If you are at #EMNLP2023, go say hi to Yang Chen and check out his work on Information Seeking VQA. 👋

Just saw this update to the ACL preprint survey, erasing a lot of the "global north vs global south" difference for the first proposed change (allowing papers to shared online at any time). Original in blue, updated in red.

EasyProject inserts markup into training datasets before translation to low-resource languages. However, inserting markup degrades translation quality. This work addresses the issue by constraining the MT decoder to only insert markup #ICLR2024

How well do language models understand neologisms? As it turns out, not very well - for example, newly coined words result in 43% lower translation quality across a range of models including GPT-4, ALMA 7B, and M2M. See our paper for more analysis: arxiv.org/pdf/2402.12261…

Interesting slide from Noah A. Smith's talk today summarizing what we know about various LLMs. Lots of question marks...

Remember when lexicons were super important to get good performance in NLP? We wondered if they can still be useful today. Turns out that lexicons are beneficial when prompting in a zero-shot setup. See Ruohao Guo's paper (accepted at ACL) for details: arxiv.org/abs/2305.14592

🚀 Excited about this new work led by Yang Chen & Vedaant Shah! Demonstrates how efforts such as NLLB that enhance MT in low-resource languages can also benefit zero-shot information extraction. 📖 Read more: arxiv.org/abs/2305.13582 Machine Learning at Georgia Tech

Congratulations to Yang Chen for successfully defending his PhD! Yang has done exciting work advancing both the multilingual and multimodal capabilities of LLMs. Many thanks to his committee: Wei Xu (co-advisor), Ming-Wei Chang, Hexiang (Frank) Hu, kartik goyal

Check out our recent work that aims to protect privacy by helping people make more informed decisions about what they share online. Drop by our poster tomorrow morning at 11am if you are at ACL in Bankok. ACL 2024 arxiv.org/abs/2311.09538