TR-Classifier and kNN Evaluation for Topic Identification tasks

International audienceThis paper focuses on studying topic identificationfor Arabic language by using two methods. The first method isthe well-known kNN (k Nearest Neighbors) which is used asbaseline. The second one is the TR-Classifier, mainly based oncomputing triggers. The experiments show that TR-Classifier hasthe advantage to give best performances compared to kNN, byusing much reduced sizes of Topic Vocabularies. TR-Classifierperformance is enhanced by increasing jointly the number oftriggers and the size of topic vocabularies. It should be noted thattopic vocabularies are used by the TR-Classifier. Whereas, ageneral vocabulary is needed for kNN, and it is obtained by theconcatenation of those used by the TR-Classifier. In addition tothe standard measures Recall and Precision used for theevaluation step, we have drawn ROC curves for some topics toillustrate more clearly the difference in performance between thetwo classifiers. The corpus used in our experiments is downloadedfrom an online Arabic newspaper. Its size is about 10 millionswords, distributed over six selected topics, in this case: culture,religion, economy, local news, international news and sports