Making something out of nothing: Building robust task-oriented dialogue systems from scratch
In this paper, we introduce Gauchobot, a task-oriented dialogue system developed for the Alexa Prize TaskBot Challenge. We identified two main obstacles to building better conversational AI assistants in real-world applications. The first is great human efforts needed in data annotation and engineering a dialogue system that provides service in a new domain from scratch. The high cost, often ignored by existing research work, has blocked the broad deployment of dialogue systems. The second is a lack of robustness when facing undesirable situations during a conversation in real scenarios. The existing paradigm, which pre-defines dialogue flows and confines the users to a box with restricted options, makes dialogue systems easily stumped by complex conversations. To solve these two issues, we invent a methodology that can automatically generate data with minimum human efforts to train a unified framework capable of handling various sub-tasks in completing a task-oriented conversation. Feedback from real-world users can be easily incorporated into the model by automatically generating more training data and thus improving the model over time. Besides, we integrate multiple generative-based and retrieval-based response generation models into our Gauchobot, making it capable of handling not only task-oriented commands, but also QA, chit-chat, and other out-of-domain cases. As a result, Gauchobot can not only help complete tasks with rich user experience, but also provide a general framework of building robust task-oriented dialogue systems quickly from scratch.