Kathy McKeown Interviews Bonnie Webber

Lifetime Achievement Award Kathy McKeown Interviews Bonnie Webber Bonnie Webber University of Edinburgh School of Informatics bonnie.webber@ed.ac.uk Because the 2020 ACL Lifetime Achievement Award presentation could not be done in person, we replaced the usual LTA talk with an interview between Professor Kathy McKeown (Columbia University) and the recipient, Bonnie Webber. The following is an edited version of the interview, with added citations. Bonnie Webber: Let me start by thanking you, Kathy, for agreeing to interview me. I think you were the first Ph.D. student I met when I joined the faculty at Penn after finishing my own Ph.D. You were Aravind Joshi’s Ph.D. student at the time, working on generating answers to definition questions. I believe your Ph.D. thesis then became the first book published in the Cambridge University Press series on Natural Language Processing, of which Aravind was the editor-in-chief. Kathy McKeown: Well, Bonnie, thanks so much for asking me to do this. I really will enjoy it. What I remember of you from that time was your arrival in the department like a breath of fresh air. You were very animated, always vocal. And to a young woman like myself at that time, you were an amazing role model. So I’m happy to have this chance to ask you more questions than I could for my keynote address. You’ve frequently asserted that we shouldn’t develop computational models based simply on standard assumptions about a linguistic phenomenon, but rather that we should first try to characterize it as fully as possible and then see which of these can be covered computationally. Given that current approaches in NLP don’t focus on subtle linguistic differences and what a model does or doesn’t cover, but instead worry about scale, do you think your approach to research still has something to contribute? Bonnie: Well, I certainly hope so! Let me answer the question in terms of standard metrics used in Natural Language Processing—in particular, F-score, Precision, and Recall. Although F-score has mostly been used as a basis for “League Table” ratings, I would contend that there is much to be learned from its components—Precision and Recall. I always encourage my students to look at them very carefully. Let me elaborate. Precision quantifies incorrect positive predictions—cases where a system has generalized incorrectly about what belongs to the class of interest, saying more things belong to it than actually do. Recall, on the other hand, quantifies incor- rect negative predictions—cases where a system has generalized incorrectly about https://doi.org/10.1162/COLI a 00393 © 2021 Association for Computational Linguistics Published under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0) license Computational Linguistics Volume 47, Number 1 what doesn’t belong to the class of interest, omitting things from the class that should be there. Now, there are at least two possible reasons for low Precision and/or low Recall: Either the system itself is unable to learn from the data because it’s too puny, or there’s something wrong with the data. What doesn’t surprise me, having spent years …

Content truncated to 500 words