Language-Specific Sentiment Polarity Biases in Encoder and Large Language Model Classification of Product Reviews

This study examines sentiment polarity biases in large language models (LLMs) and encoder models when classifying product reviews across different languages. It finds that LLMs exhibit a negative bias in French, being more accurate with negative reviews, while encoder models show a positive bias in Japanese, struggling with negative reviews that employ indirect criticism. These findings highlight critical considerations for practitioners developing multilingual sentiment analysis systems, as they reveal inherent biases that could affect the accuracy and reliability of sentiment classification in diverse linguistic contexts.