Improved Intention Discovery with Classified Emotions in A Modified POMDP
Abstract
Emotions are one of the most proactive topics in psychology, a basis of forceful conversation and divergence from the earliest philosophers and other thinkers to the present day. Human emotion classification using different machine learning techniques is an active area of research over the last decade. This investigation discusses a new approach for virtual agents to better understand and interact with the user. Our research focuses on deducing the belief state of a user who interacts with a single agent using recognized emotions from the text/speech based input. We built a customized decision tree with six primary states of emotions being recognized from different sets of inputs. The belief state at each given instance of time slice is inferred by drawing a belief network using the different sets of emotions and calculating state of belief using a POMDP (Partially Observable Markov Decision Process) based solver. Hence the existing POMDP model is customized in order to incorporate emotion as observations for finding the possible user intentions. This helps to overcome the limitations of the present methods to better recognize the belief state. As well, the new approach allows us to analyze human emotional behaviour in indefinite environments and helps to generate an effective interaction between the human and the computer.