This paper proposes a hierarchical, supervised topic model that builds on ideas from hierarchical LDA using the nested CRP and from supervised LDA, which associates topics with regression coefficients. The proposed model further extends hierarchical LDA by allowing each sentence in a document to chose its own path through the topic tree, and sharing of paths for sentences within the same document is modelled through a document-level CRP where the customers are sentences and the dishes are paths through the topic hierarchy ("combos"). Another difference to supervised LDA is that each word is also associated with its own regression coefficient to model topic-independent effects. The authors empirically demonstrate the performance of the model on 3 data sets and show that the proposed model outperforms LDA linear regression as well as supervised LDA. The idea of modelling framing in the proposed way by associating regression coefficients with hierarchically arranged topics is interesting and appears novel.