https://t.co/0PqJBIxRbN

Our #WMT2025 paper got accepted 🙌 We release a dataset for health low-res MT... and find that Gemini 2.5 beats NLLB 54B, if given full document context. Seeya in Suzhou!

Did you know? ❌77% of language models on @huggingface are not tagged for any language 📈For 95% of languages, most models are multilingual 🚨88% of models with tags are trained on English In a new blog post, @tylerachang and I dig into these trends and why they matter! 👇

paper: https://t.co/e8mIA9e5EC demo: https://t.co/jlYphZa9h6 Grateful for all the help and advice from my supervisors Hanna Suominen, Nick Thieberger, Trevor Cohn, Katerina Vylomova, and special thanks to Maluk Timor & Lois Hong for the great joint work!

in Vienna for ACL, presenting Tulun, a system for low-resource in-domain translation, using LLMs Working w 2 real use cases: medical translation into Tetun 🇹🇱 & disaster relief speech translation in Bislama 🇻🇺

Our paper on generating bilingual example sentences with LLMs got best paper award @altanlp ! https://t.co/pAULq5bv5f We work with French / Indonesian / Tetun, find that annotators don't agree about what's a "good example", but that LLMs can align with a specific annotator.

Translation is a complex task involving pre-translation research and post-translation stages. Can #LLMs handle this process step-by-step, relying solely on their internal knowledge? ✨We show that decomposing the translation process significantly improves #Gemini translation

Life update: after 10 years in industry, I'm going back to school for a PhD at Uni. of Melbourne! Started last week, lots of work to do which i'm really looking forward to !

Coming to Google Translate: Tetun, Tok Pisin, Balinese, Fijian, Acehnese, and many other languages of SEA & the Pacific! Let's test quality when they release, but potentially a small revolution in MT for our region https://t.co/sp8JMFZ95E

Also, the #EURALI folks are very cool and i'm looking forward to more work with them! https://t.co/eW1hKMs7SQ #nlproc #lreccoling2024

Anyways, goes to show the importance of working with native speakers for low resource NLP work, especially in the LLM era, when benchmarks are less trustworthy than ever!

MTOB is still interesting to evaluate LLMs language logic ability, only we need to be careful about the conclusions wrt language acquisition.

Their paper is frankly better, but our findings leave me wondering what results they'd get if producing a separate test set from a native Kalamang speaker.

In particular, our experiment is similar to Tanzer et al, working with Kalamang from Papua, and introducing the MTOB (Machine Translation from One Book) benchmark :

It works when using a test set from the same material as source, and fails miserably when using a totally different test set, which we created with the help of a native Mambai speaker.

First paper published! We create a first corpus for the Mambai language (from Timor-Leste), and try teach an LLM to translate into Mambai using examples selected to match the source, all from one language manual. https://t.co/jE6ReT052P

High internet use but low social media use, hats off to 🇩🇪! Survey by @pewresearch https://t.co/0ckNpeEiJs

This #InternationalWomensDay we pay tribute to the wonder women of Catalpa, who are not just smart and skilled but also fearless, funny and feisty! We spoke to 7 women about their jobs, their career paths and their wishes for women and girls on this day. https://t.co/zW6QYOf5zY

🇮🇩 ranks very high on economic optimism. Hope it stays this way #pemilu2024

Woke up watching this series of short videos on Papuan languages, very nice: https://t.co/GpkxaMZ1BD Lots of trivia, like the Bukiyip system having two counting systems, one in base 3 (for coconuts and fish), one in base 4 (for betel nuts and bananas)