A Comparative Study on Text Representation Models For Arabic Topic Detection

Resumen

Topic Detection (TD) plays a major role in Natural Language Processing (NLP). Its applications range from Question Answering to Speech Recognition. In order to correctly detect document’s topic, wes hall first proceed with a text representation phase to transform the electronic documents contents into an efficiently software handled form. Significant efforts have been deployed to construct effective text representation models, mainly for English documents. In this paper, we realize a comparative study to investigate the impact of using stems, multi-word terms and named entities as text representation models on Topic Detection for Arabic unvowelized documents. Our experiments indicate that using named entities as text representation model is the most effective approach for Arabic Topic Detection. The performances of the two other approaches are heavily dependent on the considered topic. In order to enhance the Topic Detection results, we use combined vocabulary vectors based on stems and named entities (respectively stems and multi-word terms) association to model topics more accurately. This approach effectiveness has been endorsed by the enhancement of the system performances.