Interviewing is a critical human-to-human task in multiple industries, including journalism, human resources, and qualitative research fieldwork. Interviewing can be broadly defined as an interactive goal-orientated knowledge acquisition task on a target source. Unlike data mining or exploration of a static document, interviewing is a dynamic, iterative process that adjusts accordingly based on the response of the individual, the current state of explored knowledge, external knowledge, and the goal (Zhou et al., 2019).

Despite the industrial interest in interviewing agents, most of the previous approaches are rule-based and domain-specific, for example the diagnostic agent MANDY (Lin et al., 2017). While a substantial amount of research has been directed towards Question-Answering agents, not much progress has been made on designing an agent that can ask multi-turn questions exploratively and strategically.

The difficulty lies in the characteristics of the interviewing task, including the difficulty of constructing multi-turn interviewing data, the nature of one-to-many action and goal relationships, in which multiple paths can achieve the same conversational goal, and the insufficiency of the available automatically computed performance metrics.

Recently, there has been emerging research on conversational question generation, which aims to make dialogue questions more interdependent than separate (Gao et al., 2019). Simultaneously, some research has started looking at ways to generate questions that are coherent with the given context, and at the same time enable exploration of answers outside the current context (Pan et al., 2019; Nakanishi et al., 2019).

The project proposed in this abstract aims to develop an interviewing agent capable of not only asking questions exploratively but also learning a strategy that aims to complete an implicit goal, such as gathering enough information for writing a news article given a topic.