to the best of my knowledge:   1. SVD 2. therefore either rotated or unrotated eigenvectors, depending on 'allow correlated topics' node property setting 3. no, the topics are somewhat correlated whilst within-vector variance is minimised, which in turn increases the likelihood of a human interpreting the topic theme with a more relevant, grammatical phrase, based on the most influential topic terms.   If I'm wrong, I'd appreciate a correction. ... View more