Tell me who you are and i'll tell you what to do: A Persona Grounded Task Oriented Dialogue Generation System

Anonymous

Abstract: Modern dialogue agents can broadly be categorized as either chit-chat or task-oriented systems. While the purpose of a chit-chat agent is to entertain and engage the user- lubricating the conversation, so to say-, the task oriented chat-bot is dedicated to fulfilling specific requests (e.g., ticket booking). Current task-oriented agents produce precise but bland and uninteresting responses. While using such agents a user may interpose personal remarks, and the failure of the agent to process and respond to such statements could be a put-off for the user. In this paper we propose a system that is persona-specific, can handle chit-chat utterances, and produces responses that add a human element to the conversation, while always remaining grounded on the task. Since current task-oriented datasets do not have persona-profiles, and do not consist of personalized remarks in utterances, we modify an existing dataset (MultiWOZ 2.1) to suit our needs. We give a semi-automated dataset creation method that uses GPT-2 model trained on the PERSONA-CHAT dataset. A small subset of the obtained data is also manually crafted to acquire a gold standard data. Our framework is based on GPT-2, Graph Convolution Network (GCN) and Memory Network that is trained on this dataset to generate persona-grounded task-oriented responses. Both automatic and manual evaluation show the effectiveness of our model and dataset. Our proposed system achieves a BLEU score of 12.12 on this new dataset.