neutral. It was built to support NLP research and model development for Sindhi, a low-resource language spoken by over 25 million people primarily in Sindh, Pakistan.

• Language: Sindhi (sd)
• Task: Text Classification (Sentiment Analysis)
• Labels: positive, negative, neutral
• Total examples: 2,833
• License: MIT

Dataset Creation

Source Data

The dataset was created by translating a balanced sample of an existing English-language social media sentiment dataset (originally labeled 0 = negative, 1 = neutral, 2 = positive) into Sindhi. A stratified sample of ~1,000 examples per label was selected from the source corpus before translation to ensure class balance.

Translation Process

Text was translated from English to Sindhi using the deep-translator Python library (Google Translate backend). The translation pipeline:

1. Loaded the original dataset and mapped numeric labels to positive / negative / neutral.
2. Sampled a balanced subset (~1,000 rows per label), filtering out very short (<20 chars) or very long (>250 chars) comments.
3. Translated each sampled comment to Sindhi (sd) via the Google Translate API, with retry logic for failed requests and incremental checkpoint saving to avoid data loss.

Cleaning Process

After translation, the dataset was cleaned to remove translation artifacts:

1. English token removal — any word retaining Latin script (untranslated words, usernames, hashtags, leftover English phrases) was stripped, while the surrounding Sindhi text was preserved.
2. URL / mention / hashtag removal — social media artifacts (http..., @user, #tag) were removed.
3. Character filtering — text was restricted to Sindhi/Arabic script, digits, and basic punctuation; stray symbols from mixed-language contamination were removed.
4. Short-row removal — rows with fewer than 3 words remaining after cleaning were dropped (these were mostly rows that were originally almost entirely in English).
5. Deduplication — exact duplicate rows were removed.

Original translated set: 2,901 rows → Final cleaned set: 2,833 rows.

Label Distribution

Label	Count
Positive	955
Negative	952
Neutral	926
Total	2,833

Dataset Structure

Data Fields

• Text (string): A Sindhi-language sentence or short passage.
• Label (string): Sentiment label — one of positive, negative, neutral.

Data Example

Text: مان ڏاڍو حسد آهيان صحيح
Label: negative

Considerations for Using the Data

Limitations

• Machine-translated text: Since the dataset was created via automated translation rather than native Sindhi authorship, some translations may contain unnatural phrasing, grammatical inconsistencies, or minor semantic drift from the original English sentiment.
• Source domain bias: The underlying source data was social-media text (short, informal, opinionated), so the dataset may not generalize well to formal or long-form Sindhi text (e.g., news articles).
• Label noise: Sentiment labels were inherited directly from the source dataset and were not independently re-verified by native Sindhi speakers after translation; some translated examples may have subtly shifted sentiment.
• Residual artifacts: Despite cleaning, a small number of rows may retain minor translation noise (e.g., broken grammar from source text ambiguity).

Recommended Uses

• Fine-tuning/evaluating text classification models for Sindhi sentiment analysis.
• Baseline benchmark for low-resource Sindhi NLP tasks.
• Research on cross-lingual transfer learning (English → Sindhi).

Citation

If you use this dataset, please cite it as:

@misc{sindhi_sentiment_2026,
  title  = {Sindhi Sentiment Analysis Dataset},
  author = {Sabiha},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Sabihaa12/sindhi-sentiment-dataset}
}

Acknowledgements

Translation pipeline built using deep-translator. Inspired by related Sindhi NLP dataset efforts such as alinawazmahar/Sindhi_Sentiment_dataset and DanishMahdi/snd_movies_sentiment_analysis.