TALK ABSTRACT:

Despite the great success of Large Language Models (LLMs) in many NLP tasks, they still suffer from several serious flaws. For instance, they struggle with tasks involving multiple documents, they are not interpretable and they are not able to develop complex plans in either solving problems or text generation. It also seems that simply making them more domain specific or just scaling them up is problematic.  In this talk, I will argue that more powerful discourse processing could come to the rescue, but only if two key challenges are addressed. First, we need to be able to train modern discourse parsers (NLU) and generators (NLG)   across domains and genres,  for both monologues and dialogues,  without requiring substantial human annotation. Secondly, we need to better  understand what discourse info is missing in LLMs and how to inject such missing info into current LLMs. To conclude, I will then discuss further longer-term open issues involving discourse processing for LLMs designed to deal with long documents and with decoder-only architectures, as well as how to integrate different theories of discourse to better enhance the understanding and generation capabilities of LLMs. 

ABOUT OUR SPEAKER: Giuseppe Carenini is a Professor in Computer Science and Director of the Master in Data Science at UBC (Vancouver, Canada). His work on natural language processing and information visualization to support decision making has been published in over 150 peer-reviewed papers (including best paper at UMAP-14 and ACM-TiiS-14). Dr. Carenini was the area chair for many conferences including recently for ACL'21 in “Natural language Generation”, as well as Senior Area Chair for NAACL'21 in “Discourse and Pragmatics”.  Dr. Carenini  was also the Program Co-Chair for IUI 2015 and  for SigDial 2016. In 2011, he published a co-authored book on “Methods for Mining and Summarizing Text Conversations”.  In his work, Dr. Carenini has also extensively collaborated with industrial partners, including Microsoft and IBM. He was awarded a Google Research Award in 2007 and a Yahoo Faculty Research Award in 2016.

LinkedIn