The Winograd Schema Challenge: Explained

The podcast episode discusses the Winograd Schema Challenge, which involves pairs of sentences that are similar but differ by one or two words that completely change the meaning. For example, changing 'feared' to 'advocated' in a sentence about the City Council and demonstrators changes the referent of the pronoun 'they.' Previous natural language processing methodologies were unable to tackle this challenge, but the advent of large language models has changed that. Researcher Vit Kotsian, who works at Kumo AI and previously at the University of Oxford, has focused on common sense reasoning and the impact of pre-training in language models. While large language models have made significant progress in solving the Winograd Schema Challenge, they still don't completely address the problem of common sense reasoning.