Feb 15, 2018 (modified: Feb 15, 2018)ICLR 2018 Conference Blind Submissionreaders: everyoneShow Bibtex
Abstract:Topic modeling of text documents is one of the most important tasks in representation learning. In this work, we propose iTM-VAE, which is a Bayesian nonparametric (BNP) topic model with variational auto-encoders. On one hand, as a BNP topic model, iTM-VAE potentially has infinite topics and can adapt the topic number to data automatically. On the other hand, different with the other BNP topic models, the inference of iTM-VAE is modeled by neural networks, which has rich representation capacity and can be computed in a simple feed-forward manner. Two variants of iTM-VAE are also proposed in this paper, where iTM-VAE-Prod models the generative process in products-of-experts fashion for better performance and iTM-VAE-G places a prior over the concentration parameter such that the model can adapt a suitable concentration parameter to data automatically. Experimental results on 20News and Reuters RCV1-V2 datasets show that the proposed models outperform the state-of-the-arts in terms of perplexity, topic coherence and document retrieval tasks. Moreover, the ability of adjusting the concentration parameter to data is also confirmed by experiments.
TL;DR:A Bayesian Nonparametric Topic Model with Variational Auto-Encoders which achieves the state-of-the-arts on public benchmarks in terms of perplexity, topic coherence and retrieval tasks.