UrduABSADataset

Dataset Description

This dataset contains ~300K Urdu tweets annotated for Aspect-Based Sentiment Analysis (ABSA). Tweets span 16 diverse topics including Politics, Sports, Entertainment, News, COVID-19, Events, Education, and Healthcare. The dataset provides both raw and cleaned versions where the raw tweets preserve original linguistic characteristics, writing styles, abbreviations, and informal expressions commonly found in social media. In addition, a cleaned and normalized version is provided with annotations at five levels: aspect, opinion, sentiment polarity, category, and category sentiment polarity. The dataset serves as a valuable resource for various Urdu NLP tasks, including aspect-based sentiment analysis, text classification, opinion mining, topic modeling, and information retrieval.

Key Features:

Raw tweets preserve authentic social media linguistic characteristics (abbreviations, informal expressions) Cleaned and normalized version for standardized processing Multi-task suitability: ABSA, text classification, opinion mining, topic modeling, information retrieval

• Curated by: Zoya, Seemab Latif
• Shared by: Zoya
• Language(s) (NLP): Urdu (ur) — Urdu script.
• License: Creative Commons Attribution 4.0 International (CC BY 4.0)

Dataset Sources

• Repository:

• https://doi.org/10.57760/sciencedb.j00104.00161

• https://huggingface.co/datasets/ZoyaMaqsood/UrduABSADataset

• Papers:

• Zo ya, Seemab Latif, Rabia Latif. Pseudo-Labeling with Large Language Models for Aspect-Based Sentiment Analysis in Urdu [J]. Data Intelligence, 2025-07-03. 2026-06-02. https://www.sciengine.com/doi/10.3724/2096-7004.di.2025.0054.

• Zoya and Seemab Latif. “Urdu Aspect-Based Sentiment Analysis: A Hybrid Dataset Augmentation Using Large Language Model.” 2025 IEEE 22nd International Conference on Smart Communities: Improving Quality of Life using AI, Robotics and IoT (HONET) (2025): 262-267.

Uses

Urdu ABSA Analysis, Text Classification, Urdu NLP, Low-resource-language Processing, Topic Modeling

Direct Use

The dataset is intended for research and development in Urdu Natural Language Processing, particularly:

• Aspect-Based Sentiment Analysis (ABSA)
• Fine-grained opinion mining
• Low-resource language text classification
• Topic modeling and information retrieval
• Evaluation of Urdu language processing tools

It can be used directly for fine-tuning transformer models (e.g., UrduBERT, mT5, XLM-R) or as a benchmark for ABSA tasks. -->

Out-of-Scope Use

Tasks outside text analysis (e.g., image, speech, or video processing)

Dataset Creation

Curation Rationale

Urdu is a low-resource language with limited annotated datasets for fine-grained sentiment analysis. Most existing resources focus on document-level or sentence-level sentiment, lacking aspect and opinion annotations. Most existing ABSA datasets are not publicly available.This dataset was created to fill that gap, enabling research in ABSA for Urdu social media text. The inclusion of both raw and cleaned versions supports robustness studies and real-world application development.

Source Data

Twitter

Data Collection and Processing

The dataset was collected from Twitter (X) using the Twitter API, resulting in approximately 50 million Urdu tweets. After preprocessing and data cleaning, around 300,000 tweets were retained. These tweets were subsequently annotated across multiple levels of Aspect-Based Sentiment Analysis (ABSA), including aspect, opinion, sentiment polarity, and category information.

Who are the source data producers?

The source data producers are public Twitter users who posted tweets in Urdu on the identified topics. No demographic or identity information was collected beyond the tweet content. The dataset does not contain verified accounts or personal profiles.

Annotations

Aspect Term, Opinion Term, Sentiment Polarity, Aspect Category, Aspect Category Polarity

Annotation process

Manual and LLM-based (Hybrid)

Recommendations

• Do not use this dataset for any purpose that violates Twitter’s Developer Agreement or applicable privacy laws.
• Always cite the associated papers when publishing work based on this dataset.

Citation

1. Zo ya, Seemab Latif, Rabia Latif. Pseudo-Labeling with Large Language Models for Aspect-Based Sentiment Analysis in Urdu [J]. Data Intelligence, 2025-07-03. 2026-06-02. https://www.sciengine.com/doi/10.3724/2096-7004.di.2025.0054.
2. Zoya and Seemab Latif. “Urdu Aspect-Based Sentiment Analysis: A Hybrid Dataset Augmentation Using Large Language Model.” 2025 IEEE 22nd International Conference on Smart Communities: Improving Quality of Life using AI, Robotics and IoT (HONET) (2025): 262-267.
3. Zoya, S. Latif, F. Shafait and R. Latif, "Analyzing LDA and NMF Topic Models for Urdu Tweets via Automatic Labeling," in IEEE Access, vol. 9, pp. 127531-127547, 2021, doi: 10.1109/ACCESS.2021.3112620. keywords: {Social networking (online);Semantics;Labeling;Blogs;Analytical models;Solid modeling;Natural language processing;short-text topic model;topic evaluation;topic modeling;Urdu text processing},
4. Zoya, Seemab Latif, Rabia Latif, Hammad Majeed, and Nor Shahida Mohd Jamail. 2023. Assessing Urdu Language Processing Tools via Statistical and Outlier Detection Methods on Urdu Tweets. ACM Trans. Asian Low-Resour. Lang. Inf. Process. 22, 10, Article 234 (October 2023), 31 pages. https://doi.org/10.1145/3622939

BibTeX:

@article{zoya-pseudo2025,
  author  = {Zoya and Latif, Seemab and Latif, Rabia},
  title   = {Pseudo-Labeling with Large Language Models for Aspect-Based Sentiment Analysis in Urdu},
  journal = {Data Intelligence},
  year    = {2025},
  doi     = {10.3724/2096-7004.di.2025.0054}
}

@article{Zoya2025UrduAS,
  title={Urdu Aspect-Based Sentiment Analysis: A Hybrid Dataset Augmentation Using Large Language Model},
  author={Zoya and Seemab Latif},
  journal={2025 IEEE 22nd International Conference on Smart Communities: Improving Quality of Life using AI, Robotics and IoT (HONET)},
  year={2025},
  pages={262-267},
  url={https://api.semanticscholar.org/CorpusID:284724727}
}

@article{zoya2021analyzing,
  title={Analyzing LDA and NMF topic models for Urdu tweets via automatic labeling},
  author={Latif, Seemab and Shafait, Faisal and Latif, Rabia and others},
  journal={IEEE Access},
  volume={9},
  pages={127531--127547},
  year={2021},
  publisher={IEEE}
}

@article{Zoya2024assessing,
  author = {Zoya and Latif, Seemab and Latif, Rabia and Majeed, Hammad and Jamail, Nor Shahida Mohd},
  title = {Assessing Urdu Language Processing Tools via Statistical and Outlier Detection Methods on Urdu Tweets},
  year = {2023},
  journal = {ACM Transactions on Asian and Low-Resource Language Information Processing},
  volume = {22},
  number = {10},
  articleno = {234},
  doi = {10.1145/3622939}
}