PROJECT TITLE :

Hierarchical Prediction and Adversarial Learning For Conditional Response Generation

ABSTRACT:

When it comes to what and how people communicate in their day-to-day lives, there are a number of underlying factors that play a role. Conversational systems are able to generate positive responses from users and establish cordial connections with them because of their capacity to recognize and make use of the aforementioned factors. In this body of work, we investigate the roles that emotion and intention play as major factors in the generation of responses. We create a hierarchical variational model that predicts in sequence the emotion and intention to be conveyed in a response so that we can investigate the dependency that exists between the two of them. Once the predictions have been made, the response can be generated word-by-word using those predictions. In addition, we make use of an innovative adversarial-augmented inference network in order to make model training easier. The outcomes of the experiments provide evidence of both the viability of the model that was proposed and the viability of the novel adversarial objective. In addition, the hypothesis that human emotions shape human behavior in Communication has been shown to be correct.