Improving Document Clustering for Short Texts by Long Documents via a Dirichlet Multinomial Allocation Model

摘要

Document clustering for short texts has received considerable interest. Traditional document clustering approaches are designed for long documents and perform poorly for short texts due to the their sparseness representation. To better understand short texts, we observe that words that appear in long documents can enrich short text context and improve the clustering performance for short texts. In this paper, we propose a novel model, namely DDMAfs, which (1) improves the clustering performance of short texts by sharing structural knowledge of long documents to short texts; (2) automatically identifies the number of clusters; (3) separates discriminative words from irrelevant words for long documents to obtain high quality structural knowledge. Our experiments indicate that the DDMA/s model performs well on the synthetic dataset and real datasets. Comparisons between the DDMA/s model and state-of-the-art short text clustering approaches show that the DDMA/s model is effective.