PARSING ARABIC TEXTS USING REAL PATTERNS OF SYNTACTIC TREES

摘要

In order to parse Arabic texts, we have chosen to use a machine learning approach. It learns from an Arabic Treebank. The knowledge enclosed in this Treebank is structured as patterns of syntactic trees. These patterns are representative models of the Arabic syntactic components. They are both layered and rich structurally and contextually. They serve as an informational source for guiding the parsing process. Our parser is progressive since it proceeded by treating a sentence into a number of stages equal to the number of its words. At every step, the parser affects the target word with the most likely patterns that represent it in the context where it is put. Then, it joins the selected patterns with those collected in the previous parsing steps in order to construct the representative syntactic tree(s) of the whole sentence. If more than one tree is proposed, all the analysis trees are sorted according to their appearance frequencies in the Treebank. The preliminary tests have yielded accuracy and f-score equal to 84.8% and 77.5%, respectively.