ACCURATE AND FAST RECURRENT NEURAL NETWORK SOLUTION FOR THE AUTOMATIC DIACRITIZATION OF ARABIC TEXT

Abstract

Arabic is mostly written now without its diacritics (short vowels). Adding these diacritics decreases reading ambiguity among other benefits. This work aims to develop a fast and accurate machine learning solution to diacritize Arabic text automatically. This paper uses long short-term memory (LSTM) recurrent neural networks to diacritize Arabic text. Intensive experiments are performed to evaluate proposed alternative design and data encoding options towards a fast and accurate solution. Our experiments involve investigating and handling problems in sequence lengths, proposing and evaluating alternative encodings of the diacritized output sequences and tuning and evaluating neural network options including architecture, network size and hyper-parameters. This paper recommends a solution that can be fast trained on a large dataset and uses four bidirectional LSTM layers to predict the diacritics of the input sequence of Arabic letters. This solution achieves a diacritization error rate of 2.46% on the LDC ATB3 dataset benchmark and 1.97% on the larger new Tashkeela dataset. This latter rate is 47% improvement over the best-published previous result.