Elamur Research⌕
☰
home / papers / 2026 / Improving Amharic Information Retrieval with Translative and Multi-Agent Debate Retrieval Augmented Generation

Improving Amharic Information Retrieval with Translative and Multi-Agent Debate Retrieval Augmented Generation

Paper · 2026 · Abel Alemu Jotie and Prasenjit Mitra

Summary

The paper tackles retrieval-augmented generation (RAG) for Amharic, a widely spoken Ethiopian language with few resources. Weak embeddings, scarce digital text, rich morphology, the Ge'ez script and culture-specific content such as proverbs and a distinct calendar all reduce retrieval quality. Unreliable retrieval in turn makes the generated answers unreliable.

The authors combine two existing ideas:

  • TraSe-style translative RAG: the Amharic query and the retrieved passages are moved into English, the LLM reasons there, and the answer is translated back to Amharic.
  • Debate-Augmented RAG: several agents propose, criticise and judge candidate contexts and answers, refining them over rounds.

To measure the effect, they build ARGB, which they present as the first comprehensive Amharic RAG benchmark. It uses whole articles as context rather than short snippets. Beyond retrieval and generation quality, it tests four robustness abilities: handling noise, rejecting unanswerable questions, resisting counterfactual information and integrating several sources.

The authors report that both techniques substantially improve several aspects of the Amharic RAG pipeline over baselines. The work offers an evaluation resource and a template that could be reused for other low-resource languages.

Key points

  • New resource: ARGB (Amharic Retrieval-Augmented Generation Benchmark). It is factoid and open-domain, spans many categories, and is designed around Amharic grammar, culture, writing system and calendar constraints. Code and data are said to be on GitHub.
  • Robustness dimensions: the four abilities tested are taken from Chen et al. (2024):
    • noise robustness,
    • negative rejection,
    • information integration,
    • counterfactual robustness.
  • Contrast with prior work: the only earlier public Amharic QA dataset, Amh-QuAD (Taffa et al., 2024, built from Amharic Wikipedia), has contexts of just a few sentences. ARGB instead uses full articles to better reflect a real retrieval space.
  • Origin of TraSe: the architecture comes from Ipa et al. (2025), who built it for Bangla. It uses translative prompting, multiple prompts and an LLM-based selector that picks the best candidate answer.
  • Origin of the debate component: it follows Hu et al. (2025), building on Du et al. (2024). Agents debate both how relevant the retrieved documents are and how faithful the answer is to them.

Interesting ideas

  • I1. Separate the retrieval language from the reasoning language. Keep the documents in the native language, but pivot the query and passages into the language where the model reasons best, then translate the result back. This sidesteps weak native-language reasoning without discarding native sources.
  • I2. Let agents debate the evidence, not only the answer. Agents argue over whether each retrieved context is relevant and whether the answer is faithful to it, which filters misleading inputs before they spread.
  • I3. Test negative rejection and counterfactual robustness explicitly. Include questions the context cannot answer and contexts containing planted false facts, so a system earns credit only for abstaining or noticing the conflict.
  • I4. Use whole articles as context instead of a few sentences. Short snippets make retrieval and synthesis look easier than they really are.
  • I5. Build culture- and script-specific traps into the benchmark. Including items such as a different calendar system exposes failures that generic multilingual test sets miss.

For Mova

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

  • I1 (back): The same split could guide translation through the meta-world: parse the source into the meaning graph, reason only on the graph, and generate the target text from it, so each pivot step can be inspected.
  • I2 (learn): When witness parsers disagree on silver data, a structured critique round could argue per arc or per frame and record the reasons for the decision.
  • I3 (answer): QA on fables and tales could add unanswerable questions and deliberately altered story facts, checking that readers say "not stated" or flag the contradiction instead of guessing.
  • I4 (world): Evaluating world-of-text construction on full documents would stress entity tracking, event ordering and coreference over long spans.
  • I5 (world): The measures and values knowledge layer could include calendar and date-system conversions, tested with dedicated items so normalisation errors show up.

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

🎓Carnegie Mellon UniversityPittsburgh · United States539 works · 179,533 citations🎓Carnegie Mellon University AfricaKigali · Rwanda2 works · 1 citation

 Fields, topics and concepts

📚 Shelves 5The 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 worksInformation Systems · 1,437 works

🧭 Topics 2Research subjects and languages the work is about.Semitic languages · 2,908 worksInformation retrieval & search · 473 works

Cite

Cite as Abel Alemu Jotie and Prasenjit Mitra. 2026. Improving Amharic Information Retrieval with Translative and Multi-Agent Debate Retrieval Augmented Generation. Proceedings of Resources for African Indigenous Languages (RAIL) 2026 @ LREC 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/3hbm9z4okcrx

@inproceedings{jotie2026improving,
  title = {{Improving Amharic Information Retrieval with Translative and Multi-Agent Debate Retrieval Augmented Generation}},
  author = {Abel Alemu Jotie and Prasenjit Mitra},
  booktitle = {Proceedings of Resources for African Indigenous Languages (RAIL) 2026 @ LREC 2026},
  year = {2026},
  doi = {10.63317/3hbm9z4okcrx},
  url = {https://doi.org/10.63317/3hbm9z4okcrx}
}

References

26 references, 11 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. 2025. rasyosef/Llama-3.2-180M-Amharic · Hugging Face. find
  2. David Adelani. 2021. xlm-roberta-base-finetuned-amharic. https://huggingface.co/Davlan/ xlm-roberta-base-finetuned-amharic. Hugging Face model card; fine-tuned XLM-RoBERTa for Amharic masked language tasks. ↗
  3. Mengistu Amberber. 2023. Amharic. In Ronny Meyer, Bedilu Wakjira, and Zelealem Leyew, editors, The Oxford Handbook of Ethiopian Languages, pages 414–442. Oxford University Press. find
  4. Israel Abebe Azime, Atnafu Lambebo Tonja, Tadesse Destaw Belay, Mitiku Yohannes Fuge, Aman Kassahun Wassie, Eyasu Shiferaw Jada, Yonas Chanie, Walelign Tewabe Sewunetie, and Seid Muhie Yimam. 2024. Walia-LLM: Enhancing Amharic-LLaMA by integrating task-specific and generative datasets. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 432–444, Miami, Florida, USA. Association for Computational Linguistics. ↗
  5. World Bank. 2023. Digital‑in‑health: Unlocking the value for everyone. Technical report, World Bank. ↗
  6. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language Models are Few-Shot Learners. In Advances in Neural Information Processing Systems, volume 33, pages 1877–1901. Curran Associates, Inc.
  7. Andreas Chari, Sean MacAvaney, and Iadh Ounis. 2025. Improving low‑resource retrieval effectiveness using zero‑shot linguistic similarity transfer. In Proceedings of the 47th European Conference on Information Retrieval (ECIR 2025). find
  8. Jiawei Chen, Hongyu Lin, Xianpei Han, and Le Sun. 2024. Benchmarking large language models in retrieval-augmented generation. In Proceedings of the Thirty-Eighth AAAI Conference on Artificial Intelligence and Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence and Fourteenth Symposium on Educational Advances in Artificial Intelligence, AAAI’24/IAAI’24/EAAI’24. AAAI Press.
  9. Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mordatch. 2024. Improving factuality and reasoning in language models through multiagent debate. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of ICML’24, pages 11733–11763, Vienna, Austria. JMLR.org. ↗
  10. Marwa Ibrahim Eid. 2021. The impact of ethiopic (ge’ez) literature on the emergence and the flourish of amharic literature. iKNiTO Journal Management System. find
  11. Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. 2024. Detecting hallucinations in large language models using semantic entropy. Nature, 630(8017):625–630. ↗
  12. Tianyu Gao, Howard Yen, Jiatong Yu, and Danqi Chen. 2023. Enabling Large Language Models to Generate Text with Citations. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6465– 6488, Singapore. Association for Computational Linguistics.
  13. Wentao Hu, Wengyu Zhang, Yiyang Jiang, Chen Jason Zhang, Xiaoyong Wei, and Li Qing. 2025. Removal of hallucination on hallucination: Debate-augmented RAG. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15839–15853, Vienna, Austria. Association for Computational Linguistics. ↗
  14. Atia Shahnaz Ipa, Mohammad Abu Tareq Rony, and Mohammad Shariful Islam. 2025. Empowering low-resource languages: TraSe architecture for enhanced retrieval-augmented generation in Bangla. In Proceedings of the 1st Workshop on Language Models for Underserved Communities (LM4UC 2025), pages 8–15, Albuquerque, New Mexico. Association for Computational Linguistics. ↗
  15. Zhengbao Jiang, Frank Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi-Yu, Yiming Yang, Jamie Callan, and Graham Neubig. 2023. Active retrieval augmented generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7969– 7992, Singapore. Association for Computational Linguistics. ↗
  16. Samreen Kazi, Shakeel Khoja, and Ali Daud. 2025. Bridging the gap: A survey of document retrieval techniques for high-resource and low-resource languages. Computer Science Review, 57:100756. ↗
  17. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. In Advances in Neural Information Processing Systems, volume 33, pages 9459–9474. Curran Associates, Inc.
  18. Zichao Li and Zong Ke. 2025. Cross-modal augmentation for low-resource language understanding and generation. In Proceedings of the 1st Workshop on Multimodal Augmented Generation via Multimodal Retrieval (MAGMaR 2025), pages 90–99, Vienna, Austria. Association for Computational Linguistics. find
  19. Tsegahun Manyazewal, Yimtubezinash Woldea-manuel, Henry M. Blumberg, Abebaw Fekadu, and Vincent C. Marconi. 2021. The potential use of digital health technologies in the african context: a systematic review of evidence from ethiopia. npj Digital Medicine, 4:125. find
  20. Melakneh Mengistu. 2018. Cross-cultural wisdom in english and amharic proverbs. Ethiopian Journal of Languages and Literature, 14:—. find
  21. Zhongtao Miao, Qiyu Wu, Kaiyan Zhao, Zilong Wu, and Yoshimasa Tsuruoka. 2024. Enhancing cross-lingual sentence embedding for low-resource languages with word alignment. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 3225–3236, Mexico City, Mexico. Association for Computational Linguistics. find
  22. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. OpenAI. Accessed: 2024-11-15. find
  23. Leonardo Ranaldi, Barry Haddow, and Alexandra Birch. 2025. Multilingual retrieval-augmented generation for knowledge-intensive task.
  24. Kurt Shuster, Spencer Poff, Moya Chen, Douwe Kiela, and Jason Weston. 2021. Retrieval Augmentation Reduces Hallucination in Conversation. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 3784– 3803, Punta Cana, Dominican Republic. Association for Computational Linguistics. ↗
  25. Tilahun Abedissa Taffa, Ricardo Usbeck, and Yaregal Assabie. 2024. Low resource question answering: An Amharic benchmarking dataset. In Proceedings of the Fifth Workshop on Resources for African Indigenous Languages @ LREC-COLING 2024, pages 124–132, Torino, Italia. ELRA and ICCL.
  26. Adam Wiemerslage, Miikka Silfverberg, Changbing Yang, Arya McCarthy, Garrett Nicolai, Eliana Colunga, and Katharina Kann. 2022. Morphological processing of low-resource languages: Where we are and what’s next. In Findings of the Association for Computational Linguistics: ACL 2022, pages 988–1007, Dublin, Ireland. Association for Computational Linguistics. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/jotie-2026-improving-amharic-information-retrieval-translative/