Short-text topic modeling with dual reinforcement from internal and external semantics

Given the prevalence of short texts as a popular form of information on the Internet, inferring latent topics from short texts has attracted increasing interest from both academia and industry. To address the data sparsity of short texts in terms of word co-occurrences, existing research efforts either try to extract more information from the given data internally or leverage externally learned semantic information such as pre-trained word embeddings. In this paper, we propose a novel model, called Dual-Reinforced Topic Model (DRTM), to identify topics from short texts by harnessing both internal and external semantic information. Improving existing internal methods that consider only first-order co-occurrence relations between words, our model exploits multi-order relations so that the relevance between words not explicitly appearing together in the given data can be captured. Addressing the limitation of existing external methods that utilize only distributed representations at the word level, we further incorporate document representations into our model to facilitate topic modeling. We have evaluated our model on multiple publicly available datasets. Our experimental results have demonstrated that DRTM clearly outperforms existing internal and external methods in terms of both topic coherence and document classification accuracy.