BERT Transformers Performance Comparison for Sentiment Analysis: A Case Study in Spanish

Despite the fact that the Spanish language is the second most spoken language in the world, research about AI and sentiment analysis is few compared with English language as research target. This paper refers to some works about sentiment analysis for the Spanish language that use BERT transformers and other technologies for this kind of sentiment analysis task; but the quality model based on indicators such as accuracy level could be different according to the tool or BERT version used. In addition, about the BERT family, it is challenging to determine which versions or subversions could perform better for sentiment analysis and also comply with the Spanish language when they are used on common platforms such as Colab. Therefore, the present study seeks to address this issue by establishing objectives, such as identifying relevant datasets based on the quality of Spanish used and having balanced subsets; also, locating different Spanish trained models; and proposing a method of comparison that involves relevant variables. We propose a weighted index that combines the F1-Score and the retraining time in different scenarios to help making better decisions. The results of this study indicate that the DistilBERT, RoBERTa, and ALBERT models have highest performances, but BERT remains in top positions as a consistent model. ©2024 Springer.