Neural variational sparse topic model for sparse explainable text representation

Texts are the major information carrier for internet users, from which learning latent representations has important research and practical value. Neural topic models have been proposed great performance in extracting interpretable topics of texts. However, there remain two limitations: (1) these methods generally ignore contextual texts limited feature representation ability due to shallow feed-forward network architecture, (2) Sparsity semantic space is ignored. To address issues, this paper, we propose a reinforcement neural variational sparse model (SR-NSTM) towards explainable text learning. Compared with existing models, SR-NSTM generative process probabilistic distributions parameterized networks incorporates Bi-directional LSTM embed at document level. It achieves posterior over documents words zero-mean Laplace distribution sparsemax. Moreover, supervised extension via adding max-margin regularization tackle tasks. The inference method utilized learn our efficiently. Experimental results on Web Snippets, 20Newsgroups, BBC, Biomedical datasets demonstrate that revisiting can improve performance, leading competitive coherent • We representation. extend tasks constraints. superiority perplexity, coherence classification accuracy.