In this article, we propose a generic method to build thematic datasets from social media. Many research works gather their data from social media, but the extraction processes used are mostly ad hoc and do not follow a formal or standardized method. We aim at extending the processes currently used by designing an iterative, generic and domain-independent approach to build thematic datasets from social media with three modulable dimensions at its core: spatial, temporal and thematic. We experiment our method using data extracted from Twitter to build a thematic dataset about tourism in a highly touristic region. This dataset is then evaluated using both quantitative and qualitative metrics to highlight the value of this method. The application to this use case shows the effectiveness of our domain-independent method to generate thematic datasets from Twitter data.