Improving Short Text Classification using Information from DBpedia Ontology

摘要

With the emergence of social networks and micro-blogs, a huge amount of short textual documents are generated on a daily basis, for which effective tools for organization and classification are needed. These short text documents have extremely sparse representation, which is the main cause for the poor classification performance. We propose a new approach, where we identify relevant concepts in short text documents with the use of the DBpedia Spotlight framework and enrich the text with information derived from DBpedia ontology, which reduces the sparseness. We have developed six variants of text enrichment methods and tested them on four short text datasets using seven classification algorithms. The obtained results were compared to those of the baseline approach, among themselves, and also to some state-of-the-art text classification methods. Beside classification performance, the influence of the concepts similarity threshold and the size of the training data were also evaluated. The results show that the proposed text enrichment approach significantly improves classification of short texts and is robust with respect to different input sources, domains, and sizes of available training data. The proposed text enrichment methods proved to be beneficial for classification of short text documents, especially when only a small amount of documents are available for training.