IC Lab wins prestigious ERC Starting Grant

Most of the benefits from powerful AI language tools, such as ChatGPT, have, for the most part, helped people who speak languages that are common in online writing—especially English. But there are more than 7,000 languages spoken in the world and most of them don’t have any written content to train language specific AI models.

To be useful for everyone, AI needs to understand and communicate in the languages people use and also know about their local culture and environment, necessitating new ways to teach and train models about less common languages.

In the RESPECT-LM project, awarded 1.5M Euros through an ERC Starting Grant, Professor Bosselut and his team will develop new, smart algorithms that will help AI to understand questions in any language, in real time and improve how the AI understands both the language and the knowledge behind the question.

“These new algorithms will use what different languages have in common to help learn new ones, while also separating out the understanding of local culture and traditions that are specific to the speakers of each language,” explained Bosselut.

The project will also create new, modular architectures to separate language from knowledge while the AI is being trained as well as developing new benchmarks and standards to measure how well the AI understands local knowledge in different languages.