Collaborative is better than adversarial: : generative cooperative networks for topic clustering

Topic modeling is a popular technique for learning the thematic structure of large corpora composed of unlabeled documents, without human supervision. In recent years, various neural network-based algorithms have been proposed to solve this task. In particular, there is an extensive literature showing how Variational AutoEncoders (VAEs) and Generative Adversarial Networks (GANs) approaches have been successful in identifying recurrent discussion topics. In this paper we propose a new neural topic detection model called Generative Cooperative Topic Modeling (GCTM), in which a Generator and a denoising AutoEncoder, rather than learning through a competitive process, act cooperatively. We show that this cooperative model has a faster convergence and surpasses the adversarial approach, as well as other popular topic detection algorithms based on VAEs, when tested on three common public datasets and with a variety of performance indicators.