Persian Ezafe Recognition Using Neural Approaches

Persian Ezafe Recognition aims to automatically identify the occurrences of Ezafe (short vowel /e/) which should be pronounced but usually is not orthographically represented. This task is similar to the task of diacritization and vowel restoration in Arabic. Ezafe recognition can be used in spelling disambiguation in Text to Speech Systems (TTS) and various other language processing tasks such as syntactic parsing and semantic role labeling.
In this paper, we propose two neural approaches for the automatic recognition of Ezafe markers in Persian texts. We have tackled the Ezafe recognition task by using a Neural Sequence Labeling method and a Neural Machine Translation (NMT) approach as well. Some syntactic features are proposed to be exploited in the neural models. We have used various combinations of lexical features such as word forms, Part of Speech Tags, and ending letter of the words to be applied to the models. These features were statistically derived using a large annotated Persian text corpus and were optimized by a forward selection method.
In order to evaluate the performance of our approaches, we examined nine baseline models including state-of-the-art approaches for recognition of Ezafe markers in Persian text. Our experiments on Persian Ezafe recognition based on neural approaches employing some optimized features into the models show that they can drastically improve the results of the baselines. They can also achieve better results than the Conditional Random Field method as the best-performing baseline. On the other hand, although the results of the NMT approach show a better performance compared to other baseline approaches, it cannot achieve better performance than the Neural Sequence Labeling method. The best achieved F1-measure based on neural sequence labeling is 96.29%