---
title: "Sentiment Analysis of Tunisian Dialects: Linguistic Ressources"
description: "They then build TSAC, a corpus of about 17k Facebook comments taken from the official pages of… 18 references, cited by 28 papers here."
authors: ["Salima Medhaffar", "Fethi Bougares", "Yannick Estève", "Lamia Hadrich-Belguith"]
year: 2017
venue: "Proceedings of the Third Arabic Natural Language Processing Workshop"
source: "https://aclanthology.org/W17-1307/"
doi: "10.18653/v1/w17-1307"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/medhaffar-2017-sentiment-analysis-tunisian-dialects-linguistic/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/medhaffar-2017-sentiment-analysis-tunisian-dialects-linguistic/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2017](https://md.research.elamur.ai/free/en/papers/2017/page/3/) / Sentiment Analysis of Tunisian Dialects: Linguistic Ressources and Experiments

# Sentiment Analysis of Tunisian Dialects: Linguistic Ressources and Experiments

**Authors:** Salima Medhaffar, [Fethi Bougares](https://md.research.elamur.ai/free/en/authors/fethi-bougares/), [Yannick Estève](https://md.research.elamur.ai/free/en/authors/yannick-esteve/) and [Lamia Hadrich-Belguith](https://md.research.elamur.ai/free/en/authors/lamia-hadrich-belguith/) · **Year:** 2017 · **Venue:** [Proceedings of the Third Arabic Natural Language Processing Workshop](https://md.research.elamur.ai/free/en/venues/wanlp/) ([WANLP 2017](https://md.research.elamur.ai/free/en/venues/wanlp/2017/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/W17-1307/) · [PDF](https://aclanthology.org/W17-1307.pdf) · [DOI 10.18653/v1/w17-1307](https://doi.org/10.18653/v1/w17-1307) · [OpenAlex](https://openalex.org/W2740432722)

**Citations:** 151 citations (OpenAlex, 2026-10-14) · [28](https://md.research.elamur.ai/free/en/papers/medhaffar-2017-sentiment-analysis-tunisian-dialects-linguistic/cited-by/) in our library, of the 149,439 papers whose bibliographies we read

## Summary
The paper tackles polarity classification for the Tunisian dialect ("Tounsi"), an under-resourced Maghrebi variety. Online it is written in both Arabic script and Latin-based Arabizi, has no standard spelling, and is often mixed with French or English. The authors first survey freely available Arabic sentiment corpora and find that most cover MSA, Egyptian or Levantine varieties. They then build TSAC, a corpus of about 17k Facebook comments taken from the official pages of Tunisian radio and TV stations and manually labelled as positive or negative. Since no Tunisian POS taggers, lemmatizers or parsers exist, each comment is represented by an unsupervised Doc2vec vector. SVM, Bernoulli Naive Bayes and MLP classifiers are trained on MSA data (OCA), mixed-dialect data (LABR), TSAC, and combinations of these. Models trained with TSAC reach an error rate of about 0.22–0.23. Models trained only on MSA or mixed-dialect data perform close to chance. Adding those corpora to TSAC gives no gain. The work shows that in-dialect labelled data matters and releases the first free resource for this dialect.

## Key points
- TSAC (Tunisian Sentiment Analysis Corpus) holds 8,215 positive and 8,845 negative comments. The comments were gathered from January 2015 through June 2016 on the Facebook pages of broadcasters including Nessma TV, Shemes FM, Mosaique FM, HiwarElttounsi TV and JawhraFM, and the corpus is freely released on GitHub.
- The corpus mixes Arabic script, Latin-script Arabizi and mixed-script comments. It contains repeated letters, ad-hoc abbreviations, onomatopoeia and emoticons.
- Before annotation, the authors removed comments written entirely in another language and deleted user names and URLs. They stripped the "#" from hashtags but kept the hashtag words.
- Training sets were TSAC (7,145 positive / 6,515 negative), sentence-segmented OCA (MSA, 4,931 / 4,931) and LABR (labelled "D Mix", 4,880 / 4,880). All systems were tested on a balanced held-out TSAC set of 1,700 / 1,700 comments.
- With Doc2vec features (gensim) and scikit-learn classifiers, the best error rates were 0.22 for MLP and 0.23 for SVM; BNB lagged at 0.42. Setups trained without TSAC scored around 0.47–0.52.

## Interesting ideas
- Before collecting new data, the authors measured how far models trained on neighbouring language varieties get on a target-variety test set; the gap they found justified the annotation effort.
- Training sources were tried alone and in every combination against one fixed test set, which showed that adding out-of-variety data gave no benefit.
- The paper shows that a single dialect word can surface in many spellings across two scripts, which makes spelling variation a core problem rather than noise.
- During cleaning, the authors dropped the hashtag symbol but kept the word, so the content stays in the text while the markup goes.
- The evaluation set is balanced between the two classes, so a single error rate is easy to read and chance level is obvious (0.5).

## For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

- **I1** (learn): Before building genre-specific silver data, score parsers trained on other English treebanks against a small held-out sample of the target genre, so the need for new data is measured rather than assumed.
- **I2** (learn): Run the same kind of source ablation when mixing silver data from different witness parsers or genres, and drop any source that does not move scores on the fixed evaluation set.
- **I3** (tokens): Add a variant-normalization step before lemmatization, so that stretched letters, digit-for-letter spellings and casual abbreviations in informal English map to one lemma that the tagger can handle.
- **I4** (tokens): The tokenizer could split "#word" into a markup token and a real word, so hashtags still take part in the UD tree instead of becoming unknown tokens.
- **I5** (learn): Build balanced evaluation sets for binary judgments such as irony or speech-act checks, so a negative-control baseline sits at a known chance level.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

## Institutions

- [Le Mans Université](https://md.research.elamur.ai/free/en/universities/le-mans-universite/) — University, Le Mans, France · 16 works · 31,097 citations
- [Laboratoire d’Informatique de l’Université du Mans](https://md.research.elamur.ai/free/en/labs/laboratoire-d-informatique-de-l-universite-du-mans/) — Research institute, Le Mans, France · 13 works · 2,995 citations

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 17,965 works · [💭 Psychology](https://md.research.elamur.ai/free/en/shelves/psychology/) · 1,649 works · [💭 Experimental and Cognitive Psychology](https://md.research.elamur.ai/free/en/shelves/psychology/experimental-and-cognitive-psychology/) · 930 works

🧭 Topics (2): [😊 Sentiment & emotion](https://md.research.elamur.ai/free/en/topics/sentiment-emotion/) · 1,035 works · [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) · 3,066 works

## Cite

Cite as: Salima Medhaffar, Fethi Bougares, Yannick Estève and Lamia Hadrich-Belguith. 2017. Sentiment Analysis of Tunisian Dialects: Linguistic Ressources and Experiments. Proceedings of the Third Arabic Natural Language Processing Workshop.

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{medhaffar2017sentiment,
  title = {{Sentiment Analysis of Tunisian Dialects: Linguistic Ressources and Experiments}},
  author = {Salima Medhaffar, Fethi Bougares, Yannick Estève and Lamia Hadrich-Belguith},
  booktitle = {Proceedings of the Third Arabic Natural Language Processing Workshop},
  year = {2017},
  doi = {10.18653/v1/w17-1307},
  url = {https://doi.org/10.18653/v1/w17-1307}
}
```

- [← previous Neural Discourse Structure for Text Categorization](https://md.research.elamur.ai/free/en/papers/ji-2017-neural-discourse-structure-text-categorization/)
- [next → Multi-Domain Neural Machine Translation through Unsupervised Adaptation](https://md.research.elamur.ai/free/en/papers/farajian-2017-multi-domain-neural-machine-translation/)
- [↑ back to papers of 2017](https://md.research.elamur.ai/free/en/papers/2017/page/3/)

## References

18 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Muhammad Abdul-Mageed, Sandra Kübler, and Mona Diab. 2012. [Samar: A system for subjectivity and sentiment analysis of arabic social media](https://md.research.elamur.ai/free/en/papers/mageed-2012-samar-system-subjectivity-sentiment-analysis/). In Proceedings of the 3rd workshop in computational approaches to subjectivity and sentiment analysis, pages 19–28. Association for Computational Linguistics.
2. Nawaf A Abdulla, Nizar A Ahmed, Mohammed A Shehab, and Mahmoud Al-Ayyoub. 2013. Arabic sentiment analysis: Lexicon-based and corpus-based. In Applied Electrical Engineering and Computing Technologies (AEECT), 2013 IEEE Jordan Conference on, pages 1–6. IEEE. [find](https://scholar.google.com/scholar?q=Nawaf+A+Abdulla%2C+Nizar+A+Ahmed%2C+Mohammed+A+Shehab%2C+and+Mahmoud+Al%2DAyyoub%2E+2013%2E+Arabic+sentiment+analysis%3A+Lexicon%2Dbased+and+corpus%2Dbased%2E+In+Applied+Electrical+Engineering+and+Computing+Technologies+)
3. Mahmoud Al-Ayyoub, Safa Bani Essa, and Izzat Alsmadi. 2015. Lexicon-based sentiment analysis of arabic tweets. International Journal of Social Network Mining, 2(2):101–114. • The OCA and LABR data sets are limited to one domain (movies and books respectively), while the evaluation set is multi-domain. [find](https://scholar.google.com/scholar?q=Mahmoud+Al%2DAyyoub%2C+Safa+Bani+Essa%2C+and+Izzat+Alsmadi%2E+2015%2E+Lexicon%2Dbased+sentiment+analysis+of+arabic+tweets%2E+International+Journal+of+Social+Network+Mining%2C+2%282%29%3A101%E2%80%93114%2E+%E2%80%A2+The+OCA+and+LABR+data+set)
4. Tareq Al-Moslmi, Mohammed Albared, Adel Al-Shabi, Nazlia Omar, and Salwani Abdullah. 2017. Arabic senti-lexicon: Constructing publicly available language resources for arabic sentiment analysis. Journal of Information Science, page 0165551516683908. • The OCA and LABR data sets are written only in Arabic character, while the evaluation set contains Latin character. [find](https://scholar.google.com/scholar?q=Tareq+Al%2DMoslmi%2C+Mohammed+Albared%2C+Adel+Al%2DShabi%2C+Nazlia+Omar%2C+and+Salwani+Abdullah%2E+2017%2E+Arabic+senti%2Dlexicon%3A+Constructing+publicly+available+language+resources+for+arabic+sentiment+analysis%2E+Journ)
5. Adel Assiri, Ahmed Emam, and Hmood Aldossari. 2015. Arabic sentiment analysis: A survey. International Journal of Advanced Computer Science and Applications, 6(12). • The lexical differences between Tunisian dialect, MSA and other dialects.For example, the English word beautiful, is written in Tunisian: TažAa§Åzi› /mizoyaAnap, in Egyptian : ÅÁw"li / Hilowapo and in MSA : ¿Talyima / jamiylapN) [find](https://scholar.google.com/scholar?q=Adel+Assiri%2C+Ahmed+Emam%2C+and+Hmood+Aldossari%2E+2015%2E+Arabic+sentiment+analysis%3A+A+survey%2E+International+Journal+of+Advanced+Computer+Science+and+Applications%2C+6%2812%29%2E+%E2%80%A2+The+lexical+differences+between+T)
6. Amine Bayoudhi, Hatem Ghorbel, Houssem Koubaa, and Lamia Hadrich Belguith. 2015. Sentiment classification at discourse segment level: Experiments on multi-domain arabic corpus. Journal for Language Technology and Computational Linguistics, page 1. Table 7 shows several outputs of our SA system with MLP classifier. We present examples for Positive and Negative classes and for both situation : when SA predict the correct polarity and when SA system fails. [find](https://scholar.google.com/scholar?q=Amine+Bayoudhi%2C+Hatem+Ghorbel%2C+Houssem+Koubaa%2C+and+Lamia+Hadrich+Belguith%2E+2015%2E+Sentiment+classification+at+discourse+segment+level%3A+Experiments+on+multi%2Ddomain+arabic+corpus%2E+Journal+for+Language+Te)
7. Mariam Biltawi, Wael Etaiwi, Sara Tedmori, and Amjad Hudaib andArafat Awajan. 2016. Sentimnt classification techniques for arabic language: A survey. Conclusions and feature work Kareem Darwish. 2014. [Arabizi detection and conversion to arabic](https://md.research.elamur.ai/free/en/papers/darwish-2014-arabizi-detection-conversion-arabic/). ANLP 2014, page 217. In this paper we have presented the first freely available annotated sentiment analysis corpus for the Tunisian dialect. We have experimented and presented several SA experiments with different training configurations. Best results for Tunisian Hady ElSahar and Samhaa R El-Beltagy. 2015. Building large arabic multi-domain resources for sentiment analysis. In International Conference on Intelligent Text Processing and Computational Linguistics, pages 23–34. Springer. User comment TVAO› An ŸyO’Až A› ks§ w˜wl’ ™fW˜ A¡ –F C y› d`q !! XFA› — C œl`› A§ Alah la trabahkom la daniya w la a5ra Rajell w m3alem tounsi wakahou Ayby˜ Ÿ› Tyt˜ ™• ¨k˜¤ šAm˜ Ÿ› T§
 rA Tn›
 •AF ¨tž¤ ¨t –ylˆ yKž ¢l˜ ºAK› whlbh›A§ Cwf` d§d˜ ¤d§rž Ÿk˜ ™ymt˜ ¨ Tnkmt›¤ EAtm› Mitrobi bsaraha . 5iiiit ech nakraha ???A¡®ˆ wyk ¨˜ ryb`t˜ T§r ¨¡ ©@¡ System output POS POS POS POS POS Reference NEG NEG NEG POS POS NEG NEG NEG NEG NEG POS POS POS NEG NEG *Table 7: Output examples of Tunisian SA system. For each example we present the predicted output and the reference.* Quoc V Le and Tomas Mikolov. 2014. Distributed representations of sentences and documents. In ICML, volume 14, pages 1188–1196.
8. Technologies and Systems (CTS), 2012 International Conference on, pages 546–550. IEEE. [find](https://scholar.google.com/scholar?q=Technologies+and+Systems+%28CTS%29%2C+2012+International+Conference+on%2C+pages+546%E2%80%93550%2E+IEEE%2E)
9. Rabih Zbib, Erika Malchiodi, Jacob Devlin, David Stallard, Spyros Matsoukas, Richard Schwartz, John Makhoul, Omar F Zaidan, and Chris Callison-Burch. 2012. [Machine translation of arabic dialects](https://md.research.elamur.ai/free/en/papers/zbib-2012-machine-translation-arabic-dialects/). In Proceedings of the 2012 conference of the north american chapter of the association for computational linguistics: Human language technologies, pages 49–59. Association for Computational Linguistics.
10. Mohamed Maamouri, Ann Bies, Seth Kulick, Michael Ciul, Nizar Habash, and Ramy Eskander. 2014. Developing an egyptian arabic treebank: Impact of dialectal morphology on annotation and tool development. In LREC, pages 2348–2354. [find](https://scholar.google.com/scholar?q=Mohamed+Maamouri%2C+Ann+Bies%2C+Seth+Kulick%2C+Michael+Ciul%2C+Nizar+Habash%2C+and+Ramy+Eskander%2E+2014%2E+Developing+an+egyptian+arabic+treebank%3A+Impact+of+dialectal+morphology+on+annotation+and+tool+development%2E)
11. Salameh Mohammad, M Mohammad Saif, and Svetlana Kiritchenko. 2015. [Sentiment after translation: A case-study on arabic social media posts](https://md.research.elamur.ai/free/en/papers/salameh-2015-sentiment-after-translation-case-study/). In Proceedings of the North American Chapter of the Association for Computational Linguistics (NAACL-2015).
12. Inès Zribi, Rahma Boujelbane, Abir Masmoudi, Mariem Ellouze, Lamia Hadrich Belguith, and Nizar Habash. 2014. [A conventional orthography for tunisian arabic](https://md.research.elamur.ai/free/en/papers/zribi-2014-conventional-orthography-tunisian-arabic/). In Proceedings of the Language Resources and Evaluation Conference (LREC), Reykjavik, Iceland, pages 2355–2361.
13. Mahmoud Nabil, Mohamed A Aly, and Amir F Atiya. 2014. [Labr: A large scale arabic book reviews dataset](https://md.research.elamur.ai/free/en/papers/nabil-2014-labr-large-scale-arabic-sentiment/). CoRR, abs/1411.6718. [↗](https://arxiv.org/abs/1411.6718)
14. Mahmoud Nabil, Mohamed Aly, and Amir F Atiya. 2015. [Astd: Arabic sentiment tweets dataset](https://md.research.elamur.ai/free/en/papers/nabil-2015-astd-arabic-sentiment-tweets-dataset/). In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2515–2519.
15. Mohammed Rushdi-Saleh, M Teresa Martı́n-Valdivia, L Alfonso Ureña-López, and José M Perea-Ortega. 2011a. Oca: Opinion corpus for arabic. Journal of the American Society for Information Science and Technology, 62(10):2045–2054. [find](https://scholar.google.com/scholar?q=Mohammed+Rushdi%2DSaleh%2C+M+Teresa+Mart%C4%B1%CC%81n%2DValdivia%2C+L+Alfonso+Uren%CC%83a%2DLo%CC%81pez%2C+and+Jose%CC%81+M+Perea%2DOrtega%2E+2011a%2E+Oca%3A+Opinion+corpus+for+arabic%2E+Journal+of+the+American+Society+for+Information+Science+and+)
16. Mohammed Rushdi-Saleh, Maria Teresa Martı́n-Valdivia, L Alfonso Ureña-López, and José M Perea-Ortega. 2011b. Bilingual experiments with an arabic-english corpus for opinion mining. [find](https://scholar.google.com/scholar?q=Mohammed+Rushdi%2DSaleh%2C+Maria+Teresa+Mart%C4%B1%CC%81n%2DValdivia%2C+L+Alfonso+Uren%CC%83a%2DLo%CC%81pez%2C+and+Jose%CC%81+M+Perea%2DOrtega%2E+2011b%2E+Bilingual+experiments+with+an+arabic%2Denglish+corpus+for+opinion+mining%2E)
17. Fadi Salem. 2017. The arab social media report 2017: Social media and the internet of things: Towards data-driven policymaking in the arab world. Dubai: MBR School of Government., 7. [find](https://scholar.google.com/scholar?q=Fadi+Salem%2E+2017%2E+The+arab+social+media+report+2017%3A+Social+media+and+the+internet+of+things%3A+Towards+data%2Ddriven+policymaking+in+the+arab+world%2E+Dubai%3A+MBR+School+of+Government%2E%2C+7%2E)
18. Amira Shoukry and Ahmed Rafea. 2012. Sentence-level arabic sentiment analysis. In Collaboration [find](https://scholar.google.com/scholar?q=Amira+Shoukry+and+Ahmed+Rafea%2E+2012%2E+Sentence%2Dlevel+arabic+sentiment+analysis%2E+In+Collaboration)

*Updated 2 November 2026*