Filtering and Classifying Relevant Short Text with a Few Seed Words

Abstract

Filtering out irrelevant documents and classifying the relevant ones into topical categories is a de facto task in many applications. However, supervised learning solutions require extravagant human efforts on document labeling. In this paper, we propose a novel seed-guided topic model for dataless short text classification and filtering, named SSCF. Without using any labeled documents, SSCF takes a few “seed words” for each category of interest, and conducts short text filtering and classification in a weakly supervised manner. To overcome the issues of data sparsity and imbalance, the short text collection is mapped to a collection of pseudodocuments, one for each word. SSCF infers two kinds of topics on pseudo-documents: category-topics and general-topics. Each category-topic is associated with one category of interest, covering the meaning of the latter. In SSCF, we devise a novel word relevance estimation process based on the seed words, for hidden topic inference. The dominating topic of a short text is identified through post inference and then used for filtering and classification. On two real-world datasets in two languages, experimental results show that our proposed SSCF consistently achieves better classification accuracy than state-of-the-art baselines. We also observe that SSCF can even achieve superior performance than the supervised classifiers supervised latent dirichlet allocation (sLDA) and support vector machine (SVM) on some testing tasks.