Sustainability-Aware Evaluating of Turkish Sentiment Analysis: Performance, Efficiency of BERTurk, DistilBERTurk, and ConvBERTurk on Humir

The performance gains of natural language processing models often come at the cost of high energy consumption and increased carbon footprint. Therefore, model evaluation should not only consider predictive accuracy but also resource usage and sustainability dimensions. This study compares pretrained Turkish language models—BERTurk, DistilBERTurk, and ConvBERTurk—for sentiment analysis through a sustainability-aware lens. The models are fine-tuned on the Humir-Sentiment dataset and evaluated using accuracy, precision, recall, F1, and Matthews Correlation Coefficient (MCC). Results show that BERTurk achieves the strongest predictive scores across accuracy, precision, recall, and F1, but at the cost of higher memory usage and longer runtimes, implying a larger estimated carbon impact. ConvBERTurk provides a balanced accuracy–efficiency profile, maintaining competitive performance while moderating resource demand. DistilBERTurk delivers the smallest memory footprint and lowest latency, translating to the lowest estimated Carbon Dioxide (CO₂) and making it well suited for resource-constrained or carbon-budgeted deployments. Overall, the findings indicate that Turkish morphological richness can be effectively modeled by modern pretrained architectures, and the proposed evaluation framework supports sustainable model selection by jointly considering accuracy, resource consumption, and carbon impact.