Persian Sentiment Analysis via a Transformer Model concerning Banking Sector

Abstract

The competitive landscape of a country's banking sector necessitates an in-depth understanding of customer satisfaction levels concerning the services provided. Presently, customers predominantly express their feedback via social media platforms in the form of posts and comments. This study endeavors to create a highly accurate sentiment detection algorithm for the Iranian banking system, utilizing a transformer model. In the initial stages, we collected data by crawling comments from Twitter, which are subsequently labeled and filtered according to the names of Iranian banks, dating from 2019. Following this, an optimized Deep Neural Network (DNN)-based pre-trained ParsBERT model, a monolingual Persian model, is fine-tuned using this data. Finally, our model is evaluated on a test dataset, and the results are validated by comparing them with the original multilingual BERT, Bidirectional Long Short-Term Memory (Bi-LSTM) network, and four other classification methods. To address the Out-Of-Vocabulary (OOV) issue, a character-level embedding is incorporated in conjunction with the word-level embedding. This approach aids in tackling the multitude of variations observed in non-native words, extracting character-level features using a character-level Bi-LSTM. The proposed model highlights the statistical superiority of our method when compared to the other methods evaluated.