In the age of social media, customers have become opinion makers: anyone interested in a product can search for opinions on internet platforms. Web-scraping is often the only way to access them, and despite the use of ETL, the heterogeneity of the data makes the task of extracting insights arduous and leads to the need for ad-hoc tools. To circumvent this problem, we apply the Embedded Dirichlet Process and the Embedded Hierarchical Dirichlet Process in an industrial setting around the case of tires. These non-parametric topic models learn topics and their number, topic embeddings and word embeddings, in order to disambiguate words and to offer more analytical levels. They can also be used to refine ETL processes. EDP and EHDP achieve similar or even higher levels of likelihood than the state-of-the-art techniques tested, without re-runs to find the number of topics.