... therefore, we create an in-dividual model for every (ambiguous) word ratherthan a global model for an entire document col-lection. We also show how multiple informationsources can be straightforwardly ... in LDA-related models (Grif-fiths and Steyvers, 2002). For simplicity, we useduniform weights for the layers. The Gibbs samplerwas run for 2,000 iterations. Due to the random-ized nature of ... addict8. company, dealer, police, enforcement, patientTable 1: Senses inferred for the word drug fromthe WSJ and BNC corpora.tail). For the model trained on WSJ, performancepeaks at four senses,...