A new framework for multilingual sentiment analysis in natural language processing using a pre-trained BERT model supported by multi-class machine learning methods

Sentiment analysis, a central application of Natural Language Processing (NLP), identifies emotions and opinions in text. Accurate interpretation requires accounting for linguistic and cultural variations across languages. Considering text as a signal and given the cyclostationarity of most human-constructed signals, we sought to enhance sentiment analysis of human-written notes by extracting the most informative features, following our previous studies. From a signal-processing perspective, we examine the cyclostationarity of text embeddings to reveal hidden periodic patterns and describe their statistical properties in the frequency domain. Our approach combines advanced feature extraction and enhancement with classification and regression models to improve BERT’s performance without modifying its architecture. Experiments on five benchmark datasets in four languages—English, Arabic, Brazilian Portuguese, and Turkish—and a mixed Arabic-English corpus showed up to a 28% accuracy improvement. Results confirm that human-generated text exhibits cyclostationary properties that can be leveraged for more effective sentiment modeling. This research represents the first comprehensive study of multilingual alignment in ML- and DL-based NLP sentiment analysis and demonstrates how signal-processing-based features can increase model robustness and cross-lingual accuracy.