Elamur Research⌕
☰
home / papers / 2019 / SenZi: A Sentiment Analysis Lexicon for the Latinised Arabic (Arabizi)

SenZi: A Sentiment Analysis Lexicon for the Latinised Arabic (Arabizi)

Paper · 2019 · Taha Tobaili, Miriam Fernandez, Harith Alani et al.

Summary

Arabizi, dialectal Arabic typed in Latin letters and digits, is common on social media but has almost no NLP resources. Arabic sentiment studies usually discard it or try to transliterate it into Arabic script first. The authors build SenZi, a sentiment lexicon for Lebanese Arabizi, and apply it to the raw text without filtering, normalisation or transliteration.

From geolocated Lebanese tweets they annotate data for two purposes: training an Arabizi language identifier and evaluating sentiment. They use the identifier to gather a 1M-comment Facebook corpus. A seed list of about 2K sentiment words, obtained by translating, annotating and transliterating existing resources, is then expanded with word embeddings trained on that corpus. The expansion adds inflected and alternatively spelled forms, giving 11.3K positive and 13.3K negative entries. A simple lexicon-based classifier reaches F1 0.72 on Arabizi tweets.

A detailed error analysis shows where lexicon methods break down: sparsity, code-switching, multi-word expressions, ambiguity and negation. The work matters because it releases the first public resources of this kind for Levantine Arabizi.

Key points

  • Collection pipeline:
    • 177K geolocated Lebanese tweets were collected over 2016–2017.
    • Removing Twitter-labelled Arabic left 97K Latin-script tweets, and cleaning reduced these to 66K.
    • 30K tweets were annotated by three Lebanese students for "mostly Arabizi?" (4.3K yes) and for sentiment, with Fleiss' Kappa reported as substantial.
  • Released resources:
    • the SenZi lexicon (11.3K positive and 13.3K negative words),
    • a 3.4K sentiment-annotated tweet set,
    • a 1M Facebook comment corpus.
  • Embedding-based expansion grew the lexicon from 2K to 24.6K words and raised recall by 0.23. The overall lexicon-based F1 is 0.72.
  • About 45% of errors were tweets assigned no sentiment at all. Twitter itself mislabelled Arabizi as Haitian, Turkish, Hindi and other languages.
  • Manual analysis of 200 misclassified tweets found these main causes:
    • about 38% were unmatched inflectional or spelling forms of words already in SenZi;
    • about 12% were English sentiment words;
    • about 14% were neutral words wrongly matched;
    • the remaining causes were multi-word expressions or sarcasm, negation, and ambiguity.

Interesting ideas

  • I1. Instead of writing normalisation rules, the authors grow a small seed lexicon by pulling embedding neighbours that turn out to be inflections and spelling variants of each seed word.
  • I2. Annotators were told to label the sentiment a tweet actually conveys rather than its surface cues, so "haha we failed the exam" counts as negative.
  • I3. The confusion matrix gives "no sentiment assigned" its own column, which made abstention visible as the dominant failure mode rather than hiding it inside errors.
  • I4. Failures were bucketed on a fixed 100+100 sample of misclassified items, with percentages per cause, which turns a single F1 score into a prioritised to-do list.
  • I5. The paper notes that negators may sit several words away from the sentiment word and sometimes only weaken it, while idioms carry sentiment without any sentiment word. Both facts defeat adjacency-based lexicon matching.

For Mova

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

  • I1 (tokens): a similar neighbour search over noisy English (luv, gr8, missin) could propose variant→lemma links for the lemmatiser, with each link kept as an explainable, reviewable entry.
  • I2 (frames): a test set built on this rule would check whether the irony/implicature layer reads the evaluated event rather than the laughter marker.
  • I3 (answer): readers on fable QA or school tests should report "could not answer" separately from wrong answers, so coverage gaps and reasoning mistakes are tracked apart.
  • I4 (learn): each evaluation round could include a manual cause taxonomy over a fixed error sample to decide which repair rules or data to target next.
  • I5 (tree): negation and idiom handling should follow UD dependency arcs (e.g. advmod/neg attachment to the head) rather than word windows, so that scope is structural and explainable.

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

🎓University of MannheimMannheim · Germany29 works · 8,347 citations🎓American University of BeirutBeirut · Lebanon34 works · 2,561 citations🎓The Open UniversityMilton Keynes · United Kingdom10 works · 338 citations🎓Lebanese American UniversityBeirut · Lebanon5 works · 37 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 2Research subjects and languages the work is about.Sentiment & emotion · 972 worksSemitic languages · 2,908 works

Cite

Cite as Taha Tobaili, Miriam Fernandez, Harith Alani, Sanaa Sharafeddine, Hazem Hajj and Goran Glavaš. 2019. SenZi: A Sentiment Analysis Lexicon for the Latinised Arabic (Arabizi). Proceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP 2019).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.26615/978-954-452-056-4_138

@inproceedings{tobaili2019senzi,
  title = {{SenZi: A Sentiment Analysis Lexicon for the Latinised Arabic (Arabizi)}},
  author = {Taha Tobaili, Miriam Fernandez, Harith Alani, Sanaa Sharafeddine, Hazem Hajj and Goran Glavaš},
  booktitle = {Proceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP 2019)},
  year = {2019},
  doi = {10.26615/978-954-452-056-4\_138},
  url = {https://doi.org/10.26615/978-954-452-056-4_138}
}

References

29 references, 12 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 and Mona T Diab. 2014. Sana: A large scale multi-genre, multi-dialect lexicon for arabic subjectivity and sentiment analysis. In LREC, pages 1162–1169.
  2. Mariam Aboelezz. 2009. Latinised arabic and connections to bilingual ability. In Papers from the Lancaster University Postgraduate Conference in Linguistics and Language Teaching. find
  3. Mohamed Al-Badrashiny, Ramy Eskander, Nizar Habash, and Owen Rambow. 2014. Automatic transliteration of romanized dialectal arabic. In Proceedings of the Eighteenth Conference on Computational Natural Language Learning, pages 30–38.
  4. Mohammed Al-Kabi, Amal Gigieh, Izzat Alsmadi, Heider Wahsheh, and Mohamad Haidar. 2013. An opinion analysis tool for colloquial and standard arabic. In The Fourth International Conference on Information and Communication Systems (ICICS 2013), pages 23–25. https://project-rbz.kmi.open.ac.uk ↗
  5. Mohammed N Al-Kabi, Amal H Gigieh, Izzat M Alsmadi, Heider A Wahsheh, and Mohamad M Haidar. 2014. Opinion mining and analysis for arabic language. International Journal of Advanced Computer Science and Applications (IJACSA), SAI Publisher, 5(5). ↗
  6. Nora Al-Twairesh, Hend Al-Khalifa, and Abdulmalik AlSalman. 2016. Arasenti: Large-scale twitter-specific arabic sentiment lexicons. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), volume 1, pages 697–705.
  7. Wid H Allehaiby. 2013. Arabizi: An analysis of the romanization of the arabic script from a sociolinguistic perspective. Arab World English Journal, 4(3):52– 62. find
  8. Gilbert Badaro, Ramy Baly, Hazem Hajj, Nizar Habash, and Wassim El-Hajj. 2014. A large scale arabic sentiment lexicon for arabic opinion mining. In Proceedings of the EMNLP 2014 workshop on arabic natural language processing (ANLP), pages 165–173.
  9. Ramy Eskander and Owen Rambow. 2015. Slsa: A sentiment lexicon for standard arabic. In Proceedings of the 2015 conference on empirical methods in natural language processing, pages 2545–2550.
  10. Joseph L Fleiss. 1971. Measuring nominal scale agreement among many raters. Psychological bulletin, 76(5):378. ↗
  11. Imane GUELLIL, Ahsan Adeel, Faical AZOUAOU, Ala-eddine Hachani, Amir Hussain, et al. 2018. Arabizi sentiment analysis based on transliteration and automatic corpus annotation. In Proceedings of the 9th Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis, pages 335–341. ↗
  12. Imane Guellil, Faiçal Azouaou, Mourad Abbas, and Sadat Fatiha. 2017. Arabizi transliteration of algerian arabic dialect into modern standard arabic. In Social MT 2017/First workshop on Social Media and User Generated Content Machine Translation. find
  13. Basis-Technology. 2012. The burgeoning challenge of deciphering arabic chat. find
  14. Yuen Chee Keong, Othman Rahsid Hameed, and Imad Amer Abdulbaqi. 2015. The use of arabizi in english texting by arab postgraduate students at UKM. The English Literature Journal, 2(2):281– 288. find
  15. Robert Michael BIANCHI. 2012. 3arabizi-when local arabic meets global english. Acta Linguistica Asiatica, 2(1):89–100. ↗
  16. J Richard Landis and Gary G Koch. 1977. The measurement of observer agreement for categorical data. biometrics, pages 159–174. find
  17. Ann Bies, Zhiyi Song, Mohamed Maamouri, Stephen Grimes, Haejoong Lee, Jonathan Wright, Stephanie Strassel, Nizar Habash, Ramy Eskander, and Owen Rambow. 2014. Transliteration of arabizi into arabic orthography: Developing a parallel annotated arabizi-arabic script sms/chat corpus. In Proceedings of the EMNLP 2014 Workshop on Arabic Natural Langauge Processing (ANLP), pages 93–103.
  18. Bing Liu. 2012. Sentiment analysis and opinion mining. Synthesis lectures on human language technologies, 5(1):1–167. ↗
  19. Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2016. Enriching word vectors with subword information. arXiv preprint arXiv:1607.04606. ↗
  20. Randa Muhammed, Mona Farrag, Nariman Elshamly, and Nady Abdel-Ghaffar. 2011. Summary of arabizi or romanization: The dilemma of writing arabic texts. In Jı̄l Jadı̄d Conference, University of Texas at Austin, pages 18–19. find
  21. Kareem Darwish. 2014. Arabizi detection and conversion to arabic. ANLP 2014, page 217. ↗
  22. Rehab M Duwairi, Mosab Alfaqeh, Mohammad Wardat, and Areen Alrabadi. 2016. Sentiment analysis for arabizi text. In 2016 7th International Conference on Information and Communication Systems (ICICS), pages 127–132. IEEE. find
  23. Rehab M Duwairi and Islam Qarqaz. 2014. Arabic sentiment analysis using supervised classification. In Future Internet of Things and Cloud (FiCloud), 2014 International Conference on, pages 579–583. IEEE. find
  24. Ramy Eskander, Mohamed Al-Badrashiny, Nizar Habash, and Owen Rambow. 2014. Foreign words and the automatic processing of arabic social media text written in roman script. EMNLP 2014, page 1.
  25. M’hamed Mataoui, Omar Zelmati, and Madiha Boumechache. 2016. A proposed lexicon-based sentiment analysis approach for the vernacular algerian arabic. Res. Comput. Sci, 110:55–70. ↗
  26. Natalie Sullivan. 2017. Writing Arabizi: Orthographic Variation in Romanized Lebanese Arabic on Twitter. Ph.D. thesis. find
  27. Taha Tobaili. 2016. Arabizi identification in twitter data. ACL 2016, page 51.
  28. Theresa Wilson, Janyce Wiebe, and Paul Hoffmann. 2005. Recognizing contextual polarity in phrase-level sentiment analysis. In Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing.
  29. Mohammad Ali Yaghan. 2008. “arabizi”: A contemporary style of arabic slang. Design Issues, 24(2):39–52. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/tobaili-2019-senzi-sentiment-analysis-lexicon-latinised/