Evaluation of Topic Identification Methods on Arabic Corpora
Résumé
Topic Identification is one of the important keys
for the success of many applications. Indeed, there are few
works in this field concerning Arabic language because of
lack of standard corpora. In this study, we will provide directly
comparable results of six text categorization methods on a
new Arabic corpus Alwatan-2004. Hence, Topic Unigram
Language Model (TULM), Term Frequency/Inverse Document
Frequency (TFIDF), Neural Network, SVM, M-SVM and TR
have been experimented, and showed that TR-Classifier is
the most efficient among the set of classifiers, nevertheless,
only binary SVM outperformed it thanks to its characteristics.
Moreover, we should note that the size of Alwatan-2004 corpus
used to achieve our experiments is considered the most
important compared to any other Arabic corpus which had
been used for topic identification experiments until now. In
addition, we aim through using small sizes of vocabularies to
reduce the time of computation. This is important for adaptive
language modeling, particularly Topic Adaptation, which is
required in real time applications such as speech recognition
and machine translation systems. Our experiments indicate
that the results are better than other works dealing with Arabic
text categorization.