SemanticGraph GuidedTopicDiscovery

Hierarchical topic models, such asgamma beliefnetwork, isstructured asatreewhere alltheleaf nodes of aparent node are on the same floor. Due tothebottom layer conceptisdefined astheword layer inthetopic model, all the children nodes of the bottom layer node concept are connected to its ancestor node. Extracting a four-layer subgraph from the WordNet that are rooted at these 736 terms, we obtain apriorWordNet with[736,314,152,52,11] nodes fromthebottom totoplayers. For Tab. 2, we choose a vocabulary of 20,000 terms, which has 4693 terms that overlap with the vocabulary ofthe WordNet.