Abstract: Neural topic models have attracted much attention for their high efficiencies in training, in which, the methods based on variational auto-encoder capture approximative distributions of data, and those based on Generative Adversarial Net (GAN) are able to capture an accurate posterior distribution. However, the existing GAN-based neural topic model fails to model the document-topic distribution of input samples, making it difficult to get the representations of data in the latent topic space for downstream tasks. Moreover, to utilize the topics discovered by these topic models, it is time-consuming to manually interpret the meaning of topics, label the generated topics, and filter out interested topics. To address these limitations, we propose a Reward-Modulated Adversarial Topic Model (RMATM). By integrating a topic predictor and a reward function in GAN, our RMATM can capture the document-topic distribution and discover interested topics according to topic-related seed words. Furthermore, benefit from the reward function using topic-related seed words as weak supervision, RMATM is able to classify unlabeled documents. Extensive experiments on four benchmark corpora have well validated the effectiveness of RMATM.
