Robust Bangla Sarcasm Detection with HFAT-v4: Multi-Seed Evaluation, Ensemble Learning and Supporting English Benchmark Analysis

Automatic sarcasm detection remains challenging in low-resource
languages because pragmatic intent is often implicit, context dependent and
weakly represented by surface-level cues. Existing Bangla studies have
largely adopted binary formulations and have reported limited evidence on
stochastic stability, class-specific reliability, confidence calibration and
systematic errors. This study introduces HFAT-v4, a BanglaBERT-based
Hybrid Fourier-Attention Transformer for three-class classification of
Neutral, Non-Sarcastic and Sarcastic text. HFAT-v4 combines scalar-mixed
contextual representations with self-attention, adaptive Fourier token mixing,
a spectral residual pathway, multi-representation pooling, hierarchical
supervision and R-Drop regularization. Following normalization, conflicting
duplicate removal and same-label deduplication, 11,911 Bangla instances
were stratified into 8,337 training, 1,787 validation and 1,787 test samples.
Performance was evaluated across five random seeds and three ensemble
strategies using classification, ranking, calibration and error-analysis
measures. The five-seed mean accuracy and macro-F1 were 0.6384 ± 0.0047
and 0.6386 ± 0.0038, respectively. Majority voting achieved the strongest
hard-label performance, with 0.6575 accuracy and 0.6584 macro-F1, whereas
probability averaging achieved the best macro-ROC-AUC (0.8254), macro
average precision (0.7068) and expected calibration error (0.0795). Sarcastic
text was recognized most reliably, while non-sarcastic text remained the
principal bottleneck; 376 test instances were misclassified by all five models,
indicating systematic ambiguity beyond seed-specific variation. A
complementary English Reddit analysis also showed gains from richer n
grams and subreddit context. These findings demonstrate the value of multi
seed and calibration-aware evaluation while identifying contextual ambiguity
as the primary target for further improvement.