---
title: "Impact of emoji exclusion on the performance — Aleryani 2024"
description: "The paper asks whether emojis actually help transformer classifiers flag sarcastic or offensive Arabic social-media posts."
authors: ["Ghalyah H. Aleryani", "Wael Deabes", "Khaled Albishre", "Alaa E. Abdel-Hakim"]
year: 2024
venue: "arXiv:2405.02195"
source: "https://arxiv.org/abs/2405.02195"
doi: "10.48550/arxiv.2405.02195"
arxiv: "2405.02195"
licence: "arXiv — non-exclusive license to distribute"
licence_url: "http://arxiv.org/licenses/nonexclusive-distrib/1.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/aleryani-2024-impact-emoji-exclusion-performance-arabic/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/aleryani-2024-impact-emoji-exclusion-performance-arabic/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2024](https://md.research.elamur.ai/free/en/papers/2024/page/3/) / Impact of emoji exclusion on the performance of Arabic sarcasm detection models

# Impact of emoji exclusion on the performance of Arabic sarcasm detection models

**Authors:** Ghalyah H. Aleryani, Wael Deabes, Khaled Albishre, Alaa E. Abdel-Hakim · **Year:** 2024 · **Venue:** arXiv:2405.02195 · **Licence:** [arXiv — non-exclusive license to distribute](http://arxiv.org/licenses/nonexclusive-distrib/1.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/2405.02195) · [PDF](https://arxiv.org/pdf/2405.02195) · [DOI 10.48550/arxiv.2405.02195](https://doi.org/10.48550/arxiv.2405.02195) · [arXiv 2405.02195](https://arxiv.org/abs/2405.02195) · [OpenAlex](https://openalex.org/W4396786006)

**Citations:** 0 citations (openalex-snapshot · 2026-10-14)

## Summary
The paper asks whether emojis actually help transformer classifiers flag sarcastic or offensive Arabic social-media posts. The authors argue that Arabic's large lexicon, many dialects and reliance on puns and context may already carry the signal that emojis add, so emojis could be redundant or even harmful. They test this by fine-tuning three AraBERT variants twice: once on data with emojis kept and once with emojis stripped during preprocessing. The variants are AraBERT_v2, AraBERT_v2_Twitter and multi_dialect_bert_base_arabert, and the datasets are SemEval, YouTube and L-HSAB. They report accuracy, recall, precision and F1 for each pairing. In most configurations, removing emojis gives equal or better scores, especially in accuracy and precision. There are some exceptions, mainly on L-HSAB. The authors conclude that dropping emojis is a positive or at least neutral preprocessing choice for Arabic. They suggest future systems that interpret emojis together with text rather than simply discarding them.

## Key points
- The study fine-tunes three models: AraBERT_v2 (v2), AraBERT_v2_Twitter (TW) and multi_dialect_bert_base_arabert (MD). Each is evaluated on SemEval, YouTube and L-HSAB, with and without emojis.
- Accuracy was higher without emojis in every case except TW on L-HSAB.
- The other metrics showed mixed results:
  - **Recall:** v2 did slightly better with emojis on YouTube and SemEval.
  - **Precision:** v2 and TW gained a little from emojis on L-HSAB.
  - **F1:** v2's score dipped slightly on L-HSAB once emojis were removed.
- The motivation rests on a root-count comparison. The paper cites about 23,090 roots for Arabic, 8,400 for English and 450 for Russian. It also cites earlier evidence that surplus input data can degrade classifiers.
- The framing mixes sarcasm with offensive and hate speech. L-HSAB, for example, is a hate-speech corpus. The paper's claim of "consistent" improvement is somewhat stronger than its own per-metric exceptions support.

## Interesting ideas
- Isolating one preprocessing step and toggling it on and off, with everything else held fixed, gives a clean causal read on whether that step earns its place. — For Mova (learn): Mova could run the same on/off ablation for individual normalisation or repair rules, with a negative control, before adopting them as defaults.
- Reporting every model × dataset × metric cell separately exposes where an intervention helps precision but costs recall, which a single averaged score would hide. — For Mova (learn): Evaluation reports for parser or repair-rule changes could break results down per treebank and per metric (e.g. LAS vs. label-specific recall) so trade-offs stay visible.
- The hypothesis that a redundant signal channel can add noise rather than information is worth testing whenever a cheap side feature duplicates what the text already says. — For Mova (frames): For irony detection in prag, Mova could check whether markers like emoticons or exclamation marks actually improve irony frames beyond what the dependency tree and lexical cues already give.
- Comparing general, social-media-trained and dialect-trained variants of one base model separates domain effects from the effect of the preprocessing change. — For Mova (tree): When comparing the perceptron parser and gparse on informal English, Mova could pair each with domain-specific training data to see whether gains come from the model or the domain.
- The authors' own future-work idea is to interpret emojis jointly with the text rather than delete them, which treats non-lexical symbols as meaningful tokens instead of noise. — For Mova (tokens): The tokenizer could keep emojis as typed tokens with a lexical gloss so that later stages can decide whether they carry sentiment or irony, instead of silently dropping them.

*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](https://md.research.elamur.ai/free/en/shelves/linguistics/)

## Cite

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

```bibtex
@misc{aleryani2024impact,
  title = {{Impact of emoji exclusion on the performance of Arabic sarcasm detection models}},
  author = {Aleryani, Ghalyah H. and Deabes, Wael and Albishre, Khaled and Abdel-Hakim, Alaa E.},
  year = {2024},
  doi = {10.48550/arxiv.2405.02195},
  eprint = {2405.02195},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2405.02195}
}
```

- [← previous HybridFlow: A Flexible and Efficient RLHF Framework](https://md.research.elamur.ai/free/en/papers/sheng-2024-hybridflow-flexible-efficient-rlhf-framework/)
- [next → Improve Mathematical Reasoning in Language Models by Automated Process Supervision](https://md.research.elamur.ai/free/en/papers/luo-2024-improve-mathematical-reasoning-language-models/)
- [↑ back to papers of 2024](https://md.research.elamur.ai/free/en/papers/2024/page/3/)

## References

27 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Vincent Dumas. Enigma machines: Deep learning algorithms as information content providers under section 230 of the communications decency act. Wis. L. Rev., page 1581, 2022. [find](https://scholar.google.com/scholar?q=Vincent+Dumas%2E+Enigma+machines%3A+Deep+learning+algorithms+as+information+content+providers+under+section+230+of+the+communications+decency+act%2E+Wis%2E+L%2E+Rev%2E%2C+page+1581%2C+2022%2E)
2. Michael Baggs. Online hate speech rose 20% during pandemic:‘we’ve normalised it.’. BBC News, November, 15:2021, 2021. [find](https://scholar.google.com/scholar?q=Michael+Baggs%2E+Online+hate+speech+rose+20%25+during+pandemic%3A%E2%80%98we%E2%80%99ve+normalised+it%2E%E2%80%99%2E+BBC+News%2C+November%2C+15%3A2021%2C+2021%2E)
3. Vidyullatha Sukhavasi and Venkatesulu Dondeti. Effective automated transformer model based sarcasm detection using multilingual data. Multimedia Tools and Applications, pages 1–32, 2023. [find](https://scholar.google.com/scholar?q=Vidyullatha+Sukhavasi+and+Venkatesulu+Dondeti%2E+Effective+automated+transformer+model+based+sarcasm+detection+using+multilingual+data%2E+Multimedia+Tools+and+Applications%2C+pages+1%E2%80%9332%2C+2023%2E)
4. Soukaina Mihi, Brahim Ait Ben Ali, Ismail El Bazi, Sara Arezki, and Nabil Laachfoubi. Automatic sarcasm detection in dialectal arabic using bert and tf-idf. In The Proceedings of the International Conference on Smart City Applications, pages 837–847. Springer, 2021. [find](https://scholar.google.com/scholar?q=Soukaina+Mihi%2C+Brahim+Ait+Ben+Ali%2C+Ismail+El+Bazi%2C+Sara+Arezki%2C+and+Nabil+Laachfoubi%2E+Automatic+sarcasm+detection+in+dialectal+arabic+using+bert+and+tf%2Didf%2E+In+The+Proceedings+of+the+International+Con)
5. José Antonio García-Díaz, Salud María Jiménez-Zafra, Miguel Angel García-Cumbreras, and Rafael Valencia-García. Evaluating feature combination strategies for hate-speech detection in spanish using linguistic features and transformers. Complex & Intelligent Systems, 9(3):2893–2914, 2023. [find](https://scholar.google.com/scholar?q=Jos%C3%A9+Antonio+Garc%C3%ADa%2DD%C3%ADaz%2C+Salud+Mar%C3%ADa+Jim%C3%A9nez%2DZafra%2C+Miguel+Angel+Garc%C3%ADa%2DCumbreras%2C+and+Rafael+Valencia%2DGarc%C3%ADa%2E+Evaluating+feature+combination+strategies+for+hate%2Dspeech+detection+in+spanish+using+lin)
6. MV Koroteev. Bert: a review of applications in natural language processing and understanding. arXiv preprint arXiv:2103.11943, 2021. [↗](https://arxiv.org/abs/2103.11943)
7. 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](https://scholar.google.com/scholar?q=Ibrahim+Abu+Farha+and+Walid+Magdy%2E+From+arabic+sentiment+analysis+to+sarcasm+detection%3A+The+arsarcasm+dataset%2E+In+Proceedings+of+the+4th+Workshop+on+Open%2DSource+Arabic+Corpora+and+Processing+Tools%2C+wi) [our card](https://md.research.elamur.ai/free/en/papers/farha-2020-arabic-sentiment-analysis-sarcasm-detection/)
8. Alaa Rahma, Shahira Shaaban Azab, and Ammar Mohammed. A comprehensive review on arabic sarcasm detection: Approaches, challenges and future trends. IEEE Access, 2023. [find](https://scholar.google.com/scholar?q=Alaa+Rahma%2C+Shahira+Shaaban+Azab%2C+and+Ammar+Mohammed%2E+A+comprehensive+review+on+arabic+sarcasm+detection%3A+Approaches%2C+challenges+and+future+trends%2E+IEEE+Access%2C+2023%2E)
9. (b) (d) Jayashree Subramanian, Varun Sridharan, Kai Shu, and Huan Liu. Exploiting emojis for sarcasm detection. In Social, Cultural, and Behavioral Modeling: 12th International Conference, SBP-BRiMS 2019, Washington, DC, USA, July 9–12, 2019, Proceedings 12, pages 70–80. [find](https://scholar.google.com/scholar?q=%28b%29+%28d%29+Jayashree+Subramanian%2C+Varun+Sridharan%2C+Kai+Shu%2C+and+Huan+Liu%2E+Exploiting+emojis+for+sarcasm+detection%2E+In+Social%2C+Cultural%2C+and+Behavioral+Modeling%3A+12th+International+Conference%2C+SBP%2DBRiMS+2)
10. Alaa E Abdel-Hakim and Wael A Deabes. Impact of sensor data glut on activity recognition in smart environments. In 2017 IEEE 17th International Conference on Ubiquitous Wireless Broadband (ICUWB), pages 1–5. [find](https://scholar.google.com/scholar?q=Alaa+E+Abdel%2DHakim+and+Wael+A+Deabes%2E+Impact+of+sensor+data+glut+on+activity+recognition+in+smart+environments%2E+In+2017+IEEE+17th+International+Conference+on+Ubiquitous+Wireless+Broadband+%28ICUWB%29%2C+pag)
11. Mohamed Tahar Ben Othman, Mohammed Abdullah Al-Hagery, and Yahya Muhammad El Hashemi. Arabic text processing model: Verbs roots and conjugation automation. IEEE Access, 8:103913–103923, 2020. [find](https://scholar.google.com/scholar?q=Mohamed+Tahar+Ben+Othman%2C+Mohammed+Abdullah+Al%2DHagery%2C+and+Yahya+Muhammad+El+Hashemi%2E+Arabic+text+processing+model%3A+Verbs+roots+and+conjugation+automation%2E+IEEE+Access%2C+8%3A103913%E2%80%93103923%2C+2020%2E)
12. Andrew Biemiller and Naomi Slonim. Estimating root word vocabulary growth in normative and advantaged populations: Evidence for a common sequence of vocabulary acquisition. Journal of educational psychology, 93(3):498, 2001. [find](https://scholar.google.com/scholar?q=Andrew+Biemiller+and+Naomi+Slonim%2E+Estimating+root+word+vocabulary+growth+in+normative+and+advantaged+populations%3A+Evidence+for+a+common+sequence+of+vocabulary+acquisition%2E+Journal+of+educational+psyc)
13. George Zinovei Patrick. Roots of the russian language. (No Title), 1989. [find](https://scholar.google.com/scholar?q=George+Zinovei+Patrick%2E+Roots+of+the+russian+language%2E+%28No+Title%29%2C+1989%2E)
14. Moumen El-Melegy, Asmaa Abdelbaset, Alaa Abdel-Hakim, and Gamal El-Sayed. Recognition of arabic handwritten literal amounts using deep convolutional neural networks. In Pattern Recognition and Image Analysis: 9th Iberian Conference, IbPRIA 2019, Madrid, Spain, July 1–4, 2019, Proceedings, Part II 9, pages 169–176. [find](https://scholar.google.com/scholar?q=Moumen+El%2DMelegy%2C+Asmaa+Abdelbaset%2C+Alaa+Abdel%2DHakim%2C+and+Gamal+El%2DSayed%2E+Recognition+of+arabic+handwritten+literal+amounts+using+deep+convolutional+neural+networks%2E+In+Pattern+Recognition+and+Image+A)
15. Wissam Antoun, Fady Baly, and Hazem Hajj. Arabert: Transformer-based model for arabic language understanding. arXiv preprint arXiv:2003.00104, 2020. [↗](https://arxiv.org/abs/2003.00104) [our card](https://md.research.elamur.ai/free/en/papers/antoun-2020-arabert-transformer-based-model-arabic/)
16. Maha Jarallah Althobaiti. Bert-based approach to arabic hate speech and offensive language detection in twitter: Exploiting emojis and sentiment analysis. International Journal of Advanced Computer Science and Applications, 13(5), 2022. [find](https://scholar.google.com/scholar?q=Maha+Jarallah+Althobaiti%2E+Bert%2Dbased+approach+to+arabic+hate+speech+and+offensive+language+detection+in+twitter%3A+Exploiting+emojis+and+sentiment+analysis%2E+International+Journal+of+Advanced+Computer+Sc)
17. Hamdy Mubarak, Sabit Hassan, and Shammur Absar Chowdhury. Emojis as anchors to detect arabic offensive language and hate speech. Natural Language Engineering, 29(6):1436–1457, 2023. [find](https://scholar.google.com/scholar?q=Hamdy+Mubarak%2C+Sabit+Hassan%2C+and+Shammur+Absar+Chowdhury%2E+Emojis+as+anchors+to+detect+arabic+offensive+language+and+hate+speech%2E+Natural+Language+Engineering%2C+29%286%29%3A1436%E2%80%931457%2C+2023%2E) [our card](https://md.research.elamur.ai/free/en/papers/mubarak-2022-emojis-anchors-detect-arabic-offensive/)
18. Fatemah Husain and Ozlem Uzuner. Investigating the effect of preprocessing arabic text on offensive language and hate speech detection. Transactions on Asian and Low-Resource Language Information Processing, 21(4):1–20, 2022. [find](https://scholar.google.com/scholar?q=Fatemah+Husain+and+Ozlem+Uzuner%2E+Investigating+the+effect+of+preprocessing+arabic+text+on+offensive+language+and+hate+speech+detection%2E+Transactions+on+Asian+and+Low%2DResource+Language+Information+Proc)
19. Saja Al-Dabet, Ahmed ElMassry, Ban Alomar, and Abdullah Alshamsi. Transformer-based arabic offensive speech detection. In 2023 International Conference on Emerging Smart Computing and Informatics (ESCI), pages 1–6. [find](https://scholar.google.com/scholar?q=Saja+Al%2DDabet%2C+Ahmed+ElMassry%2C+Ban+Alomar%2C+and+Abdullah+Alshamsi%2E+Transformer%2Dbased+arabic+offensive+speech+detection%2E+In+2023+International+Conference+on+Emerging+Smart+Computing+and+Informatics+%28ESC)
20. Muhammad Okky Ibrohim, Muhammad Akbar Setiadi, and Indra Budi. Identification of hate speech and abusive language on indonesian twitter using the word2vec, part of speech and emoji features. In Proceedings of the 1st International Conference on Advanced Information Science and System, pages 1–5, 2019. [find](https://scholar.google.com/scholar?q=Muhammad+Okky+Ibrohim%2C+Muhammad+Akbar+Setiadi%2C+and+Indra+Budi%2E+Identification+of+hate+speech+and+abusive+language+on+indonesian+twitter+using+the+word2vec%2C+part+of+speech+and+emoji+features%2E+In+Procee)
21. Sadam Al-Azani and El-Sayed M El-Alfy. Combining emojis with arabic textual features for sentiment classification. In 2018 9th International Conference on Information and Communication Systems (ICICS), pages 139–144. [find](https://scholar.google.com/scholar?q=Sadam+Al%2DAzani+and+El%2DSayed+M+El%2DAlfy%2E+Combining+emojis+with+arabic+textual+features+for+sentiment+classification%2E+In+2018+9th+International+Conference+on+Information+and+Communication+Systems+%28ICICS%29)
22. Hanane Elfaik et al. Combining context-aware embeddings and an attentional deep learning model for arabic affect analysis on twitter. IEEE Access, 9:111214–111230, 2021. [find](https://scholar.google.com/scholar?q=Hanane+Elfaik+et+al%2E+Combining+context%2Daware+embeddings+and+an+attentional+deep+learning+model+for+arabic+affect+analysis+on+twitter%2E+IEEE+Access%2C+9%3A111214%E2%80%93111230%2C+2021%2E)
23. Mohammad Ali Humayun, Hayati Yassin, and Pg Emeroylariffion Abas. Dialect classification using acoustic and linguistic features in arabic speech. IAES International Journal of Artificial Intelligence, 12(2):739, 2023. [find](https://scholar.google.com/scholar?q=Mohammad+Ali+Humayun%2C+Hayati+Yassin%2C+and+Pg+Emeroylariffion+Abas%2E+Dialect+classification+using+acoustic+and+linguistic+features+in+arabic+speech%2E+IAES+International+Journal+of+Artificial+Intelligence%2C)
24. Ali Saleh Alammary. Bert models for arabic text classification: a systematic review. Applied Sciences, 12(11):5720, 2022. [find](https://scholar.google.com/scholar?q=Ali+Saleh+Alammary%2E+Bert+models+for+arabic+text+classification%3A+a+systematic+review%2E+Applied+Sciences%2C+12%2811%29%3A5720%2C+2022%2E)
25. Marcos Zampieri, Preslav Nakov, Sara Rosenthal, Pepa Atanasova, Georgi Karadzhov, Hamdy Mubarak, Leon Derczynski, Zeses Pitenis, and Çağrı Çöltekin. Semeval-2020 task 12: Multilingual offensive language identification in social media (offenseval 2020). arXiv preprint arXiv:2006.07235, 2020. [↗](https://arxiv.org/abs/2006.07235)
26. Azalden Alakrot, Liam Murray, and Nikola S Nikolov. Dataset construction for the detection of anti-social behaviour in online communication in arabic. Procedia Computer Science, 142:174–181, 2018. [find](https://scholar.google.com/scholar?q=Azalden+Alakrot%2C+Liam+Murray%2C+and+Nikola+S+Nikolov%2E+Dataset+construction+for+the+detection+of+anti%2Dsocial+behaviour+in+online+communication+in+arabic%2E+Procedia+Computer+Science%2C+142%3A174%E2%80%93181%2C+2018%2E)
27. Thuat Nguyen, Chien Van Nguyen, Viet Dac Lai, Hieu Man, Nghia Trung Ngo, Franck Dernoncourt, Ryan A Rossi, and Thien Huu Nguyen. Culturax: A cleaned, enormous, and multilingual dataset for large language models in 167 languages. arXiv preprint arXiv:2309.09400, 2023. [↗](https://arxiv.org/abs/2309.09400) [our card](https://md.research.elamur.ai/free/en/papers/nguyen-2023-culturax-cleaned-enormous-multilingual-dataset/)