---
title: "Hate Speech Detection and Classification in — Gashe 2024"
description: "Amharic is Ethiopia's long-standing lingua franca and is increasingly used online, but it has very little tooling for detecting hateful posts."
authors: ["Samuel Minale Gashe", "Seid Muhie Yimam", "Yaregal Assabie"]
year: 2024
venue: "arXiv:2408.03849"
source: "https://arxiv.org/abs/2408.03849"
doi: "10.48550/arxiv.2408.03849"
arxiv: "2408.03849"
licence: "CC BY-NC-ND 4.0 (arXiv)"
licence_url: "http://creativecommons.org/licenses/by-nc-nd/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/gashe-2024-hate-speech-detection-classification-amharic/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/gashe-2024-hate-speech-detection-classification-amharic/)

[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/) / Hate Speech Detection and Classification in Amharic Text with Deep Learning

# Hate Speech Detection and Classification in Amharic Text with Deep Learning

**Authors:** Samuel Minale Gashe, [Seid Muhie Yimam](https://md.research.elamur.ai/free/en/authors/seid-muhie-yimam/), [Yaregal Assabie](https://md.research.elamur.ai/free/en/authors/yaregal-assabie/) · **Year:** 2024 · **Venue:** arXiv:2408.03849 · **Licence:** [CC BY-NC-ND 4.0 (arXiv)](http://creativecommons.org/licenses/by-nc-nd/4.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/2408.03849) · [PDF](https://arxiv.org/pdf/2408.03849) · [DOI 10.48550/arxiv.2408.03849](https://doi.org/10.48550/arxiv.2408.03849) · [arXiv 2408.03849](https://arxiv.org/abs/2408.03849) · [OpenAlex](https://openalex.org/W4403965720)

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

## Summary
Amharic is Ethiopia's long-standing lingua franca and is increasingly used online, but it has very little tooling for detecting hateful posts. Such content can inflame ethnic and religious tensions there. Earlier Amharic studies did not separate types of hate and did not normalize their text, and the authors connect both gaps to false positives. This short paper contributes a new labeled dataset and a classifier. The authors gathered over a million posts and comments from Twitter, Facebook and YouTube (2014–2022). They cleaned the collection with the PYCLD2 language identifier and used keyword filtering to reduce it to 5k items. One hundred native speakers then labeled these items as racial, religious, gender, or non-hate with a purpose-built annotation tool. The pipeline normalizes and tokenizes the text, rebalances classes with SMOTE, builds features with fastText and TF-IDF, and trains a stacked bidirectional LSTM (SBi-LSTM). The model reaches an F1 of 94.8, far above rule-based (40.1) and classical ML (80.3) baselines. This is an early multi-class Amharic hate speech resource. However, the experimental details are thin, and the data release was only promised upon acceptance.

## Key points
- The raw pool of over 1 million social media items was reduced to 5k annotated posts and comments in four classes: racial, religious, gender, and non-hate.
- 100 native Amharic annotators with varied demographic backgrounds did the labeling in a custom "Amharic Hate Speech Annotation Tool".
- Preprocessing combined PYCLD2 language filtering, normalization and tokenization. SMOTE was used for class balancing, and fastText and TF-IDF provided the features.
- On the same data, the SBi-LSTM reached F1 94.8, compared with 80.3 for classical ML and 40.1 for rule-based methods.
- Proposed future work includes identifying the targeted ethnicity or religion, adding more categories, and integrating the system with platforms.

## Interesting ideas
- The authors build a rule-based, a classical ML and a deep model on one shared dataset, so the gain from each added layer of complexity is directly visible. — For Mova (learn): Score self-induced repair rules, the perceptron parser and gparse on the same held-out split as a tiered ladder, so each component's contribution is measurable rather than assumed.
- They run a language identifier before annotation so that off-language or mixed noise never reaches the costly human labeling step. — For Mova (learn): Gate silver-data candidates through language ID before the witness parsers vote, so their agreement is not spent on non-English fragments.
- They use keyword-targeted sampling to pull rare categories out of a huge, mostly irrelevant pool before labeling, which concentrates annotation effort where the signal is. — For Mova (frames): Mine candidate sentences for rare pragmatic phenomena such as irony or indirect speech acts with cue lists before sending them to LLM teacher tiers.
- The authors attribute earlier misclassifications partly to missing normalization, so canonicalizing spelling variants becomes a first-class step rather than an afterthought. — For Mova (tokens): Add an explicit, logged normalization pass for informal spellings ahead of the tagger and lemmatizer, keeping each rewrite traceable for explanations.
- For subjective labels, they recruit a large annotator pool from varied backgrounds so that one group's perception does not define the categories. — For Mova (frames): When collecting gold judgments for irony or implicature, draw from diverse annotators and treat their disagreement as a signal of genuinely ambiguous readings.

*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{gashe2024hate,
  title = {{Hate Speech Detection and Classification in Amharic Text with Deep Learning}},
  author = {Gashe, Samuel Minale and Yimam, Seid Muhie and Assabie, Yaregal},
  year = {2024},
  doi = {10.48550/arxiv.2408.03849},
  eprint = {2408.03849},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2408.03849}
}
```

- [← previous HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal](https://md.research.elamur.ai/free/en/papers/mazeika-2024-harmbench-standardized-evaluation-framework-automated/)
- [next → HATFormer: Historic Handwritten Arabic Text Recognition with Transformers](https://md.research.elamur.ai/free/en/papers/chan-2024-hatformer-historic-handwritten-arabic-text/)
- [↑ back to papers of 2024](https://md.research.elamur.ai/free/en/papers/2024/page/3/)

## References

21 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Mesfin Abate and Yaregal Assabie. 2014. Development of amharic morphological analyzer using memory-based learning. In Advances in Natural Language Processing: 9th International Conference on NLP, PolTAL 2014, Warsaw, Poland, September 17-19, 2014. Proceedings 9, pages 1–13. Springer. [find](https://scholar.google.com/scholar?q=Mesfin+Abate+and+Yaregal+Assabie%2E+2014%2E+Development+of+amharic+morphological+analyzer+using+memory%2Dbased+learning%2E+In+Advances+in+Natural+Language+Processing%3A+9th+International+Conference+on+NLP%2C+PolT)
2. Gabe Adugna. 2020. Language learning – amharic. Technical report, Boston University Libraries. [find](https://scholar.google.com/scholar?q=Gabe+Adugna%2E+2020%2E+Language+learning+%E2%80%93+amharic%2E+Technical+report%2C+Boston+University+Libraries%2E)
3. Nuha Albadi, Maram Kurdi, and Shivakant Mishra. 2018. Are they our brothers? analysis and detection of religious hate speech in the arabic twittersphere. In 2018 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), pages 69–76. IEEE. [find](https://scholar.google.com/scholar?q=Nuha+Albadi%2C+Maram+Kurdi%2C+and+Shivakant+Mishra%2E+2018%2E+Are+they+our+brothers%3F+analysis+and+detection+of+religious+hate+speech+in+the+arabic+twittersphere%2E+In+2018+IEEE%2FACM+International+Conference+on+A)
4. Ika Alfina, Rio Mulia, Mohamad Ivan Fanany, and Yudo Ekanata. 2017. Hate speech detection in the indonesian language: A dataset and preliminary study. In 2017 international conference on advanced computer science and information systems (ICACSIS), pages 233–238. IEEE. [find](https://scholar.google.com/scholar?q=Ika+Alfina%2C+Rio+Mulia%2C+Mohamad+Ivan+Fanany%2C+and+Yudo+Ekanata%2E+2017%2E+Hate+speech+detection+in+the+indonesian+language%3A+A+dataset+and+preliminary+study%2E+In+2017+international+conference+on+advanced+comp)
5. Pinkesh Badjatiya, Shashank Gupta, Manish Gupta, and Vasudeva Varma. 2017. Deep learning for hate speech detection in tweets. In Proceedings of the 26th international conference on World Wide Web companion, pages 759–760. [find](https://scholar.google.com/scholar?q=Pinkesh+Badjatiya%2C+Shashank+Gupta%2C+Manish+Gupta%2C+and+Vasudeva+Varma%2E+2017%2E+Deep+learning+for+hate+speech+detection+in+tweets%2E+In+Proceedings+of+the+26th+international+conference+on+World+Wide+Web+comp)
6. CARD. 2020. Anti-Fake News and Hate Speech Guideline, volume 1. Center for Advancement of Rights and Democracy (CARD) Ethiopia, Addis Ababa, Ethiopia. [find](https://scholar.google.com/scholar?q=CARD%2E+2020%2E+Anti%2DFake+News+and+Hate+Speech+Guideline%2C+volume+1%2E+Center+for+Advancement+of+Rights+and+Democracy+%28CARD%29+Ethiopia%2C+Addis+Ababa%2C+Ethiopia%2E)
7. Nitesh V Chawla, Kevin W Bowyer, Lawrence O Hall, and W Philip Kegelmeyer. 2002. Smote: synthetic minority over-sampling technique. Journal of artificial intelligence research, 16:321–357. [find](https://scholar.google.com/scholar?q=Nitesh+V+Chawla%2C+Kevin+W+Bowyer%2C+Lawrence+O+Hall%2C+and+W+Philip+Kegelmeyer%2E+2002%2E+Smote%3A+synthetic+minority+over%2Dsampling+technique%2E+Journal+of+artificial+intelligence+research%2C+16%3A321%E2%80%93357%2E) [our card](https://md.research.elamur.ai/free/en/papers/chawla-2002-smote-synthetic-minority-over-sampling/)
8. Fabio Del Vigna12, Andrea Cimino23, Felice Dell’Orletta, Marinella Petrocchi, and Maurizio Tesconi. 2017. Hate me, hate me not: Hate speech detection on facebook. In Proceedings of the first Italian conference on cybersecurity (ITASEC17), pages 86–95. [find](https://scholar.google.com/scholar?q=Fabio+Del+Vigna12%2C+Andrea+Cimino23%2C+Felice+Dell%E2%80%99Orletta%2C+Marinella+Petrocchi%2C+and+Maurizio+Tesconi%2E+2017%2E+Hate+me%2C+hate+me+not%3A+Hate+speech+detection+on+facebook%2E+In+Proceedings+of+the+first+Italian+c)
9. Hang Thi-Thuy Do, Huy Duc Huynh, Kiet Van Nguyen, Ngan Luu-Thuy Nguyen, and Anh Gia-Tuan Nguyen. 2019. Hate speech detection on vietnamese social media text using the bidirectional-lstm model. arXiv preprint arXiv:1911.03648. [↗](https://arxiv.org/abs/1911.03648)
10. Alex Graves, Navdeep Jaitly, and Abdel-rahman Mohamed. 2013. Hybrid speech recognition with deep bidirectional lstm. In 2013 IEEE workshop on automatic speech recognition and understanding, pages 273–278. IEEE. [find](https://scholar.google.com/scholar?q=Alex+Graves%2C+Navdeep+Jaitly%2C+and+Abdel%2Drahman+Mohamed%2E+2013%2E+Hybrid+speech+recognition+with+deep+bidirectional+lstm%2E+In+2013+IEEE+workshop+on+automatic+speech+recognition+and+understanding%2C+pages+273%E2%80%93)
11. Sylvia Jaki and Tom De Smedt. 2019. Right-wing german hate speech on twitter: Analysis and automatic detection. arXiv preprint arXiv:1910.07518. [↗](https://arxiv.org/abs/1910.07518)
12. Till Jünger, Jakob. 2019. Facepager. an application for automated data retrieval on the web. [https://github.com/strohne/Facepager/](https://github.com/strohne/Facepager/). [Online; accessed 21 Mar 2021]. [↗](https://github.com/strohne/Facepager/)
13. Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. 2015. Deep learning. nature, 521(7553):436–444. [find](https://scholar.google.com/scholar?q=Yann+LeCun%2C+Yoshua+Bengio%2C+and+Geoffrey+Hinton%2E+2015%2E+Deep+learning%2E+nature%2C+521%287553%29%3A436%E2%80%93444%2E)
14. Son T Luu, Hung P Nguyen, Kiet Van Nguyen, and Ngan Luu-Thuy Nguyen. 2020. Comparison between traditional machine learning models and neural network models for vietnamese hate speech detection. In 2020 RIVF International Conference on Computing and Communication Technologies (RIVF), pages 1–6. IEEE. [find](https://scholar.google.com/scholar?q=Son+T+Luu%2C+Hung+P+Nguyen%2C+Kiet+Van+Nguyen%2C+and+Ngan+Luu%2DThuy+Nguyen%2E+2020%2E+Comparison+between+traditional+machine+learning+models+and+neural+network+models+for+vietnamese+hate+speech+detection%2E+In+202)
15. Zewdie Mossie, Jenq-Haur Wang, et al. 2018. Social network hate speech detection for amharic language. Computer Science & Information Technology, pages 41–55. [find](https://scholar.google.com/scholar?q=Zewdie+Mossie%2C+Jenq%2DHaur+Wang%2C+et+al%2E+2018%2E+Social+network+hate+speech+detection+for+amharic+language%2E+Computer+Science+%26+Information+Technology%2C+pages+41%E2%80%9355%2E)
16. Gudbjartur Ingi Sigurbergsson and Leon Derczynski. 2019. Offensive language and hate speech detection for danish. arXiv preprint arXiv:1908.04531. [↗](https://arxiv.org/abs/1908.04531) [our card](https://md.research.elamur.ai/free/en/papers/sigurbergsson-2020-offensive-language-hate-speech-detection/)
17. Stéphan Tulkens, Lisa Hilte, Elise Lodewyckx, Ben Verhoeven, and Walter Daelemans. 2016. A dictionary-based approach to racism detection in dutch social media. arXiv preprint arXiv:1608.08738. [↗](https://arxiv.org/abs/1608.08738)
18. William Warner and Julia Hirschberg. 2012. Detecting hate speech on the world wide web. In Proceedings of the second workshop on language in social media, pages 19–26. [find](https://scholar.google.com/scholar?q=William+Warner+and+Julia+Hirschberg%2E+2012%2E+Detecting+hate+speech+on+the+world+wide+web%2E+In+Proceedings+of+the+second+workshop+on+language+in+social+media%2C+pages+19%E2%80%9326%2E) [our card](https://md.research.elamur.ai/free/en/papers/warner-2012-detecting-hate-speech-world-wide/)
19. Seid Muhie Yimam, Abinew Ali Ayele, and Chris Biemann. 2019. Analysis of the ethiopic twitter dataset for abusive speech in amharic. arXiv preprint arXiv:1912.04419. [↗](https://arxiv.org/abs/1912.04419)
20. Kenenisa Yonas. 2019. Hate speech detection for amharic language on social media using machine learning techniques. Ph.D. thesis, ASTU. [find](https://scholar.google.com/scholar?q=Kenenisa+Yonas%2E+2019%2E+Hate+speech+detection+for+amharic+language+on+social+media+using+machine+learning+techniques%2E+Ph%2ED%2E+thesis%2C+ASTU%2E)
21. Steven Zimmerman, Udo Kruschwitz, and Chris Fox. 2018. Improving hate speech detection with deep learning ensembles. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA). [find](https://scholar.google.com/scholar?q=Steven+Zimmerman%2C+Udo+Kruschwitz%2C+and+Chris+Fox%2E+2018%2E+Improving+hate+speech+detection+with+deep+learning+ensembles%2E+In+Proceedings+of+the+Eleventh+International+Conference+on+Language+Resources+and+E)