B1327
Title: What is love? Describing emotions using prediction models
Authors: Yuval Benjamini - Hebrew University of Jerusalem (Israel) [presenting]
Abstract: New architectures for prediction models have proven that bigger is better in terms of predictive abilities on industry-scale challenges such as natural language processing and image understanding. Incorporating these advances into small lab science poses two outstanding challenges: (a) the data sets are typically too small to fit these models, and (b) it is unknown how to explain the predictions, which is necessary for scientific applications. Prediction models are discussed for characterizing emotion from a rich data set of autobiographical stories. The team has recorded and annotated several hours of emotional autobiographical stories and collected fMRI brain recordings while rehearing these recordings. Publicly available models (large language models, facial expression annotators) are adapted to characterize how different perceived emotions are constructed through varied modalities such as semantic content, facial expression, and speech tone. Both the technical challenges of working with this relatively small data set, as well as the dilemmas in interpreting the results are reviewed.
