[4Xin1] Poster session 2

[4Xin1-02] Designing Instructions in Crowdsourcing for Collecting Diverse Task-Oriented Dialogue Data

Clinic reservation dialog

Asahi Hentona^1,2, 〇Yuta Tomomatsu¹, Shota Sasaki^3,4, Kaori Abe³, Kentaro Inui^3,4 (1.AI Shift,inc., 2.CyberAgent, Inc., 3.Tohoku University, 4.RIKEN)

Keywords:task oriented dialogue, dialogue corpus, crowdsourcing, dialogue system

In customer service to accomplish some tasks, such as booking a clinic appointment, the operator offers suggestions, questions, and explanations to the customer, and in some cases negotiates compromises. To realize such a flexible task-oriented dialog system, a diverse and large set of dialog data is required. Crowdsourcing is a means of collecting dialogue data on a large scale, but care must be taken in instruction design to avoid monotonous dialogue. In this study, a dialogue corpus was constructed by generating and presenting different instructions to each of the crowdsourcers playing the operator and customer roles, to collect a variety of dialogue data. This corpus consists of data from approximately 100 dialogs for clinic appointments. We also analyze the constructed dialogue corpus and report the challenges and findings for collecting more diverse task-oriented dialogue data.