Is Deep Learning a Valid Approach for Inferring Subjective Self-Disclosure in Human-Robot Interactions?

摘要

One limitation of social robots has been the ability of the models they operate on to infer meaningful social information about people's subjective perceptions, specifically from non-invasive behavioral cues. Accordingly, our paper aims to demonstrate how different deep learning architectures trained on data from human-robot, human-human, and human-agent interactions can help artificial agents to extract meaning, in terms of people's subjective perceptions, in speech-based interactions. Here we focus on identifying people's perceptions of their subjective self-disclosure (i.e., to what extent one perceives to be sharing personal information with an agent). We approached this problem in a data-first manner, prioritizing high quality data over complex model architectures. In this context, we aimed to examine the extent to which relatively simple deep neural networks could extract non-lexical features related to this kind of subjective self perception. We show that five standard neural network architectures and one novel architecture, which we call a Hopfield Convolutional Neural Network, are all able to extract meaningful features from speech data relating to subjective self-disclosure.