As in demo 5A, I suggest using sklearn's implementation of LDA, which is the best option for generating topical relevance at the document level. We need to do a little tuning for our model. I recommend doing this on a smaller sample of 5,000 records to save time. You should try between 40 and 150 topics, counting by 10s. You should set topic_word_prior equal to 0.15 and doc_topic_prior to 25 divided by the number of topics. Finally, we will use the coherence score, u_mass, as in demo 5A to evaluate topic quality