The Effect of Data Augmentation Techniques on Persian Sentiment Analysis
Ali Nazarizadeh, Minoo Sayyadpour, Touraj Banirostam · 2023
The growth of the Internet and increased use of social media have led to the production of a large amount of unstructured data. These data have many variations, such as text, image, video, and audio. Among them, textual data has the largest share. Considerably textual data comprises the viewpoints of users regarding a wide range of services and products. In order to achieve high accuracy in sentiment classification, Data Augmentation techniques have been used in this paper, which has significantly increased the volume and variety of data. Seventeen different Deep learning models have been trained on the final dataset. The results show that using the Data Augmentation approach has increased the accuracy by more than 6% in Persian sentiment analysis. Methods used in this paper include translating sentences from Farsi to English and vice versa, sentence displacement, replacing synonyms, and data balancing. Also, in this paper, two new Lexicons have been manually created for Persian sentiment analysis. The first is a sentiment lexicon containing 660 positive and negative words, and the second is a lexicon of 390 stop words. Finally, the accuracy obtained using the Data Augmentation approach has reached 96.73%, which is sufficient compared to previous research.