Barry Haddow

📣 Excited to share our latest research: "Demystifying Multilingual Chain-of-Thought in Process Reward Modeling" where we explore process reward models beyond English to improve multi-step reasoning in 11 languages! Link: arxiv.org/abs/2502.12663 Code: github.com/weixuan-wang12…

New paper on the HPLT v2 dataset making-of: - pipeline documentation and code - extensive analysis of the quality and characteristics - evaluation of the performance of language models and machine translation systems trained on it 🤓Happy reading! arxiv.org/pdf/2503.10267

We are happy to announce the second release of HPLT bilingual datasets: - 50 English-centric language pairs = 380M parallel sentences (HPLT) 🤩 - 1,275 non-English-centric language pairs = 16.7B parallel sentences (MultiHPLT) 😮 Available at the HPLT dataset catalogue and OPUS.

MT Summit 2025 - deadline extended! The deadline for all papers (technical/user/translator/products/projects) has been extended to February 10th. MT Summit will be in Geneva, June 23--27. mtsummit2025.unige.ch/index.html

EAMT best thesis award - closes on January 31st. Completed an MT-related PhD in 2024? In Europe, Africa or Middle East. Then why not submit your thesis. eamt.org/2024/11/28/the…

🥳 Amazing performance of the #HPLT v2 dataset! HuggingFace multilingual evaluation + HPLT English internal evaluation show that HPLT v2 is one of the best datasets to train LLMs. Downloads and more at either HPLT ➡️ hplt-project.org/hplt-v2-datase… or HF ➡️huggingface.co/datasets/HPLT/…

Very exciting to see the 9B EuroLLM model released - made in Europe and supporting all official EU languages. More and bigger models to come ...

EAMT Best thesis award - now open! Have you defended an MT-related thesis in 2024, in EMEA? Then why not submit to the prestigious EAMT BTA? eamt.org/2024/11/28/the… . Deadline: 2025-01-31

Join us on a new edition of the Winter School! "Pretraining Data Quality 🧐 and Multilingual Evaluation of LLMs👀" 🪂Feb. 3–5, 2025, Norway More info and registration: wiki.nlpl.eu/Community/trai… Jointly organised by @hplt_eu and the Nordic Language Processing Laboratory (NLPL)

The 18th MT marathon will be organized in beautiful Helsinki in the end of August, 2025. We invite you to a week-long gathering of researchers, developers and students with lectures, labs and hacking projects. More information will come - stay tuned!

*Update:* Deadline for EAMT project grants is extended by 1 week - to November 25th. Details here: eamt.org/2024/10/21/eam…

Only 5 days left to apply for EAMT project grants

📢 𝗠𝗧 𝗦𝘂𝗺𝗺𝗶𝘁 𝟮𝟬𝟮𝟱: Calls for Papers, Workshops and Tutorials 𝗮𝗿𝗲 𝗢𝘂𝘁! You'll find all the details on our website mtsummit2025.unige.ch Deadlines: 📆WS&Tutorials = 25 Nov 2024 📆CfP = 27 Jan 2025 #machinetranslation #users #researchers #translators #AI #LLM

Have an MT-related idea and looking for funding? EAMT are offering project grants of up to €10k (main track) or €4k (students). Apply by Nov. 18th. eamt.org/2024/10/21/eam… eamt.org/2024/10/21/eam…

New HPLT data release is out!

Have you recently used COMET for MT evaluation? ☄️ - Did you report the specific model? ≥12% of papers don't! - Did you report the package version? Makes a difference. - `pip install sacrecomet` generates a nice version+model signature. Not too late for WMT/EMNLP camera-ready!

❗Are We Truly Achieving Multilingualism in LLMs or Just Relying on Translation?❗ Need multilingual instruction data and benchmarks? Just translate from English. LLM multilingualism can be easily solved! If you agree, check out our #EMNLP 2024 paper which says this is sub-optimal. arxiv.org/abs/2406.12822 🧵Below

Today we release the first EuroLLM paper and models: EuroLLM-1.7B and EuroLLM-1.7B-Instruct! The EuroLLM project will develop open-weight multilingual LLMs that understand and generate text in all official EU languages. Stay tuned for the bigger and stronger EuroLLMs (9B, 22B)!

We know LLMs are poor at MT in low-resource languages (LRLs): curious how to adapt them to perform better? 🚀 Our new paper explores the interplay between scale (of MT data) and diversity (of tasks/langs) in instruction tuning in determining LLM-MT performance for LRLs💡 arxiv.org/abs/2408.12780