Study And Develop Generative Model Chatbots

Abstract

In the field of open-domain dialog systems, latent variable based models have shown to deliver promising performances. However, response generation in these models cannot be explicitly controlled. Therefore, researchers have come up with SPHRED, inspired by a latent variable model called VHRED. SPHRED addresses the problem and controls generation by generating responses based on attributes like level of genericness, or different sentiment modes, or any attribute defined in accordance with specific contexts. In addition, SPHRED separately models dialog states for both speakers to reflect personal features. The researchers tested SPHRED in two different scenarios, where the attribute refers to genericness and to sentiment states. This thesis focuses on building SPHRED model in TensorFlow; testing its performance to verify its capabilities on Ubuntu dataset [1] in terms of global structure capturing, response generation control, and respective communication style replication; and addressing problems faced by the model, then suggesting improvements to overcome them