This work focuses on analyzing online forum threads by characterizing them as conversation trees of topics. It introduces models that perform two tasks simultaneously: segmenting a thread into subparts and assigning topics to each segment. The approach uses probabilistic grammars, specifically Context-Free Grammars and Linear Context-Free Rewriting Systems, to create hierarchical and coherent topic structures. These models are shown to be effective, outperforming several baseline methods in generating structured and topic-labeled forum threads.