Associate Professor, Dr. Mieradilijiang Maimaiti,
School of Computer Science, Xinjiang University, China
LARGE LANGUAGE MODELS FOR UNDER-RESOURCED LANGUAGES: FROM TEXT UNDERSTANDING TO MULTIMODAL RETRIEVAL
Large language models (LLMs) have demonstrated remarkable capabilities across a wide range of natural language processing and multimodal tasks. However, their performance remains limited in under-resourced language scenarios due to insufficient training data, weak linguistic coverage, and a lack of multimodal supervision. This talk presents our recent efforts to advance LLMs for under-resourced languages in both text-only and multimodal settings. The presentation begins with a brief overview of the foundations of LLMs, examining their applications, challenges, and limitations in under-resourced NLP tasks. Subsequently, we explore recent research on LLM-based machine translation, covering instruction data construction, prior knowledge integration, chain-of-thought prompting, and linguistically aware translation strategies for under-resourced and morphologically rich languages. The focus then shifts to our recent progress in multimodal learning, highlighting the automatic construction of multimodal corpora, efficient multimodal fusion, and vision-aware retrieval mechanisms for under-resourced translation and caption generation. Finally, the discussion turns to our ongoing explorations of retrieval-augmented generation (RAG) and the development of custom large language models for multilingual understanding and generation in under-resourced languages.
