Cyclostationary Signal Modeling Enhances Multilingual BERT Sentiment Analysis Accuracy

Modeling text embeddings as potentially cyclostationary sequences improves BERT-based multilingual sentiment analysis accuracy by up to 28% relative to baseline performance without modifying the underlying model architecture, according to a study by Nazila Panahi and Mehdi Chehel Amirani published in the International Journal of Data Science and Analytics. The authors investigated text representations from a signal-processing perspective to address performance challenges across languages that arise from linguistic and cultural variation, aiming to identify polarity, emotions, and opinions expressed in text.

By formulating text embeddings as cyclostationary sequences, the framework enables the characterization of hidden periodic and statistical patterns within the frequency domain to extract informative features for sentiment prediction. The researchers integrated this cyclostationarity-based analysis with dimensionality-reduction techniques, specifically Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA), and subsequently applied the extracted features to downstream classification and regression models.

The framework was tested on five benchmark datasets spanning four languages—English, Arabic, Brazilian Portuguese, and Turkish—alongside a mixed Arabic–English corpus, demonstrating accuracy gains of up to 28% compared to the baseline. In addition, the framework was evaluated against two contemporary Large Language Models, DeepSeek and Gemini, under a consistent experimental setting, indicating that cyclostationarity-informed feature extraction enhances sentiment analysis across linguistically diverse datasets.

Subject of Research: Signal processing methods for multilingual sentiment analysis in natural language processing

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

Article References: Panahi, N., & Chehel Amirani, M. (2026). A new framework for multilingual sentiment analysis in natural language processing using a pre-trained BERT model supported by multi-class machine learning methods. International Journal of Data Science and Analytics, 22(1), Article 374. https://doi.org/10.1007/s41060-026-01287-3

Image Credits: AI Generated

DOI: 10.1007/s41060-026-01287-3

Keywords: sentiment analysis, natural language processing, BERT, cyclostationarity, signal processing, multilingual sentiment analysis, dimensionality reduction, machine learning, framework, multilingual, sentiment, analysis