Title: Keeping Track of Entities Over Time, Minds, and Knowledge Sources
Abstract: Transformer-based language models form the basis of many state-of-the-art NLP systems across a wide range of tasks. Yet there is now ample evidence that these systems fail to systematically generalize to rare events, and make mistakes in reasoning and logical inferences. I argue that many of these issues stem from the fact that the training tasks of language modelling and finetuning are not sufficient for keeping track of entity and world states over a discourse, or for comparing or updating world states based on newly ingested text. I'll discuss my lab's work on creating probes, datasets, and evaluations that demonstrate these specific problems. I'll then present work that highlights the promise of explicitly modelling entities in different contexts, from integrating external knowledge for the prediction of rare events, to adjudicating between conflicting world states to solve theory-of-mind questions.
Speaker Bio: Jackie Chi Kit Cheung is an associate professor at McGill University's School of Computer Science where he co-directs the Reasoning and Learning Lab, and a Canada CIFAR AI Chair at the Mila Quebec AI Institute. His research focuses on natural language generation tasks such as automatic summarization, and on integrating diverse sources of knowledge into NLP systems for pragmatic and common-sense reasoning. He is motivated in particular by how the structure of the world can be reflected in the structure of language processing systems. He is a consulting researcher at Microsoft Research Montreal.
Date and Time:
Thursday, March 3, 2022
1:00 PM - EST
Recording: https://youtu.be/FGxa9xszJHY