Date: Friday, March 15,  2024
Time: 12:00pm – 1:00pm
Location: Simons Center Conference room 46-6011 + Zoom Meeting (https://mit.zoom.us/j/98903444024)

Speaker: Thomas Clark
Affiliation: PhD student, TedLab and Computational Psycholinguistics Lab, Department of Brain and Cognitive Sciences, MIT

Talk title: Quantifying correlations between linguistic surprisal, prosody, and backchannels in naturalistic conversation

Abstract: Research indicates that more predictable (i.e. less surprising) words are generally reduced in duration. The relationship between surprisal and other prosodic features, as well as listener backchannels (short interjections such as “yeah” or “mhm” that are common in conversation) remains understudied. Using large language models to compute word surprisal, we investigate a new large corpus of naturalistic conversations (CANDOR). We find that surprisal robustly predicts duration, pitch, and intensity even when controlling for possible confounds, and is negatively correlated with backchannels. These preliminary results indicate a sensitivity of both speaker and listener behavior to the information content of words in everyday conversation.