Toward Safe and Interpretable Adaptation of Multilingual and Multimodal AI

Abstract: Despite advances in multilingual and multimodal language models, two challenges remain prominent as these models grow larger and more capable: they are biased toward high-resource languages, and their internal adaptation mechanisms are poorly understood. In this talk, I will present two studies addressing these challenges. First, I will demonstrate how targeted adaptation can significantly improve downstream task performance on underrepresented African languages. To achieve this, we performed multilingual adaptive fine-tuning on a carefully curated corpus of African languages, resulting in AfroXLM-R, a transformer model that better represents these languages. Second, I will show how we analyzed the hidden space of a common method for extending language models to new languages, transformer language adapters, to understand their effect in language model adaptation. By analyzing adapter representations, we can see how they steer language model predictions from the source language toward the target language. Together, these studies combine model development with scientific investigation to improve and better understand NLP models. Finally, I will discuss my future research agenda, which focuses on building multilingual and multimodal AI systems that are adaptable and safe.