Improving asynchronous interview interaction with follow-up question generation

Abstract

The user experience of an asynchronous video interview system, conventionally is not reciprocal or conversational. Interview applicants expect that, like a typical face-to-face interview, they are innate and coherent. We posit that the planned adoption of limited probing through follow-up questions is an important step towards improving the interaction. We propose a follow-up question generation model (followQG) capable of generating relevant and diverse follow-up questions based on the previously asked questions, and their answers. We implement a 3D virtual interviewing system, Maya, with capability of follow-up question generation. Existing asynchronous interviewing systems are not dynamic with scripted and repetitive questions. In comparison, Maya responds with relevant follow-up questions, a largely unexplored feature of virtual interview systems. We take advantage of the implicit knowledge from deep pre-trained language models to generate rich and varied natural language follow-up questions. Empirical results suggest that followQG generates questions that humans rate as high quality, achieving 77% relevance. A comparison with strong baselines of neural network and rule-based systems show that it produces better quality questions. The corpus used for fine-tuning is made publicly available.