A Statistical Approach for the Best Deep Neural Network Configuration for Arabic Language Processing

The widespread of the computer technology and the Internet lead to a massive amount of textual information being available in written Arabic. This that more is available, it becomes more difficult to extract the relevant information. To meet this challenge, many researchers are directed to the development of information retrieval systems based on syntactic and semantic parsing. In Arabic, this field is restricted by the lack of labeled datasets. Thus, it is important to build systems for part-of-speech tagging and language modeling and use their results for further syntactic and semantic parsing in fields like chunking, semantic role labeling, information extraction, named entity recognition and statistical machine translation. Deep neural networks have proved efficient in fields like imaging or acoustics and recently in natural language processing. In this study, we used the Taguchi method to find the optimal parameter combination for a deep neural network architecture. Therefore, the neural network obtained the most accurate results. The main use of the Taguchi method in our work is to help us to choose the best context which is the number of words before and after the word on which the training is made.