Toward Artificial Social Intelligence: A Semi-supervised, Split Decoder Approach to EQ in a Conversational Agent

Abstract

This paper explores the idea of making text-based conversational agents more human-like by engineering emotional capabilities in them. The emotional quotient is still missing in the current generation of conversational chatbots. Since this is a broad problem, this research focuses on a key building block of such agents, i.e., ability to respond with a particular emotion given an input text embedded with a certain emotion. This work achieves this research goal by using a few innovative strategies, i.e., a layering of the system architecture, semi-supervised learning to workaround lack of labeled data, an innovative agent model based on a classical decision-maker coupled with a deep learning-based split decoder architecture, greedy training of the individual components of the model, an innovative approach for model evaluation based on emotion intensity, and finally, selection of final weights of the components of the model based on experimental results measured in terms of perplexity. The results obtained are extremely promising.