Currently, most surveys ask for occupation with open-ended questions. The verbatim responses are coded afterwards, which is error-prone and expensive. When textual answers have a low level of detail, exact coding may be impossible. We improve this process by asking response-dependent questions during the interview. Candidate job categories are predicted with a machine learning algorithm. When chances are high that the forecasted job is correct, the interviewer presents a short list of related jobs to the respondent. We compare the new technique to traditional coding practice and describe and quantify different error sources for both methods.
