| This paper presents a comparative evaluation of LLMs for Sentiment Analysis in English and Greek. Utilizing a balanced Kaggle dataset of 1950 X comments, labeled with the sentiments "positive", "negative" or "neutral", we created a Zero-Shot and a Few-Shot strategy and used it across GPT-4.1 mini, GPT-4o mini, and GPT-5 mini. Results indicate that the Few-Shot strategy consistently improves performance, scoring a 13.1% increase for English and 11.7% for Greek. Deviating from the field’s standard, the Greek Few-Shot setup outperformed the English baseline (peak Macro-F1: 0.688), a phenomenon we attribute to the morphological richness of the Greek Language. This gap is mostly driven by an "Optimism Bias", where English Models frequently misclassify "neutral" sentiment as "positive". To our knowledge, this is the first comparative study of the GPT-4 and GPT-5 models on paired English–Greek sentiment data that jointly examines zero-shot vs few-shot prompting and rich qualitative phenomena such as sarcasm, code-switching (triggering a 4.99% accuracy drop), and the impact of sentence length on model resilience. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.