Elamur Research⌕
☰
home / papers / 2023 / Content-Localization based Neural Machine Translation for Informal Dialectal Arabic: Spanish/French to Levantine/Gulf Arabic

Content-Localization based Neural Machine Translation for Informal Dialectal Arabic: Spanish/French to Levantine/Gulf Arabic

Summary

Dialectal Arabic is heavily used on social media, yet it has few labelled resources for analysing online social behaviour such as sentiment or hate speech. Spanish and French already have such resources. The authors propose reusing them through "content localization." Instead of translating literally, the approach carries a message's context, tone, slang and cultural cues into a specific target dialect. They build SF-ArLG, a parallel corpus in which native Spanish and French tweets were rendered into Levantine and Gulf Arabic by professional translators who are native dialect speakers and active social-media users. On this corpus they train neural MT models in both directions. They then use the localized data to train sentiment and hate-speech classifiers for each dialect, with no new annotation. The authors report that the resulting classifiers work reliably. Levantine and Gulf models behaved differently on hate content, which suggests that merging dialects of one language can distort behavioural analysis. This matters because it offers a cheap way to bootstrap resources for under-resourced varieties.

Key points

  • SF-ArLG is presented as the first parallel corpus between informal Spanish/French and informal Arabic dialects, specifically Levantine and Gulf.
  • Source data: the Spanish and French tweets were themselves translations of English tweets from the DFMSD dataset.
  • Translation guidelines set several rules:
    • keep emojis and emoticons;
    • render slang and abbreviations;
    • translate hashtag words while keeping their order and context;
    • allow code borrowing, such as writing "lol" in Arabic script;
    • transliterate proper nouns;
    • finish with a proofreading pass.
  • Models: NMT systems were trained for French↔Levantine, Spanish↔Levantine and Spanish↔Gulf, with sample outputs produced via a transitive translation route.
  • Evaluation: the models were tested downstream on sentiment and hate-speech classification. The Levantine hate model caught Levantine hate tweets that the Gulf model misclassified, because some insults are used only in Levantine speech.

Interesting ideas

  • I1. Evaluate a translation resource by what it enables downstream: train task classifiers on the localized data and check that they still separate the classes, which is more telling than surface overlap metrics. — For Mova (back): Mova's translation through the meta-world could be judged by whether readers answering questions on the translated text keep the same answers as on the original.
  • I2. Annotate once in a rich language and move the labels into a poor variety by localization, avoiding a fresh labelling effort for every dialect. — For Mova (learn): silver training data for rarer English registers (tweets, dialect text) could be produced by localizing already-labelled standard text, with the labels carried over.
  • I3. Treat dialect as a meaning variable: the same word can denote different things (a Gulf "friend" means "boyfriend" in Lebanese), so dialects should not be pooled blindly. — For Mova (world): lexical senses in the knowledge layer could carry a dialect or register tag, so that entity and relation readings depend on the detected variety.
  • I4. Write explicit guidelines for social-media artefacts: keep emojis, translate hashtag words in place and order, and allow borrowed words in native script. — For Mova (tokens): the tokenizer and tagger could split hashtags into words and tag emojis and borrowed slang as first-class tokens rather than noise.
  • I5. A per-dialect classifier that fails on another dialect's insults works as a natural diagnostic for hidden distribution shift. — For Mova (frames): irony and speech-act detection could be tested on register-shifted inputs as a control, exposing where pragmatic cues are dialect-specific.

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

Fields: Linguistics & languages

Topics: Machine translation

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2312.06926 · arXiv 2312.06926

@misc{alzamzami2023content,
  title = {{Content-Localization based Neural Machine Translation for Informal Dialectal Arabic: Spanish/French to Levantine/Gulf Arabic}},
  author = {Alzamzami, Fatimah and Saddik, Abdulmotaleb El},
  year = {2023},
  doi = {10.48550/arxiv.2312.06926},
  eprint = {2312.06926},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2312.06926}
}

References

35 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Mukesh Kumar Saini, Fatimah Al-Zamzami, and Abdulmotaleb El Saddik. Towards storytelling by extracting social information from osn photo’s metadata. In Proceedings of the First International Workshop on Internet-Scale Multimedia Management, pages 15–20, 2014. find
  2. Fatimah Alzamzami, Mukesh Saini, and Abdulmotaleb El Saddik. Dst: days spent together using soft sensory information on osns—a case study on facebook. Soft Computing, 21(15):4227–4238, 2017. find
  3. Vineet John and Olga Vechtomova. Sentiment analysis on financial news headlines using training dataset augmentation. arXiv preprint arXiv:1707.09448, 2017. ↗
  4. Hind Saleh Alatawi, Areej Maatog Alhothali, and Kawthar Mustafa Moria. Detecting white supremacist hate speech using domain specific word embedding with deep learning and bert. arXiv e-prints, pages arXiv–2010, 2020. find
  5. Meshrif Alruily and Osama R Shahin. Sentiment analysis of twitter data for saudi universities. International Journal of Machine Learning and Computing, 10(1), 2020. find our card
  6. Fatimah Alzamzami and Abdulmotaleb El Saddik. Monitoring cyber sentihate social behavior during covid-19 pandemic in north america. IEEE Access, 2021. find
  7. Hassan Sajjad, Ahmed Abdelali, Nadir Durrani, and Fahim Dalvi. Arabench: Benchmarking dialectal arabic-english machine translation. In Proceedings of the 28th International Conference on Computational Linguistics, pages 5094–5107, 2020. find
  8. Hadry. French twitter sentiment analysis. https://kaggle.com/hbaflast/french-twitter-sentiment-analysis, 2020. Accessed: (August 15, 2023). ↗
  9. Ramy Baly, Alaa Khaddaj, Hazem Hajj, Wassim El-Hajj, and Khaled Bashir Shaban. Arsentd-lev: A multi-topic corpus for target-based sentiment analysis in arabic levantine tweets. arXiv preprint arXiv:1906.01830, 2019. ↗ our card
  10. Rana Abaalkhail, Fatimah Alzamzami, Samah Aloufi, Rajwa Alharthi, and Abdulmotaleb El Saddik. Affectional ontology and multimedia dataset for sentiment analysis. In International Conference on Smart Multimedia, pages 15–28. Springer, 2018. find
  11. Hatem Haddad, Hala Mulki, and Asma Oueslati. T-hsab: A tunisian hate speech and abusive dataset. In International Conference on Arabic Language Processing, pages 251–263. Springer, 2019. find
  12. Ibrahim Abu Farha and Walid Magdy. From arabic sentiment analysis to sarcasm detection: The arsarcasm dataset. In Proceedings of the 4th Workshop on Open-Source Arabic Corpora and Processing Tools, with a Shared Task on Offensive Language Detection, pages 32–39, 2020. find our card
  13. Muhammad Abdul-Mageed and Mona T Diab. Sana: A large scale multi-genre, multi-dialect lexicon for arabic subjectivity and sentiment analysis. In LREC, pages 1162–1169, 2014. find
  14. Hala Mulki and Bilal Ghanem. Let-mi: an arabic levantine twitter dataset for misogynistic language. arXiv preprint arXiv:2103.10195, 2021. ↗
  15. Fatimah Alzamzami, Mohamad Hoda, and Abdulmotaleb El Saddik. Light gradient boosting machine for general sentiment classification on short texts: A comparative evaluation. IEEE Access, 2020. find
  16. Surangika Ranathunga, En-Shiun Annie Lee, Marjana Prifti Skenduli, Ravi Shekhar, Mehreen Alam, and Rishemjit Kaur. Neural machine translation for low-resource languages: A survey. ACM Computing Surveys, 55(11):1–37, 2023. find
  17. Mohamed El-Madkouri and Beatriz Soto Aranda. Readability and communication in machine translation of arabic phraseologisms into spanish. In International Conference on Computational and Corpus-Based Phraseology, pages 78–89. Springer, 2022. find
  18. Arwa Alqudsi, Nazlia Omar, and Khalid Shaker. A hybrid rules and statistical method for arabic to english machine translation. In 2019 2nd International Conference on Computer Applications & Information Security (ICCAIS), pages 1–7. IEEE, 2019. find
  19. Arwa Alqudsi, Nazlia Omar, and Khalid Shaker. Arabic machine translation: a survey. Artificial Intelligence Review, 42:549–572, 2014. find
  20. Jezia Zakraoui, Moutaz Saleh, Somaya Al-Maadeed, and Jihad Mohamad AlJa’am. Evaluation of arabic to english machine translation systems. In 2020 11th International Conference on Information and Communication Systems (ICICS), pages 185–190. IEEE, 2020. find
  21. Rabih Zbib, Erika Malchiodi, Jacob Devlin, David Stallard, Spyros Matsoukas, Richard Schwartz, John Makhoul, Omar Zaidan, and Chris Callison-Burch. Machine translation of 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, 2012. find our card
  22. Houda Bouamor, Nizar Habash, and Kemal Oflazer. A multidialectal parallel corpus of arabic. In LREC, pages 1240–1245, 2014. find our card
  23. Houda Bouamor, Nizar Habash, Mohammad Salameh, Wajdi Zaghouani, Owen Rambow, Dana Abdulrahim, Ossama Obeid, Salam Khalifa, Fadhl Eryani, Alexander Erdmann, et al. The madar arabic dialect corpus and lexicon. In Proceedings of the eleventh international conference on language resources and evaluation (LREC 2018), 2018. find
  24. Mohamed Elmahdy, Mark Hasegawa-Johnson, and Eiman Mustafawi. Development of a tv broadcasts speech recognition system for qatari arabic. In LREC, volume 14, pages 3057–3061, 2014. find
  25. Fatimah Alzamzami and Abdulmotaleb El Saddik. Osn-mdad: Machine translation dataset for arabic multi-dialectal conversations on online social media. arXiv preprint arXiv:2309.12137, 2023. ↗ our card
  26. Fatimah Alzamzami and Abdulmotaleb El Saddik. Content-localization based system for analyzing sentiment and hate behaviors in low-resource dialectal arabic: English to levantine and gulf. arXiv preprint arXiv:2312.03727, 2023. ↗ our card
  27. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998–6008, 2017. find
  28. Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. Multilingual denoising pre-training for neural machine translation. Transactions of the Association for Computational Linguistics, 8:726–742, 2020. find
  29. Yuqing Tang, Chau Tran, Xian Li, Peng-Jen Chen, Naman Goyal, Vishrav Chaudhary, Jiatao Gu, and Angela Fan. Multilingual translation with extensible multilingual pretraining and finetuning. arXiv preprint arXiv:2008.00401, 2020. ↗
  30. Fatimah Alzamzami and Abdulmotaleb El Saddik. Transformer-based feature fusion approach for multimodal visual sentiment recognition using tweets in the wild. IEEE Access, 2023. find
  31. Alexander Rietzler, Sebastian Stabinger, Paul Opitz, and Stefan Engl. Adapt or get left behind: Domain adaptation through bert language model finetuning for aspect-target sentiment classification. arXiv preprint arXiv:1908.11860, 2019. ↗
  32. Go Inoue, Bashar Alhafni, Nurpeiis Baimukan, Houda Bouamor, and Nizar Habash. The interplay of variant, size, and task type in Arabic pre-trained language models. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, Kyiv, Ukraine (Online), April 2021. Association for Computational Linguistics. find our card
  33. Marwan Al Omari. Oclar: logistic regression optimisation for arabic customers’ reviews. International Journal of Business Intelligence and Data Mining, 20(3):251–273, 2022. find
  34. Wajid Hassan Moosa and Najiba. Multi-lingual hatespeech dataset, 2022. find
  35. Hala Mulki, Hatem Haddad, Chedi Bechikh Ali, and Halima Alshabani. L-hsab: A levantine twitter dataset for hate speech and abusive language. In Proceedings of the third workshop on abusive language online, pages 111–118, 2019. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/alzamzami-2023-content-localization-based-neural-machine/