Abstract

AI based on Large language models (LLMs) is often presented as technology that serves and benefits everyone. In practice, however, the opportunities to benefit from it (and, of course, to build it) are far from equitably distributed. In this talk, I first visit the many faces of linguistic exclusion in language-based AI: from "token tax" for speakers of non-major languages and models' Anglo-centric cultural biases to the fact that LLMs truly support well under 1% of the world's 7,000+ languages. I then zoom in and ask whether the only remedy is more data and more compute, which are exactly the resources that excluded societies lack. Finally, I'll present a set of resource-lean (read: "cheap") strategies that improve multilingual language technology across three fronts: (1) text-based LLMs, (2) vision-language models, and (3) code language models. The takeaway? Scaling isn't the only answer. Truly inclusive language technology needs stronger inductive biases and more conceptual innovation, not just bigger models.

Biography

Goran Glavaš is a Full Professor for Natural Language Processing at the University of Würzburg (Germany), Center for AI and Data Science (CAIDAS). He obtained his Ph.D. in 2014 at the Text Analysis and Knowledge Engineering Lab (TakeLab), Faculty of Electrical Engineering and Computing, University of Zagreb. His research interests are in the areas of Natural Language Processing (NLP) and Information Retrieval (IR), with focus on multilinguality and cross-lingual transfer, trustworthy and reliable language-based AI, and NLP applications (primarily in social sciences and humanities). He has (co-)authored over 150 publications in the areas of NLP and Information Retrieval (IR), publishing regularly at top-tier venues (e.g., ACL, EMNLP, SIGIR). He was an Editor-in-Chief of the ACL Rolling Review (ARR), a central reviewing service of the Association for Computational Linguistics. He is a Scholar of the NLP Program in ELLIS, an elite pan-European AI network of excellence, and a member of the Association for Computational Linguistics and German Society for Computational Linguistics (GSCL).