Accuracy–Efficiency Trade-Offs in Deep Learning for Twitter Emotion Classification: A Hardware-Aware Benchmark of Transformer and Neural Models
Year: 2026
Type: article
Abstract: The widespread use of social media platforms such as Twitter enables large-scale, real-time analysis of public emotions, yet the short and noisy nature of tweets continues to challenge emotion classification systems. While much prior work emphasizes binary or ternary sentiment prediction, controlled benchmarking for six-class emotion classification across diverse deep learning architectures remains limited, particularly regarding hardware efficiency under controlled experimental conditions. This study benchmarks five models—BERTweet, BERTweet + BiLSTM, MobileBERT, TextCNN, and BiLSTM + Conv1D—on a Kaggle dataset of 416,809 English tweets labeled with six emotions: sadness, joy, love, anger, fear, and surprise. All models were trained under a unified experimental protocol with stratified splitting and consistent preprocessing. Performance was evaluated using accuracy and macro F1-score, supported by confusion matrices, computational cost analysis, GPU memory profiling, and inference latency benchmarking. Multi-seed evaluation (seeds: 21, 42, 123) was conducted for the top two models to assess performance stability across random seeds. Results show that transformer-based architectures consistently outperform classical baselines: BERTweet achieved 96.84% accuracy and 0.9493 macro F1 with 2.9 GB GPU memory usage, while classical models (TextCNN, BiLSTM + Conv1D) trained 7– $11\times $ faster with less than 250 MB memory footprint and only 1–2% lower macro F1. Across the three evaluated seeds, no statistically significant difference was detected between BERTweet and BERTweet + BiLSTM (p > 0.68). Given the limited number of seeds, this result should be interpreted cautiously, although the hybrid model showed only marginal practical benefit relative to its additional computational cost. Limited cross-dataset evaluation on BERTweet provides preliminary evidence of generalization under distribution shift (93.03% accuracy). These findings provide a controlled comparative basis for model selection in Twitter emotion classification. BERTweet is suitable for applications prioritizing predictive performance, while classical models offer computationally efficient alternatives under constrained resource settings. Deployment-oriented conclusions should be interpreted together with the study’s emphasis on training-time cost, memory profiling, and inference latency benchmarking, which reveal that classical models offer substantially lower inference latency than transformer-based alternatives.
Source: IEEE Access
Authors: Sabah Dawod, Burcu Çarkli Yavuz
Institution: Sakarya University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,075
OpenAlex ID: W7155071989