AIN SHAMS UNIVERSITY               
                  Faculty of Computer  
                 & Information Sciences          
              Computer Science Department 
                                                 
                      AN INTELLIGENT SYSTEM FOR 
                         AUTOMATED ARABIC TEXT 
                                 CATEGORIZATION 
                                                 
                           A Thesis Submitted to Computer Science Department,  
                      Faculty of Computer & Information Sciences, Ain Shams University 
                     In partial fulfillment of the requirements for Master of Science Degree 
                                                 
                                               By 
                                                 
                                   Mena Badieh Habib 
                                   B.Sc. in Computer Science, 2002. 
                              Demonstrator, Computer Science Department, 
                              Faculty of Computer & Information Sciences,  
                                  Ain Shams University, Cairo, Egypt. 
                                                 
                                      Under Supervision of 
                                                 
                         Prof. Dr. Mostafa Mahmoud Syiam 
                                    Professor of Computer Science, 
                                    Computer Science Department, 
                               Faculty of Computer & Information Sciences,  
                                  Ain shams University, Cairo, Egypt. 
                                                 
                                   Dr. Zaki Taha Fayed 
                                Associate Professor of Computer Science, 
                                    Computer Science Department, 
                              Faculty of Computer & Information Sciences, 
                                  Ain shams University, Cairo, Egypt. 
                                                 
                                Dr. Tarek Fouad Gharib 
                               Associate Professor of Information Systems, 
                                   Information Systems Department, 
                               Faculty of Computer & Information Sciences,  
                                  Ain shams University, Cairo, Egypt. 
                                                 
                                                 
                                              2008 
                      Acknowledgements 
           
           First  and  foremost,  I  could  never  forget  the  late  Prof  Dr.  Mosatafa 
          Syiam who walked with me on the first steps with this work. I dedicate 
          this work to his soul. 
            
           I would like to express my sincere gratitude to my chief supervisor Dr. 
          Tarek Gharib from whom I have learned a lot, due to his supervision, 
          guidance, support and advising till this work come to light.  
            
           I  would like to thank Dr. Zaki Taha, for his valuable scientific and 
          technical notes. 
            
           Also I would like to express my gratitude to Prof Dr. Abdel-Badeeh 
          Salem the head of computer Science department who gave me the basic 
          idea of this thesis and helped me with his great experience. 
            
           My great thanks also go Prof Dr. Essam Khalifa and Prof Dr. Said 
          Ghoniemy for their encouragement. 
            
           Finally, my deepest thanks go to my parents for their unconditional 
          love, and to my friends for their support. 
            
           This  thesis  would  have  been  much  different  (or  would  not  exist) 
          without these people. 
            
                                                Mena
                                
                                
                              ii 
           
                       Abstract 
           
          New technological developments have resulted in a dramatic increase 
        in  the  availability  of  on-line  text-newspaper  articles,  incoming 
        (electronic) mail, technical reports, etc. This led to the need for methods 
        that help users organize such information. Text Categorization may be the 
        solution  for  the  increased  need  for  advanced  techniques.  Text 
        Categorization is the classification of units of natural language texts with 
        respect to a set of predefined categories. Categorization of documents is 
        challenging, as the number of discriminating words can be very large. 
        Machine  learning  approaches  are  applied  to  build  an  automatic  text 
        classifier by learning from a set of previously classified documents. 
           
          Few researches have tackled the area of Arabic text categorization till 
        the time we start working on this research. Arabic language is a Semitic 
        language that has a complex and much morphology than English. It needs 
        a  set  of  preprocessing  routines  to  be  suitable  for  manipulation.  Stop 
        words like prepositions and particles are considered insignificant words 
        and must be removed; Words must be stemmed after stop words removal. 
        Stemming is  the  process  of  removing  the  affixes  from  the  word  and 
        extracting  the  word  root.  After  applying  preprocessing  routines, 
        document is represented as  a  weighted  vector.  Representation  process 
        consists of two phases: 
          a)  Term  selection  which  can  be  seen  as  a  form  of  dimensionality 
        reduction by selecting a subset of terms from the full original set of terms 
        according to some criteria,  
          b) Term weighting in which, for every term selected in phase (a) and 
        for every document, a weight is computed which represents how much 
        this term contributes to the discriminative semantics of the document. 
                          iii 
         
           
          Finally, the classifier is constructed by learning the characteristics of 
        every category from a training set of documents, and tested by applying it 
        to the test set and checking the degree of correspondence between the 
        decisions of the classifier and those encoded in the corpus. 
           
          This thesis presents an intelligent Arabic text categorization system. 
        Experimental results performed on a text collection of 1132 document 
        collected from the local newspapers show that using light stemming along 
        with  trigram  stemmer  is  the  most  appropriate  stemming  approach  for 
        Arabic  language.  The  main  problem  with  the  traditional  methods  of 
        feature selection is founding a large set of sparse documents (most of the 
        documents does not contain any term in the list of the selected terms). To 
        solve this problem we removed words that rarely appear in the documents 
        before using information gain, this gives better results. Also we combined 
        global  and  local  feature  selection  to  reduce  the  number  of  empty 
        documents  without  affecting  the  performance.  Normalized  term 
        frequency inverse document frequency (normalized-tfidf) was the most 
        suitable weighting criteria for representing the documents as a vector of 
        the  set  of  selected  terms  (words).  Finally  after  testing  four  famous 
        classifiers, it has been shown that Rocchio classifier performs better when 
        the  number  of  terms  is  small  while  Support  Vector  Machines  (SVM) 
        outperforms the other classifiers when the number of is large enough. 
        Classification accuracy exceeds 90% when using over than 4500 feature 
        to represent documents. 
                          iv