Mahnoornaz/Sindhi-sentiment-analysis-dataset-100k

Sindhi Sentiment Analysis Dataset (100K)

Dataset Description

The Sindhi Sentiment Analysis Dataset (100K) is a large-scale, labeled text classification dataset built to support sentiment analysis research and applications in the Sindhi language, a low-resource language spoken by over 25 million people primarily in the Sindh province of Pakistan and parts of India.

The dataset contains 100,000 samples, each consisting of a Sindhi text snippet and a corresponding sentiment label (positive, negative, or neutral). It was built using a hybrid data creation approach, combining translated real-world social media sentiment data with synthetically generated sentences, in order to achieve both authenticity and scale while keeping the label distribution balanced.

This dataset is intended to help close the resource gap for Sindhi natural language processing (NLP) and to enable the training and evaluation of sentiment classification models for the language.

Motivation

Sindhi is classified as a low-resource language in NLP research — very few labeled datasets exist for tasks such as sentiment analysis, making it difficult to train or benchmark machine learning models for Sindhi text understanding. Most sentiment resources are concentrated in high-resource languages such as English, leaving regional and minority languages underserved.

This dataset was created to:

• Provide a sizeable, balanced, and openly available sentiment dataset for Sindhi.
• Encourage NLP research, tool-building, and benchmarking for Sindhi and other low-resource South Asian languages.
• Serve as a foundation for downstream applications such as social media monitoring, customer feedback analysis, and content moderation in Sindhi.

Language

• Language: Sindhi (سنڌي)
• ISO 639-1 code: sd
• Script: Perso-Arabic (Sindhi script)
• Region: Primarily Pakistan (Sindh province), with speakers also in India and the Sindhi diaspora worldwide

Dataset Structure

Data Instances

Each row in the dataset represents a single text sample with its associated sentiment label. Example:

Text	Label
هي فلم تمام بورنگ هئي.	negative
موسيقي فلم کي وڌيڪ سهڻو بڻايو.	positive
معمول موجب بئنڪ جي وقت شام پنجين وڳي تائين آهي.	neutral

Data Fields

• Text (string): A sentence or short passage written in Sindhi script.
• Label (string): The sentiment associated with the text. One of positive, negative, or neutral.

Data Splits

The dataset is released as a single unified file. It does not currently ship with predefined train/validation/test splits; users are encouraged to create their own splits (e.g., 80/10/10) using stratified sampling on the Label column to preserve class balance.

• Total samples: 100,000

Labels

The dataset uses three sentiment classes:

Label	Description
positive	Text expressing favorable opinions, happiness, satisfaction, praise, or approval.
negative	Text expressing unfavorable opinions, dissatisfaction, criticism, sadness, or complaints.
neutral	Factual, descriptive, or informational text with no clear positive or negative sentiment.

The label distribution is approximately balanced across all three classes, with each class representing roughly one-third of the dataset.

Data Collection

This dataset was created using a hybrid approach, combining two sources:

1. Real-world, translation-derived data (~25,000 samples): Sourced from the publicly available Twitter Entity Sentiment Analysis dataset on Kaggle (twitter_training.csv), which contains English-language tweets labeled with sentiment.
2. Synthetically generated data (~75,000 samples): Additional Sindhi sentences generated programmatically using template- and vocabulary-based text generation techniques to expand the dataset's size and topical diversity while preserving label balance.

Data Cleaning and Preprocessing

The Twitter-derived portion of the dataset underwent the following preprocessing steps before translation:

• Removal of duplicate and near-duplicate tweets.
• Removal of offensive, abusive, hateful, or otherwise inappropriate content.
• Removal of irrelevant noise such as URLs, user handles/mentions, hashtags used as metadata, and excessive punctuation or emojis that did not contribute to sentiment meaning.
• Filtering out of very short, ambiguous, or non-sentiment-bearing tweets.
• Normalization of text spacing and encoding (UTF-8) prior to translation.

After translation, all Sindhi text (from both the translated and synthetic portions) was checked to ensure:

• No blank or null values in the Text or Label columns.
• Labels were restricted strictly to positive, negative, or neutral.
• Consistent UTF-8 encoding suitable for Sindhi (Perso-Arabic) script rendering.

Translation Process

The ~25,000 samples derived from the Twitter Entity Sentiment Analysis dataset were originally in English. These samples went through the following translation pipeline:

1. Filtering: Tweets were filtered for sentiment clarity and to remove offensive or low-quality content (see Data Cleaning and Preprocessing above).
2. Translation: The remaining English text was translated into natural, fluent Sindhi, aiming to preserve the original meaning, tone, and — critically — the original sentiment label.
3. Label preservation: The original sentiment label (positive, negative, or neutral) from the source dataset was retained for each translated sample, under the assumption that sentiment polarity is preserved across translation.
4. Quality review: Translated samples were reviewed for fluency and naturalness in Sindhi, so that the resulting text reads as native Sindhi rather than a literal machine translation.

Note: As with any translation-based dataset, some nuance, sarcasm, idiomatic expression, or culturally specific sentiment cues from the original English tweets may not translate perfectly. Users should be aware of this limitation when using the translated subset.

Synthetic Data Generation

To reach the target size of 100,000 samples while keeping the dataset balanced and diverse, approximately 75,000 additional samples were synthetically generated. The synthetic generation process involved:

• Template-based sentence construction: A large set of Sindhi sentence templates was designed to reflect natural, everyday statements across a range of topics (e.g., shopping, banking, travel, food, education, health, family, weather, and social interactions).
• Vocabulary substitution: Templates were populated using curated pools of Sindhi names, city/place names, objects, food items, and common phrases, producing a wide combinatorial variety of unique sentences.
• Sentiment-conditioned generation: Separate template and vocabulary sets were used for each sentiment class (positive, negative, neutral) so that generated text authentically reflects the intended sentiment.
• Gender and grammatical agreement: Sentence templates account for grammatical gender agreement (e.g., verb forms) based on the subject name used, to preserve grammatical correctness.
• Deduplication: Generated samples were checked against both the existing dataset and previously generated samples to avoid duplicate or near-duplicate entries being introduced during this generation phase.
• Balance control: Generation was balanced across the three sentiment labels to maintain an approximately equal class distribution across the full 100,000-sample dataset.

Because this portion of the data is synthetically generated rather than collected from real user-generated content, it may exhibit more repetitive sentence structures and less lexical/topical diversity than naturally occurring text, even though duplicate exact-text entries were actively filtered out during generation.

Intended Use

This dataset is intended for:

• Training and fine-tuning text classification / sentiment analysis models for Sindhi.
• Benchmarking NLP models on a low-resource language task.
• Academic research on Sindhi computational linguistics, sentiment analysis, and low-resource NLP methods.
• Building downstream applications such as social media sentiment monitoring, customer feedback classification, or content moderation tools for Sindhi text.

It is not intended to be used as a sole source of ground truth for high-stakes decisions (e.g., moderation actions, legal, medical, or financial decisions) without human review, given its partially synthetic and translated nature.

Limitations

Users of this dataset should be aware of the following limitations:

• Synthetic majority: The majority (~75%) of the dataset is synthetically generated using templates, which may lead to repetitive sentence patterns, limited stylistic variety, and lower linguistic naturalness compared to fully organic, human-written text.
• Translation artifacts: The translated portion (~25%) may contain translation artifacts, and some sentiment nuance, sarcasm, humor, or cultural context from the original English tweets may be lost or altered during translation.
• Domain skew: The translated subset originates from Twitter/X, so it may reflect the topical and stylistic biases of that platform (e.g., commentary on public figures, brands, current events) rather than a fully representative sample of everyday Sindhi text.
• Dialectal coverage: Sindhi has regional dialectal variation; this dataset may not equally represent all dialects or regional forms of the language.
• Label subjectivity: Sentiment labeling, especially for translated text, involves inherent subjectivity, and label preservation across translation is an assumption rather than a guarantee of perfect accuracy.
• No named-entity or demographic annotations: The dataset does not include additional metadata (e.g., topic, source, timestamp, or annotator information) beyond text and label.

Ethical Considerations

• Content filtering: Efforts were made to remove offensive, hateful, abusive, or otherwise harmful content from the source Twitter data prior to translation and inclusion. However, no automated or manual filtering process can guarantee the complete absence of such content, and users should perform their own review if deploying this dataset in sensitive contexts.
• Bias: As with any dataset derived from social media, the original Twitter-based subset may carry biases present in the source platform's user base, topics of discussion, and prevailing sentiment expressions. The synthetic subset, while designed for balance, reflects the choices made in template and vocabulary design and may not capture the full diversity of authentic Sindhi expression.
• Privacy: The Twitter-derived subset originates from a public, pre-existing Kaggle dataset intended for research use; no additional personal or identifying information was collected, and translated text was generalized/cleaned rather than tied to specific individuals.
• Representation: As a low-resource language dataset, this resource aims to improve representation of Sindhi in NLP research. Users are encouraged to use it responsibly and to pair it with human review when building applications that affect real users.

License

This dataset is released under the MIT License. Users are free to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the dataset, for any purpose, including commercial use, provided the original copyright and license notice is included.

MIT License

Copyright (c) 2026 [Your Name / Organization]

Permission is hereby granted, free of charge, to any person obtaining a copy
of this dataset and associated documentation files (the "Dataset"), to deal
in the Dataset without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Dataset, and to permit persons to whom the Dataset is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Dataset.

THE DATASET IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE DATASET OR THE USE OR OTHER DEALINGS IN THE
DATASET.

Note: The Twitter-derived subset is based on the Twitter Entity Sentiment Analysis dataset from Kaggle; users should also review the original dataset's license/usage terms on Kaggle before redistribution.

Citation

If you use this dataset in your research or applications, please cite it as:

@dataset{sindhi_sentiment_2026,
  title        = {Sindhi Sentiment Analysis Dataset (100K)},
  author       = {[Mahnoor Naz Baloch/ Proxima AI ]},
  year         = {2026},
  note         = {A hybrid dataset combining translated Twitter sentiment data and synthetically generated Sindhi text for sentiment classification.},
  howpublished = {Hugging Face Datasets}
}

Please also cite the original source dataset used for the translated portion:

Twitter Entity Sentiment Analysis Dataset. Kaggle. https://www.kaggle.com/datasets/jp797498e/twitter-entity-sentiment-analysis

References

• Twitter Entity Sentiment Analysis Dataset. Kaggle. https://www.kaggle.com/datasets/jp797498e/twitter-entity-sentiment-analysis