The Global and local attention for automatic Arabic text diacritization
Ali Mijlad, Yacine El Younoussi · International Journal of Engineering and Applied Physics · 2023
Automatic Arabic diacritization is the task to restore diacritic or vowel marks for a non-vocalized Arabic text. This task showed its importance in the natural language processing NLP field and it helps people with specific learning difficulties to access Arabic web content. To tackle the problem, we suggest a letter-based encoder-decoder that uses previous deep learning attention models known as Luong attention. The training of the models knew unstable loss. And, as was expected — from the proposed models — the model that uses local predictive attention achieved the best word and letter error rates. The best-achieved diacritic error rate in the test data is about 26.80%. Nevertheless, the models need improvements in future work.