In this study, we explore the use of a hybrid approach in online surveys, combining traditional form-based closed-ended questions with open-ended questions administered by a chatbot.
We trained a chatbot using OpenAI's GPT-3 language model to produce context-dependent probes to responses given to open-ended questions. The goal was to mimic a typical professional survey interviewer scenario where the interviewer is trained to probe the respondent when answering an open-ended question.
For example, assume this initial exchange: “What did you find hard to use or frustrating when using Google Maps?” “It wasn't easy to find the address we were looking for”
The chatbot would follow-up with “What made it hard to find the address?” or “What about it made it difficult to find?” or “What steps did you take to find it?”.
The experiment consisted of a Qualtrics survey with 1,200 participants, who were randomly assigned to one of two groups. Both groups answered closed-ended questions, but the final open-ended question differed between the groups, with one group receiving a chatbot and the other group receiving a single open-ended question.
The results showed that using a chatbot resulted in higher quality and more detailed responses compared to the single open-ended question approach, and respondents indicated a preference towards using a chatbot to open-ended questions. However, respondents also noted the importance of avoiding repetitive probes and expressed dislike for the uncertainty around the number of required exchanges.
This hybrid approach has the potential to provide valuable insights for survey practitioners, although there is room for improvement in the conversation flow.