Biterm-based multilayer perceptron network for tagging short text

With the emergence of online social media, usergenerated content (UGC) in short text form is becoming the most popular and valuable information available on the web. Mining potential resources from large scale short texts is particularly critical but challenging for many content analysis tasks. Among the text mining, one of useful techniques is tagging. Conventional approach to tag short text only use simple text-level word co-occurence. However, different from long articles, short texts have limited words, which do not provide sufficient information on tags. In this paper, we propose a novel method for tagging short texts by using multilayer perceptron network (MLP), which takes full advantage of the generation of word co-occurrences by biterms whose definition will be given in the text. With the help of biterm, we make the inference effective with rich training information in MLP. Experiments on real-word short text collections show that the proposed methods outperforms the tranditional approach on ZHIHU datasets.