A Framework for Classical Arabic Parsing Using Deep Learning

Abstract

Arabic language is one of the most challenging languages for parsing because of the rich morphology and free word order system and the depth of the syntactical structure. This challenge increases with Classical Arabic, which is the old form of the Arabic language with more complex syntax. Moreover, Classical Arabic suffers from a lack of resources for parsing such as Treebank. Accordingly, this thesis has to main objectives 1) to the extending of the only Treebank of Classical Arabic; 2) and to develop a new framework for parsing. To achieve the first objective, the three layers within Quranic Treebank have been expanded. Various new algorithms have been introduced to achieve layers expansion. In the orthographical layer, a new script has been added that fits all the Classical Arabic texts instead of the existing script. Instead of the current numbering based on the concept of division of the Quranic Treebank, a new sequential numbering of sentences was added. In the morphological layer, the annotation scheme was moved towards more fine-grained tags by adding new tags and dividing some existing tags. Besides, approximately 0.73% of the morphological annotation was corrected, thus moving to a higher level in the gold-standard morphological annotation. In the syntactical layer, the entire layer was built according to a hybrid representation of the phrase-structure and dependency relations. To formalize this hybrid representation, the CoNLL-U format has been expanded. To achieve the second goal, this thesis presented a deep model for Classical Arabic parsing contains carefully selected standard techniques based on the best existing literature. The pipeline model is organized into three lines. In the pre-processing line, word embedding is pre-trained to achieve dense features representation. Also, the elliptical and phrase construction has been preprocessed to achieve a pure representation of sentences, the purpose of this is to make training data standard for dependency parsing. In this regard, the thesis provides new resources that have been introduced to Classical Arabic, the largest corpora for NLP tasks, and two models (Skip-gram and CWOB) for word embedding. In deep parsing lines, simple and effective parser features are extracted by BiLSTM training along with the parser algorithm. Syntactic units are represented in sequence with their contexts by using Deep-BiLSTM vectors. Propagating errors are reduced through Oracle Dynamic and an aggressive-exploration scheme. We introduce a new transition system that allows direct syntax information to be used in transition decisions. In the pre-processing line, a deep visual representation was presented that reflected the complexities of the Classical Arabic structure. To remove the conflicts within the grammatical structure of the sentence, a semantic dimension describing syntax has been added. Finally, we demonstrate the effectiveness of the pipeline parsing model by comparing it with previous work. The proposed framework has proven to be the best in all different performance measures.