||In a questionnaire with a variety of questions for consumers to answer, there may be a small number of specific answers that are far-off from the distribution and different from most of the respondents. These anormal responses have a significant impact on the results. However, in a large-scale questionnaire survey, it is generally difficult to identify such respondents individually. Therefore, in this paper, we report a result of consideration of anormal respondent detection in the feature space using variational autoencoder as a dimension reduction method. The large-scale questionnaire response data of approximately 1,000 questions (approximately 46,000 including sub-questions), which was conducted for approximately 15,000 people, are divided into question categories. And focusing on the answers that are deviated from the majority of others in each of those categories, we detect the respondents who gave anormal answers in each question category. As a result, we find multiple cases where a person detected as an anormal respondent in one question category is also detected as an anormal respondent in another question category. Therefore, it is inferred that these respondents should be excluded from the aggregated results.