The increasing volume of large, multi-thematic text corpora in social sciences presents a challenge in selecting relevant documents for qualitative and mixed-methods research. Traditional sample selection methods require extensive manual coding or prior dataset knowledge, while unsupervised methods can yield inconsistent results with theory-driven coding. To address this, we propose purposive semantic sampling – a Natural Language Processing approach using document-level embeddings created by a weighted average of word vectors with term frequency-inverse document frequency (tf-idf). We demonstrate its effectiveness using the example of democracy, a complex topic difficult to retrieve from parliamentary corpora. This method applies to any multi-thematic research area within big data, offering a reliable, efficient sample selection method for social research texts. Our contribution includes validating this NLP approach for social sciences and humanities as well as providing a robust tool for researchers, facilitating deeper qualitative analysis and exploration of big data corpora within the computational grounded theory framework.