Effect of artificial intelligence translation tools for sentiment analysis in Turkish texts

In sentiment analysis, it is often difficult to determine the most efficient model for a particular task. Criteria such as time, cost and performance play a decisive role in model selection. Therefore, it is important to compare the performances of the models. Some of the pre-trained language models were trained only in English and translation is required to perform sentiment analysis on Turkish texts with these models. Since correct performances cannot be achieved in models trained in English with incorrect or incomplete translations, the translation phase is very important. In this study, the English translation performances of GPT, Google Translate and Amazon Translate services were observed. Cyclic translation operations were applied to the texts in the form of Turkish→English→Turkish. Vector representations of the original and translated Turkish texts were obtained, and the Levenshtein distance, Cosine Similarity and semantic similarity rates between them were obtained. The averages of these values and their performances were calculated and criteria such as time and cost were taken into account. Google Translate was used as the translation method of the data sets for the performance evaluation of English models due to its high semantic similarity rate, free and fast. In the sentiment analysis model performance evaluations conducted in Turkish and English with the data set labeled as Positive, Negative or Neutral, GPT and Amazon Comprehend services, which can be accessed via online API, fell behind the pre-trained models Google BERT and Facebook RoBERTa. It was concluded that the most successful model was the Google BERT model trained for Turkish. Well-trained Turkish models are more successful than English models.