We introduce a novel discriminative latentvariablemodel for the task of bilingual lexiconinduction. Our model combines the bipartitematching dictionary prior of Haghighiet al. (2008) with a state-of-the-art embeddingbasedapproach. To train the model, we derivean efficient Viterbi EM algorithm. We provideempirical improvements on six language pairsunder two metrics and show that the prior theoreticallyand empirically helps to mitigate thehubness problem. We also demonstrate howprevious work may be viewed as a similarlyfashioned latent-variable model, albeit with adifferent prior.1