A Recurrent Neural Network-Based Approach for Word-Level Diacritization of Arabic Text
SALIH TAQIUDDIN · Research Publication Repository of King Fahd University of Petroleum and Minerals (King Fahd University of Petroleum and Minerals) · 2020
Automatic Arabic text diacritization (AATD) is an active and important field of research.It has numerous applications including text-to-speech systems and as an aid for learners of Arabic.Although it has attracted a lot of research in the last three decades, there is still need for improvement.Currently, the deep learning area of machine learning field is enjoying great success and spread in a wide range of applications.More specifically, the recurrent neural network architecture is regaining high interest and achieving state-of-the-art published results in various research fields.Research employing deep learning approaches to the AATD problem has already been conducted in multiple publications.However, most of these works built their approaches on a character-level basis.This thesis aims to explore a deep learning approach to the problem that is developed on a pure word-level basis.Its contributions are (1) developing a formal definition of the research problem that is independent of solution approaches, and (2) assembling a coherent framework for the research problem, and (3) introducing a new metric for evaluating the performance of a diacritization system or approach, and (4) the automatic detection and processing of multiple cases of Arabic declension.The results achieved in this work are 16.70% diacritic error rate, 16.38% character error rate, and 38.53% word error rate.This means that around 84% of the text's characters are correctly diacritized.xxi