Title: Dealing with uncertainty in language-based AI
Authors: Christian Hardmeier - IT University of Copenhagen (Denmark) [presenting]
Abstract: Language-based artificial intelligence (AI) in the form of large generative language models has achieved breakthroughs with enormous public impact in the recent past. Reliable uncertainty estimation is essential to create trustworthy models. However, this is rendered difficult by the sheer size of the models, the complexity of human language, and the reasoning capabilities expected in such models. The challenges of dealing with uncertainty in large language models are discussed, and recent work on harnessing evidential deep learning for uncertainty estimation in natural language processing is presented.