White Paper by the Stanford Institute for Human-Centered AI (HAI), the Asia Foundation and the University of Pretoria: “…maps the LLM development landscape for low-resource languages, highlighting challenges, trade-offs, and strategies to increase investment; prioritize cross-disciplinary, community-driven development; and ensure fair data ownership…
- Large language model (LLM) development suffers from a digital divide: Most major LLMs underperform for non-English—and especially low-resource—languages; are not attuned to relevant cultural contexts; and are not accessible in parts of the Global South.
- Low-resource languages (such as Swahili or Burmese) face two crucial limitations: a scarcity of labeled and unlabeled language data and poor quality data that is not sufficiently representative of the languages and their sociocultural contexts.
- To bridge these gaps, researchers and developers are exploring different technical approaches to developing LLMs that better perform for and represent low-resource languages but come with different trade-offs:
- Massively multilingual models, developed primarily by large U.S.-based firms, aim to improve performance for more languages by including a wider range of (100-plus) languages in their training datasets.
- Regional multilingual models, developed by academics, governments, and nonprofits in the Global South, use smaller training datasets made up of 10-20 low-resource languages to better cater to and represent a smaller group of languages and cultures.
- Monolingual or monocultural models, developed by a variety of public and private actors, are trained on or fine-tuned for a single low-resource language and thus tailored to perform well for that language…(More)”
