Discrete-State Variational Autoencoders for Joint Discovery and Factorization of Relations

摘要: We present a method for unsupervised open-domain relation discovery. In contrast to previous (mostly generative and agglomerative clustering) approaches, our model relies on rich contextual features makes minimal independence assumptions. The is composed of two parts: feature-rich extractor, which predicts semantic between entities, factorization model, reconstructs arguments (i.e., the entities) relying predicted relation. components are estimated jointly so as minimize errors in recovering arguments. study models inspired by work selectional preference modeling. Our substantially outperform agglomerative-clustering counterparts achieve state-of-the-art performance.