تشخيص و تصحيح خودكار خطاهاي دستوري رايج در زبان فارسي با استفاده از يادگيري عميق

Nowadays, the use of handwritten letters or texts for daily tasks has significantly decreased and most people prefer to write and edit their texts electronically. However, due to lack of time, lack of attention while writing, or even lack of knowledge, written texts may suffer from grammatical errors. Even those who are fully proficient in their native language can make mistakes. These mistakes can reduce the quality of the texts and prevent the correct transmission of the meaning of the text. Therefore, adhering to grammar is considered essential for text writers. Grammatical error detection and correction systems can do this with less time and more accuracy. In the Persian language, this work has often been done with rule-based methods. But with the advancement of artificial intelligence and the proof of its ability to understand natural language, it is possible to provide solutions based on this technology to achieve better results in detecting and correcting grammatical errors. This research has presented a method based on deep learning for detecting Persian grammar error types and correcting them. In this method, the type of error in the text is detected first using the ParseBERT model, and then it uses appropriate trained models to correct errors. To enhance performance, the proposed system first detects the type of error in the text and then focuses on the natural language understanding component during the grammatical error correction phase and converts the input text into a sequence of tags as edit operators using sequence-to-edit methods and finally generates the output sentence through post-processing. Additionally, due to the lack of an appropriate dataset, this research has created a dataset for detecting and correcting grammatical errors in the Persian language. For this purpose, sentences were initially collected from the Internet, and artificial errors were introduced into grammatically correct sentences to be used in training the model. Based on the eva‎luation of this model, F1 score for error detection and F0.5 for error correction were obtained as 93 and 87.94, respectively. By considering that this approach is the first approach based on deep learning in Persian language to detect and correct grammar errors, the results are acceptable.

6