Utilizing unsupervised crowdsourcing to develop a machine learning model for virtual human animation prediction

One type of experiential learning in the medical domain is chat interactions with a virtual human. These virtual humans play the role of a patient and allow students to practice skills such as communication and empathy in a safe, but realistic sandbox. These interactions last 10-15 min, and the typical virtual human has approximately 200 responses. Part of the realism of the virtual human's response is the associated animation. These animations can be time consuming to create and associate with each response. We turned to crowdsourcing to assist with this problem. We decomposed the process of creating basic animations into a simple task that nonexpert workers can complete. We provided workers with a set of predefined basic animations: six focused on head animation and nine focused on body animation. These animations could be mixed and matched for each question/response pair. Then, we used this unsupervised process to create a machine learning model for animation prediction: one for head animation and one for body animation. Multiple models were evaluated and their performance was assessed. In an experiment, we evaluated participant perception of multiple versions of a virtual human suffering from dyspepsia (heartburn-like symptoms). For the version of the virtual human that utilized our machine learning approach, participants rated the character's animation on par with a commercial expert. Head animation specifically was rated more natural and typically expected than other versions. Additionally, analysis of time and cost show the machine learning approach to be quicker and cheaper than an expert alternative.