Lexicon–Context Divergence for Reliability-Aware Hate Speech Detection in Low-Resource Bangla Memes: A BanglaBERT Hybrid with Calibration and Selective Prediction Analysis

Hate speech detection in Bangla memes is complicated by a mismatch between surface lexical cues and contextual
meaning: profane vocabulary is often jocular, and clean vocabulary can carry hateful intent once sarcasm or implication is
read into it. We study this mismatch on a 5,126-meme Bangla
dataset (OCR text, English translation, and a 441-term offensive
lexicon) by defining a Lexicon–Context Divergence (LCD) score,
the absolute difference between a normalized lexicon-hit density
and a fine-tuned BanglaBERT model’s own predicted probability
of hate speech. We fuse LCD and lexicon features into a
BanglaBERT classifier and attach a learned abstention head
trained to predict its own error probability, evaluating against five
baselines (lexicon-only logistic regression, BanglaBERT text-only,
BanglaBERT with lexicon fusion, entropy-based abstention, and
XLM-RoBERTa). The lexicon-fusion baseline reaches the highest
point-estimate test F1 (0.807), with the full hybrid model close
behind (0.803, bootstrap ranges overlapping, paired-bootstrap
p = 0.767); the hybrid model obtains the lowest expected
calibration error of the four calibration-evaluated models (0.071
versus 0.104–0.118) and the highest ROC-AUC (0.927). On our
two research questions, results are mixed and reported honestly:
LCD does not significantly separate memes with unanimous
annotator agreement from memes with majority disagreement on
a 200-item triple-annotated subset (Mann-Whitney p = 0.219),
and the learned abstention head’s risk-coverage trade-off (AURC
= 0.232) is markedly worse than a simple entropy heuristic
(AURC = 0.053) at aggressive rejection rates, even though the
same head’s predicted error probability is a highly significant
separator of correct from incorrect predictions in aggregate
(p < 10−4
). This suggests a learned P(error) score can carry real
signal on average while still ranking items poorly at the extremes,
producing an unfavorable low-coverage operating curve; we
report the full ablation, calibration, and code-mixed robustness
analysis needed to reproduce and stress-test this distinction.