This paper presents a Wizard of Oz (WOZ) data collection method that uses dialogue examples (or utterances) from one domain for use in a target domain. Providing a text-based dialogue system with empathy requires providing the system with a wide range of expressions, with expressions corresponding best to users. However, there are few dialogue examples available and the variation of utterances is limited. We have to collect wider range of example utterances. A typical method to collect dialogue data is the WOZ method. The use of WOZ for dialogue data collection often requires substantial cognitive load for participating wizards. To alleviate this problem, an utterance suggestion mechanism using a portable corpus is introduced. We investigated differences in the response times of a wizard when utterance suggestions from a portable corpus are offered. We also evaluated the ratio of utterance suggestions selected versus free utterances. The experimental results indicate that using a portable dialogue corpus to suggest utterances for wizards has a potential to be helpful in data collection.
展开▼