arXiv:2607. 15861v1 Announce Type: cross Abstract: Moderation systems increasingly rely on external toxicity tools, but those tools are unreliable under code-mixing, transliteration, slang, and language mismatch.
The paper introduces BanglaSafe, a benchmark of 879 Bengali prompts that covers 17 culturally grounded harm categories and five prompting conditions. Evaluation of 18 frontier LLMs shows that 53.6% of responses are unsafe or partially unsafe, with 14.7% containing strictly harmful content. The study finds that the writing style within Bengali has a stronger impact on safety than the language switch itself, and that current safety classifiers struggle to reliably evaluate Bengali content.
arXiv:2610.15795v1 Announce Type: cross Abstract: Automated moderation systems are deployed at scale, yet they remain opaque to the users and moderators affected by their decisions, and their coverag...
Lipi-Ghor is a large-scale, weakly-supervised Bengali speech corpus containing 885 hours of audio from 998 YouTube videos across 17 domains, each annotated in SSTT format with speaker diarization and timestamps. It is the first Bengali corpus to combine extensive coverage with speaker-diarized, long-form audio annotations, and the authors provide a replicable pipeline for building similar resources in other languages. The dataset includes a human-annotated gold evaluation split and is released under CC‑BY‑4.0, with raw audio withheld.
arXiv:2604.10632v2 Announce Type: replace-cross Abstract: Music and food traditions are both intangible cultural heritage, and the links between them, how a sound can make a taste seem sweeter or mor...
arXiv:2607. 03981v1 Announce Type: cross Abstract: Memes have become influential communication tools on social media, combining viral visuals with concise messaging to convey impactful ideas.
Bengali, spoken by over 230 million people, remains severely under-resourced in spoken language research. Existing open Bengali ASR corpora are confined to short, read or crowd-sourced speech and lack...
arXiv:2503.16072v5 Announce Type: replace-cross Abstract: Toxicity detection has become core safety infrastructure for online moderation, dataset filtering, and deployed language-model systems. Yet m...
arXiv:2610.15350v1 Announce Type: cross Abstract: We present Gugak-VocalSet, the first publicly available audio dataset for Korean traditional singing with dense annotations of sigimsae, a structured...
BanglishDepYT is a native Bengali depression detection dataset comprising 191,000 unlabeled YouTube comments and 1,000 bilingual‑annotated samples. The authors introduce Emotion‑Preserving Hybrid preprocessing (EPH), which retains emoticons, Banglish tokens, negations, and emotional elongation, improving SVM F1 from 0.7062 to 0.8570. Domain‑adaptive pretraining of SahajBERT on in‑domain comments yields the best performance (F1 = 0.8161, MCC = 0.6272), and LLM‑assisted annotation enables pseudo‑labeling that further boosts AUC from 0.8499 to 0.8731.
The paper "Ghaib in Translation" investigates how large language models (LLMs) handle Urdu, a widely spoken language that is largely absent from safety evaluations. Five prominent LLMs—GPT‑4o, Claude Sonnet 4.5, Gemini 2.5 Flash, Qwen‑2.5, and Llama‑3.1—were tested on six datasets covering Nastaliq Urdu, Roman Urdu, English, and code‑switched Urdu‑English. The study found significant label instability between original‑script and English‑translation classifications, with missed‑in‑Urdu rates ranging from 2.4% to 9.9% (median 4.3%). A review of 205 papers across nine ALW/WOAH editions revealed no dedicated Urdu research, underscoring the language’s neglect in current safety research.
The paper adapts the LLM‑SC framework to extract structured information from visitor reviews of Pashupatinath Temple, Nepal, by adding eight sacred‑heritage entity types, four relation types, and nine aspect‑sentiment dimensions. Using GPT‑4o on 2,500 Tripadvisor reviews, the authors produced over 11,000 entity, 6,200 aspect‑sentiment, and 5,600 relation records, with the new sacred‑heritage types comprising 64.8% of entities. Human evaluation on a 200‑review subset yielded moderate precision and recall for entities and aspect‑sentiment extraction, and a 78.9% agreement on first‑listed relation candidates.