4arXiv cs.CL (Computation and Language)·
Study exposes generalization failures in Bangla hate speech detection models
A new arXiv paper evaluates six NLP architectures for Bangla hate speech detection, finding severe performance degradation when models trained on benchmark datasets are applied to real-world social media content. BanglaBERT drops from 91.4% F1 on benchmarks to 63.4% on implicit hate speech involving sarcasm and emojis. The study highlights a broader generalization crisis for low-resource language moderation systems and finds that emoji-aware preprocessing recovers up to 12% of lost performance.
Related guides (2)
Related events (8)
Adversarial robustness evaluation of Arabic language models across attack granularities
A new arXiv preprint evaluates five Arabic language models (including MARBERT and AraBERT) against adversarial attacks at character, word, and sentence levels, finding that diacritic insertion can reduce model accuracy by up to 92% and paraphrasing attacks degrade performance by an average of 76%. The study also tests adversarial training as a defense, finding MARBERT most robust and AraBERT showing the greatest relative gains. Results highlight particular challenges for morphologically rich languages where character-level noise remains difficult to defend against.
Tokenizer transplantation fixes autoregressive collapse in Bengali edge ASR
Researchers identify that English-centric byte-level tokenizers cause catastrophic autoregressive collapse when lightweight ASR models like Moonshine are applied to morphologically rich languages like Bengali. They propose a vocabulary transplantation pipeline that swaps in BanglaBERT's WordPiece vocabulary, reducing token fertility from 9.16 to 1.30 and cutting autoregressive sequence length by 85.8%. The modified model achieves 21.54% WER and an RTF of 0.0053 on the 882-hour Lipi-Ghor dataset, offering a reproducible blueprint for cross-script adaptation of compact ASR models without retraining from scratch.
Multilingual study finds LLMs are not uniformly robust to non-canonical tokenizations, with up to 23.7% performance drops
A new arXiv paper investigates how language models behave when given alternative (non-canonical) tokenizations of the same input string across 27 languages and six downstream tasks. While prior work showed English models are largely invariant to such perturbations, the study finds this does not generalize: Llama-3.1-8B drops 23.7% on average, Qwen3-8B 11.4%, and Gemma-3-12B 9.9% in relative performance. Languages with higher token fragmentation are systematically more sensitive, and the authors show LoRA fine-tuning on multi-tokenization data—including English-only data—provides meaningful mitigation.
LLMs fail to consistently simulate demographic perspective-taking in hate speech annotation
A new arXiv paper evaluates whether persona-conditioned LLMs can replicate how different demographic groups perceive hate speech, testing three dimensions: inter-group disagreement, in-group sensitivity, and vicarious prediction. No model consistently captures all three dimensions, and performance is highly model-dependent rather than emerging reliably from identity prompts alone. Vicarious prompting with Llama 3.1 provides the closest approximation to human disagreement patterns across demographic axes. The findings have implications for using LLMs as proxies for diverse human annotators in content moderation tasks.
Text Analytics Evaluation Framework: Benchmarking LLMs on Social Media NLP Tasks
Researchers introduce a 470-question evaluation framework to assess LLM performance on aggregated social media text, applied to Twitter datasets across sentiment analysis, hate speech detection, and emotion recognition. Results show performance degrades substantially as input scale exceeds 500 instances, particularly for open-weights models on numerical tasks. Multi-label and target-dependent scenarios also show notable performance drops, and task complexity progressively erodes accuracy from basic semantic identification to comparison and counting operations. The findings point to architectural bottlenecks in current LLMs for rigorous quantitative analysis over large text collections.
When Does Demographic Information Help? Data and Modeling Regimes for Perspective-Aware Hate Speech Detection
This paper investigates when demographic features improve hate speech detection models that account for annotator perspectives, finding that gains are not universal but depend on specific data and modeling conditions. The authors identify that demographic information helps most in regimes with low training disagreement, high test disagreement, sufficient training data, and strong demographic overlap between train and test sets. They introduce a gated demographic residual model that selectively applies demographic adjustments to text-only predictions, demonstrating effectiveness on high-disagreement and low-confidence examples using the MHS and POPQUORN datasets. The work cautions against assuming demographic features are universally beneficial in subjective NLP tasks.
Disagreeing Rationales: Rethinking Classification and Explainability Evaluation in Hate Speech Detection
This paper investigates human disagreement in token-level rationale annotations for hate speech detection, a dimension less studied than label disagreement. The authors unify diverse models, training strategies, loss functions, and evaluation metrics under a single protocol, systematically comparing hard and soft label/rationale representation spaces. Results show that both hard and soft metrics favor softer representations, suggesting that soft supervision better captures human reasoning variation in subjective NLP tasks. The work calls for rethinking evaluation frameworks for classification and explainability in subjective NLP.
RDS Fusion: Hybrid neuro-symbolic gating with compressed CoT for zero-shot irony detection
Researchers introduce the Robust Dual-Signal (RDS) Fusion framework, a hybrid neuro-symbolic architecture that compresses Chain-of-Thought reasoning without supervised fine-tuning for irony and sarcasm detection in social media text. Evaluated on TweetEval (N=734) and iSarcasm, the zero-shot system matches fine-tuned BERTweet performance and outperforms supervised SemEval transformer ensembles on the imbalanced iSarcasm dataset. A statistical ablation shows that only the full concurrent fusion of all three signals yields a validated improvement, with individual components providing no significant standalone gain.