Enhancing Efficiency and Transparency in Transformer-Based Emotion Classification for Indonesian Tweets

Abstract

The growing use of social media in Indonesia has led to an abundance of emotionally rich content, offering valuable insights into public sentiment while also creating challenges for computational analysis due to linguistic variability, informal expressions, and diverse writing styles. This study aims to evaluate and enhance emotion classification on Indonesian tweets through the lens of efficient and explainable AI. Four Transformer-based architectures, such as IndoBERTweet, IndoBERT-base, IndoBERT-large, and Multi-BERT, were compared to establish a baseline. Among these models, IndoBERT-base achieved the best baseline performance with an F1-macro score of 0.8386. To improve computational efficiency and reduce deployment cost, two compression techniques were explored, which is mixed-precision and unstructured pruning with varying sparsity ratios. Experimental results showed that the configuration combining FP16 mixed-precision and 50% pruning achieved the best trade-off, maintaining an F1-macro of 0.8536 while reducing model size by 50% and achieving a 5.3× inference speedup compared to the FP32 baseline. Finally, explainable AI techniques were applied to interpret model predictions and analyze misclassifications between the baseline and efficient models. Overall, the findings demonstrate that efficiency-oriented optimization can enhance performance without sacrificing the model interpretability and evaluation score, leading to a more transparent and deployable emotion analysis systems for Indonesian social media.