Using a pre-trained Network to recognize the “Kan group”

Abstract

One of the most important fields of artificial intelligence is processing the Arabic language; its goal is to enable computers to understand and analyze texts written in human language. The arrival of pre-trained neural networks has made it feasible to enhance the precision of Arabic text analysis while identifying "kāna" and its sisters in sentences with greater efficiency. Therefore, in this work a framework for obtaining a qualitative identification of the “kāna” and its sisters in Arabic texts through the use of the pre-trained neural networks (DNNs) using the computed Zernike moments was developed. An extensive corpus of Arabic text examples that includes all instances of the use of "kāna" and its sisters in varying fonts and sizes was compiled. Arabic Zernike Moments were calculated to use as the base model using pre-trained DNNs. We first performed pre-training upon the collected dataset, then fine-tuning on the specific recognition task of “kāna” and its sister. Metrics such as accuracy, recall and F1 score are used to evaluate the models performance. The trained until now model detected well "kāna" and his sisters (those words behaved like "kāna") and got high accuracy on identifying these grammatical tools in Arabic texts. The results also reflect a good performance of the model in dealing with the variety of contexts in which kana and its sisters are found.