Context reinforced neural topic modeling over short texts

As one of the prevalent topic mining methods, neural modeling has attracted a lot interests due to advantages low training costs and strong generalisation abilities. However, existing models may suffer from feature sparsity problem when applied short texts, lack context in each message. To alleviate this issue, we propose Context Reinforced Neural Topic Model (CRNTM), whose characteristics can be summarized as follows. First, by assuming that text covers only few salient topics, proposed CRNTM infers for word narrow range. Second, our model exploits pre-trained embeddings treating topics multivariate Gaussian distributions or mixture embedding space. Extensive experiments on two benchmark corpora validate effectiveness both discovery classification.