| Project Detail |
Intelligent chatbots (ICs) such as ChatGPT have revolutionized the generation of content for a few languages such as English, but there are 7099 currently spoken languages in the world. EPICAL will, for the first time, determine how to add new low resource languages (LRLs) to ICs. We will make six advances to revolutionize the capabilities of ICs, unifying different areas of research that are incorrectly studied separately. We will: 1) determine how to generate hallucination-free text using ICs, and how to enter a virtuous cycle where LRL text is created using cross-lingual knowledge from ICs and then quickly post-edited and trained upon, resulting in a better LRL representation in the IC. 2) develop more powerful encoding and language adaptation approaches which combine the benefits of fine-tuning and adapters, taking full advantage of linguistically related languages to model LRLs. 3) enable ICs to reason about their own LRL capabilities and determine what they know and do not know. 4) unify research on machine translation and ICs to obtain ICs which can translate to LRLs with state-of-the-art accuracy. 5) enable high quality text-to-speech and automatic speech recognition of LRLs with ICs, thereby unifying the research on low resource speech processing with research on LRL text processing. 6) develop a novel evaluation methodology including a robust method for automatically measuring fact hallucination. My research group is well-known for LRL research, which differs from large commercial labs focusing only on the top 200 languages. Our work is critical for a multilingual Europe which values the role of minority languages, culture and heritage. Our innovations will benefit natural language processing beyond text generation and machine translation and strongly impact other areas of machine learning research suffering from data bottlenecks. |
