Resolving the Paradoxes of Cross-lingual Transfer in Multilingual Language Models

The technical advances, and resulting societal opportunities, of Large Language Models (LLMs) have principally benefited communities whose primary languages are well-represented in the written data used for training LLMs (e.g., English). While these few high-resource languages are used by many around the world, they do not cover large segments of the global population of 8.2 billion, who collectively speak over 7000 languages. For intelligent natural language systems to be adopted and useful, they must enable interaction in the preferred languages of their users and be knowledgeable of the environments of those users. This expansion of LLM functionality requires re-thinking the cross-lingual transfer paradigm for enabling systems in low-resource languages. In an era where LLMs are knowledge bases, naive reasoners, and interactive agents, our intuitions that held for cross-lingual transfer to linguistic tasks will not extend to transferring regional and cultural knowledge understanding, which may differ even among similar languages.In this proposal, we reformulate cross-lingual transfer using inference-time algorithms that dynamically localize, augment, and adapt implicit language and knowledge representations of multilingual LLMs for queries presented in any language. These new algorithms will leverage shared linguistic knowledge for cross-lingual transfer to new languages while disentangling regional and cultural knowledge that is tied to language but unique to individual language environments. Second, we will develop novel modular architectures to catalyze our adaptation algorithms by disentangling language and knowledge representations within multilingual LLMs during pretraining. Finally, we will develop new benchmarks, settings, and standards for reliable evaluation of regional knowledge in multilingual contexts.