Great work by Andreea Iana, well deserved!

Great work by Andreea Iana to provide the first truly multilingual news recommendation dataset to date, enabling research on multilingual and cross-lingual news recommendation!

Happy for the opportunity to talk about xMIND, our new #multilingual #news #dataset , joint work w/ Goran Glavaš & Heiko Paulheim, at the Academic Speed Dating event organized by the Mannheim Center for Data Science!

Check out xMIND at: github.com/andreeaiana/xM…

Pretty neat but would be great if they compare with (and cite) mBLIP, an open multilingual LVLM predating this by half a year arxiv.org/abs/2307.06930.
Also maybe evaluate on established benchmarks like IGLUE, MAXM, or XM3600 to compare with prior work?

📢New Preprint
🤔️Balancing the tradeoff between in-task performance and model calibration: thrilled to share our findings with self-ensembled in-context learning which enhances⬆️ both aspects! 🎉

📄Paper: arxiv.org/abs/2312.13772
📎Code: github.com/cambridgeltl/e…

With LLMs, it's easier than ever to do solid NLP for standard languages, BUT can your LLM *reason* in micro- and nano-dialects? Check our DIALECT-COPA shared task for more details. w. Nikola Ljubešić and Ivan Vulić

For research works in #multilingual news #recommendersystems , we are looking for people with skills in the following languages: Guarani, Haitian Creole, Quechua, Somali, Thai, Indonesian, Swahili, Tamil, Georgian, Vietnamese, Japanese. Get in touch with Andreea Iana. Please RT!

Llava 1.5 struggles with other languages. That's why we need more multilingual multimodal models like our mBLIP (arxiv.org/abs/2307.06930)

Also why F1 and chrF++ and not the established accuracy and CIDEr? Just makes comparison harder.

Great effort by Andreea Iana! An amazing resource & codebase for anyone working on news recommendation. Easy to train SotA models and comparatively evaluate and ablate them (true apple-to-apple comparisons)!

We (WüNLP Universität Würzburg #UniWürzburg ) are hiring! Looking for a postdoc to work on LLM alignment.

More info at tinyurl.com/5n7xy9c9 and tinyurl.com/2murbute

Come join our young and energetic team in Würzburg, one of the prettiest and most livable cities in Germany!

Writing the EMNLP rebuttals. I'm now convinced (also after having served for a year as EiC for ACLRollingReview) that nothing short of publicly releasing reviews *with reviewer identities* will substantially improve the (currently appaling) average review quality in #NLProc .

'We thank the reviewers for their insightful comments'. Happy EMNLP rebuttal, everyone!😊 #EMNLP2023

I've developed a handy way of streamlining our EMNLP 2024 reviewer criticisms on social media

Great work by Gregor Geigle! SoTA massively multilingual Vision-and-Language models obtained with few days of training on consumer-grade GPUs.

Thanks to Gregor Geigle, you can now evaluate exactly how multilingual your favorite 'multilingual' V&L model truly is.

Happy to share that I presented our work about Demographic Adaptation at Insights Workshop today at EACL2023.
🙌Thanks to my co-authors:
Anne Lauscher (she/her) Dirk Hovy Goran Glavaš @spponzius
📖Check here: aclanthology.org/2023.findings-…
(1/2)

Gladly sharing that our work with Goran Glavaš and Ivan Vulić is accepted to #acl2023 , takeaways: (1) averaging training checkpoints outperforms conventional model selection for cross-lingual transfer; (2) averaging checkpoints from multiple runs brings further gains if heads are aligned

Come join us at CAIDAS Universität Würzburg #UniWürzburg!

Still training #news #recommendersystems with explicit user models? Check our SIGIR 2024 paper 'Simplifying Content-Based Neural News Recommendation: On User Modeling and Training Objectives' to learn about alternatives! arxiv.org/abs/2304.03112 Andreea Iana Goran Glavaš Data and Web Science Group

Ajmo, MUP-RH, radite svoj posao. Kein platz für Nazis!