Thomas Demeester, Véronique Hoste and Sofie Labat
Request access


On this page, you can download the EmoWOZ-CS corpus. EmoWOZ-CS is a text corpus of 2,148 customer service interactions that were collected through a Wizard of Oz experiment. In our experiment, 179 Dutch-speaking participants were told to be chatting to different versions an autonomous conversational agent, while the system was in reality fully controlled by a student worker (henceforth: wizard) who steered conversations towards a given sentiment (positive, neutral, negative). Each participant had 12 conversations that were grounded in different events related to the domain of customer service. Given an event, participants had to assume the role of customers and contact a company's chatbot to discuss the event.

In EmoWOZ-CS, participant turns are annotated for emotions (emotion labels, valence, arousal, dominance), while wizard turns are annotated for response strategies. The corpus also contains anonymized profiling information on the participants: age, gender, big-5 personality traits, and prior experience in interacting with conversational agents. Finally, the resulting Dutch corpus is automatically translated to English with DeepL Pro and subsequently post-edited by student workers from the Department of Translation, Interpreting and Communication at Ghent University. Both the Dutch and English versions of the corpus are made available on this page.

The Wizard of Oz experiment we conducted was approved by the Ethics Committee of the Faculty of Arts and Philosophy at Ghent University. Upon terminating their 12th conversation, participants were debriefed and the Wizard of Oz setup was revealed to them. At this point, they explicitly received the option to withdraw their consent and data from the corpus.

The data collection and annotation procedure are described in the following paper:

Labat, S., Demeester, T., & Hoste, V. (2024). A Customer Journey in the Land of Oz: Leveraging the Wizard-of-Oz Technique to Model Emotions in Customer Service Interactions. (Paper submitted to Computers in Human Behavior).

Please note that by downloading the data you agree to the following terms and conditions:

  • The authors and their affiliated institutions make no warranties regarding the datasets provided. They cannot be held liable for providing access to the datasets or the usage of the datasets.
  • The dataset should only be used for scientific or research purposes. Any other use is explicitly prohibited.
  • The datasets must not be redistributed or shared in part or full with any third party. Redirect interested parties to this page.
  • If you use any of the datasets, you agree to cite the associated paper.