Fariya Tabassum | Searcharxiv

arXiv2026

Beyond Benchmarks: Exposing the Hidden Crisis in Bangla Hate Speech Detection

The spread of hate speech (HS) across different social media platforms (SMPs) poses a major concern for online safety and ethical moderation. Automatic detection of HS remains a challenging task, especially in under-resourced languages like Bangla, due to cultural context, implicit expressions, and informal linguistic patterns. This study aimed to expose the crisis of Bangla HS detection systems by diagnosing how and why benchmark-trained models fail to identify implicit, context-dependent HS. Six architectures (FastText + CNN, FastText + LSTM, FastText + BiLSTM, BanglaBERT, BanglaBERT + CNN, and BanglaBERT + BiLSTM) were trained on benchmark datasets (about 75,000 posts) and a merged multi-source dataset (about 120,000 posts), then externally validated on an annotated dataset (about 200 posts) collected from Facebook, Twitter, and YouTube, labeled as HS and non-HS, where HS was further categorized as explicit and implicit. BanglaBERT achieved an F1-score of 91.4% on benchmark datasets but declined to 75.3% on the external set and 63.4% for implicit HS involving sarcasm and emojis. The accuracy of FastText + CNN dropped from 78.0% to 51.2% under similar conditions. Emoji-aware preprocessing improved implicit HS detection by up to 12%, whereas emoji removal caused a notable decline in performance (F1: 0.75 to 0.63). Frequent misclassifications in politically charged or satirical comments revealed over-policing risks. This study not only exposes the generalization crisis due to implicit, culturally embedded, and emoji-laden expressions but also underscores the need for developing adaptive, emoji-aware, and culturally grounded frameworks that ensure ethical moderation while preserving freedom of expression. Findings of this study provide insights for researchers, SMPs, and policymakers to design more context-sensitive HS detection systems for low-resource languages.

cs.CL↗

arXiv2026

PIE-ADA: Physics-Informed Ensemble with Adaptive Data Augmentation for Photometric Transient Classification

The upcoming Large Synoptic Survey Telescope (LSST) is expected to observe approximately 10 million astronomical transient events per night, creating an urgent need for automated classification systems. A key challenge is the extreme class imbalance in transient datasets, where rare event types represent less than 1% of all observations. This paper presents PIE-ADA (Physics-Informed Ensemble with Adaptive Data Augmentation), a framework that generates physically realistic synthetic light curves for underrepresented classes using astrophysically motivated transformations. PIE-ADA applies four augmentation operations, namely correlated noise injection, cosmological time dilation, wavelength-dependent dust extinction, and observation phase shifting, while enforcing physical constraints to prevent unrealistic samples. We extract 271 multi-scale features from six photometric passbands covering statistical, temporal, peak, color, and frequency-domain properties. Evaluated on the PLAsTiCC dataset (7,848 original objects augmented to 8,148 across 14 classes), five classifiers were compared using stratified 5-fold cross-validation. LightGBM achieved the best performance with a weighted log loss of 0.5763 ($\pm$0.0083) and 80.33% accuracy, improving over Random Forest, Extra Trees, and Neural Network baselines by 24-49% in log loss. The framework is computationally efficient, completing the full pipeline in under 37 minutes and classifying individual objects in less than 0.05 seconds, making it suitable for real-time LSST alert processing.

astro-ph.IM↗