How People Guess Others' Interests from Short Profile Descriptions

Abstract

This paper reports a study on predicting others' interests from short profile descriptions, such as "Ms. Fitzgerald is a retired woman." We first conducted surveys using a crowd-sourcing service to investigate people's tendencies in predicting others' interests over 46 profiles and 49 subjects of interest. The survey results of 315,560 responses from 2,129 different participants show that people demonstrate biased but quite stable tendencies within and between gender/age groups. We then had BERT and RoBERTa perform the same predictions using masked language modeling and found that it could very weakly emulate people's tendencies. The presented methodology shows a potential as a unique metric to measure the extent to which an artificial agent would behave in accordance with a part of social common sense.