Using Computational Methods for Persian Collocations Identification and Extraction

In this article we study collocations in Persian. Previous researches in this field have been mostly statistical and comparative. The purpose of this research is to identify collocations using a corpus-based and computational method. In this research, after reviewing the definition of a collocation given by Iranian and non-Iranian linguists, researches conducted by Iranians or non-Iranians in this field are presented. In this paper, the Persian language database is used as the corpus. Also, as no dictionary of Persian collocations exist, a dataset of collocations has been compiled based on the Advanced Learners' Persian Dictionary. Using FastText embedded vectors, the language model is trained with a Long Short-Term Memory network. The results are then evaluated using several methods. Also, by fine-tuning ParsBert, the call of this language model is calculated using a thousand-item lists of collocations and non-collocations. Finally, a comparative analysis of collocation translation in Google Translate is conducted by translating a thousand Persian sentences into English. The following results are obtained from the examination of collocations in the language model trained with the Long Short-Term Memory network and ParsBert: in both models, collocations can be predicted, but ParsBert proved a stronger model in investigating language problems such as collocation examination. In the comparative analysis of the accuracy of Google Translate's collocation translation, three results were obtained: (1) the translation was correct; (2) the translation was literal and word for word; (3) The translation of collocations was ignored.