Melhoria na classificação de tópicos em textos curtos usando background knowledge

Abstract

The power of interaction between internet users has grown since the appearance of tools aligned with the principles of WEB 2.0, including blogs, forums and social networks like Twitter and Facebook. This kind of application is based on frequent message exchanges between users, generating large quantities of textual data comprised of small messages. Text classification techniques allow the extraction of relevant information from such messages. In this context, the challenges are related to the fact that the short messages common in social networks contain, individually, too little data for the traditional analyses. In this work a new technique for improving topic classification in short texts is proposed. This technique is based on the idea of combining a standard text classifier with a keywords-based simple classification scheme; the novelty here beyond the combination of two classification schemes is the use of a semi-automated, unsupervised technique for building the list of keywords reated to the desired topic; this technique is based on the use of topic modeling using the LDA algorithm. To demonstrate the validity of the proposed approach, a Corpus of twitter messages was built around the topic “violence”. This Corpus was used in experiments to assess the performance of the proposed classification technique. Results show that topic classification for short texts is improved by the proposed technique.