Datasets:

tunis-ai
/

tsac

dataset_info:
  features:
    - name: sentence
      dtype: string
    - name: label
      dtype: int64
    - name: __index_level_0__
      dtype: int64
  splits:
    - name: train
      num_bytes: 1017693
      num_examples: 13669
    - name: test
      num_bytes: 269633
      num_examples: 3400
  download_size: 713636
  dataset_size: 1287326
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
license: lgpl-3.0
task_categories:
  - text-classification
language:
  - ar
tags:
  - arabic
  - tunisian
  - sentiment_analysis
pretty_name: Tunisian Sentiment Analysis Corpus (TSAC)
size_categories:
  - 10K<n<100K

Tunisian Sentiment Analysis Corpus (TSAC)

The Tunisian Sentiment Analysis Corpus (TSAC) is a collection of approximately 17,000 Tunisian Arabic user comments manually annotated for sentiment polarity (positive or negative).
It was collected from Facebook comments written on the official pages of Tunisian radio and TV stations between January 2015 and June 2016.

This cleaned and Hugging Face–ready version of TSAC provides train/test splits in a simple format compatible with any modern NLP framework.

Dataset Details

Dataset Description

Name: Tunisian Sentiment Analysis Corpus (TSAC)
Curated by: Salima Medhaffar, Fethi Bougares, Yannick Estève, and Lamia Hadrich-Belguith
Language: Tunisian Arabic (Arabic script)
License: Apache License 2.0
Original Repository: https://github.com/fbougares/TSAC
Hugging Face Maintainer: tunis-ai

Dataset Sources

Data collected from: Official Facebook pages of
- Mosaique FM
- Jawhara FM
- Shems FM
- Hiwar Ettounsi TV
- Nessma TV
Timeframe: January 2015 – June 2016
Paper: Medhaffar et al., 2017 — Sentiment Analysis of Tunisian Dialects: Linguistic Resources and Experiments

Uses

Direct Use

The dataset is suitable for:

Sentiment analysis in Tunisian Arabic
Dialectal Arabic language modeling
Evaluation of cross-dialectal or multilingual sentiment models

Out-of-Scope Use

Not suitable for general Modern Standard Arabic (MSA) tasks.
Not recommended for topic classification or sarcasm detection without adaptation.

Dataset Structure

Data Fields

Field	Type	Description
`sentence`	string	User comment written in Tunisian Arabic
`label`	int	Sentiment label (1 = positive, 0 = negative)

Splits

Split	# Examples
Train	13,669
Test	3,400

Splits were created using a stratified partition to maintain class balance.

Dataset Creation

Curation Rationale

The dataset was built to support research in sentiment analysis for Tunisian Arabic, a dialect that differs significantly from Modern Standard Arabic and lacks large-scale annotated resources.

Data Collection and Processing

Comments were collected from Facebook public pages using web scraping tools, manually filtered for relevance, and annotated by native Tunisian speakers into two polarity classes: positive and negative.

Preprocessing steps include:

Removing URLs, emojis, and metadata
Normalizing Arabic characters
Deduplicating sentences

Source Data Producers

Public Facebook users posting on the mentioned Tunisian media pages between 2015 and 2016.

Annotations

Annotation Type: Binary sentiment classification (positive/negative)
Annotators: Native Tunisian Arabic speakers
Validation: Manual cross-checking and agreement verification by linguists

Personal and Sensitive Information

Comments are publicly available and were anonymized by removing any identifiable information (e.g., usernames, mentions).

Bias, Risks, and Limitations

The dataset reflects opinions expressed on public Facebook pages and may include demographic, temporal, or topical biases.
It should not be used to infer general population sentiment or to train systems that make sensitive decisions about individuals.

Citation

If you use this dataset, please cite the following paper:

BibTeX:

@inproceedings{medhaffar-etal-2017-sentiment,
    title = "Sentiment Analysis of {T}unisian Dialects: Linguistic Resources and Experiments",
    author = "Medhaffar, Salima  and Bougares, Fethi  and Estève, Yannick  and Hadrich-Belguith, Lamia",
    booktitle = "Proceedings of the Third Arabic Natural Language Processing Workshop",
    month = apr,
    year = "2017",
    address = "Valencia, Spain",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/W17-1307/",
    pages = "55--61",
    doi = "10.18653/v1/W17-1307"
}