arXiv:2510.10961v3 Announce Type: replace 
Abstract: As language models become increasingly deployed in online environments, toxicity detection and detoxification have received growing attention. Existing studies primarily focus on non-obfuscated text, which limits robustness when users intentionally disguise toxic expressions. In particular, Korean toxic expressions can be easily disguised through agglutinative morphology and Hangeul-specific orthographic variation. However, obfuscation in Korean remains largely unexplored, which motivates us to introduce a KOTOX: Korean toxic dataset for deobfuscation and detoxification. We categorize Korean obfuscation patterns into linguistically grounded classes, define transformation rules derived from real-world examples, and provide the resulting obfuscation framework as an open transformation package. Using these rules, we provide paired neutral and toxic sentences alongside their obfuscated counterparts. Models trained on our dataset better handle obfuscated text without sacrificing performance on non-obfuscated text. This is the first dataset that simultaneously supports deobfuscation and detoxification for the Korean language. We expect the dataset to facilitate better understanding and mitigation of obfuscated toxic content in LLM for Korean. Our code and data are available at https://github.com/leeyejin1231/KOTOX.

تم تقديم إطار عمل جديد للكشف عن التعبيرات السامة وإزالة سمومها باللغة الكورية، مع التركيز على أنماط التعتيم الفريدة في اللغة. يُعرف هذا الإطار باسم KOTOX، ويصنف هذه الأنماط ويوفر قواعد تحويل مستندة إلى أمثلة من العالم الحقيقي، بهدف تحسين قوة نماذج الكشف عن السمية.

Se ha introducido un nuevo marco para detectar y desintoxicar expresiones tóxicas en coreano, centrándose en los patrones de ofuscación únicos del idioma. Este marco, denominado KOTOX, categoriza estos patrones y proporciona reglas de transformación basadas en ejemplos del mundo real, con el objetivo de mejorar la robustez de los modelos de detección de toxicidad.

Un nouveau cadre pour détecter et détoxifier les expressions toxiques en coréen a été introduit, en se concentrant sur les motifs d'obfuscation uniques de la langue. Ce cadre, nommé KOTOX, catégorise ces motifs et fournit des règles de transformation basées sur des exemples du monde réel, visant à améliorer la robustesse des modèles de détection de toxicité.

A new framework for detecting and detoxifying toxic expressions in Korean has been introduced, focusing on the unique obfuscation patterns in the language. This framework, named KOTOX, categorizes these patterns and provides transformation rules based on real-world examples, aiming to enhance the robustness of toxicity detection models.

Obfuscation Rules for Detecting and Detoxifying Korean Toxicity

arXiv:2606.06065v2 Announce Type: replace 
Abstract: Second-language (L2) speech recognition often requires transcriptions of pronunciations and intended meanings. Multi-task learning (MTL) is a natural approach because it assumes that shared representations benefit both outputs. However, this paper shows that this assumption does not hold across Korean and English. MTL improves meaning but degrades surface transcription, especially in English, where the degradation scales with surface-meaning divergence measured by Levenshtein edit distance. Encoder analysis links these patterns to encoder-level entanglement, with Korean preserving distinct task representations while English produces nearly identical ones. Cross-task decoder analysis shows that the meaning dual-output decoder adapts with a unique representation, while the surface dual-output decoder remains constrained by the encoder. These findings motivate the design of MTL frameworks that mitigate encoder-level entanglement to reduce surface degradation in dual-output L2 automatic speech recognition.

تسلط دراسة حديثة الضوء على قيود التعلم المتعدد المهام (MTL) في التعرف على الكلام بلغة ثانية، خاصة بين الكورية والإنجليزية. تشير الأبحاث إلى أنه بينما يمكن أن يحسن MTL التعرف على المعاني، فإنه يؤثر سلبًا على دقة النسخ السطحي، خاصة في الإنجليزية، حيث ترتبط التدهورات بالتباين بين التمثيلات السطحية والمعنوية.

Un estudio reciente destaca las limitaciones del aprendizaje multitarea (MTL) en el reconocimiento de voz de segundas lenguas, especialmente entre el coreano y el inglés. La investigación indica que, aunque el MTL puede mejorar el reconocimiento de significados, afecta negativamente la precisión de la transcripción superficial, especialmente en inglés, donde la degradación se correlaciona con la divergencia entre las representaciones superficiales y de significado.

Une étude récente met en lumière les limites de l'apprentissage multitâche (MTL) dans la reconnaissance de la parole en langue seconde, en particulier entre le coréen et l'anglais. La recherche indique que, bien que le MTL puisse améliorer la reconnaissance des significations, il affecte négativement la précision de la transcription de surface, surtout en anglais, où la dégradation est corrélée à la divergence entre les représentations de surface et de signification.

A recent study highlights the limitations of multi-task learning (MTL) in second language speech recognition, particularly between Korean and English. The research indicates that while MTL can enhance meaning recognition, it adversely affects surface transcription accuracy, especially in English, where the degradation correlates with the divergence between surface and meaning representations.

Multi-task Learning is Not Enough: Representational Entanglement in Dual-output Second Language Speech Recognition

arXiv:2606.06797v1 Announce Type: new 
Abstract: Cultural-aspect work on large language models is dominated by a negative target: which outputs to suppress. We argue that a constructive counterpart is also needed, a working definition of what a culturally coherent response is rather than only what it must avoid, and instantiate it for Korean. We design an alignment-data pipeline around a prompt-based LLM seed generator that expands a Korean harm taxonomy, with a Korean-culturally-adapted safe-response policy at its centre: a per-category guideline grounded in Korean legal frameworks, social norms, and interpretive conventions, against which three frontier models each produce a candidate response. DPO fine-tuning on the resulting triplets improves the Korean cultural safe rate across six open-weight LLMs while causing no large degradation on Korean general-capability benchmarks, and qualitative outputs show fine-tuned models naming Korean statutes and institutional procedures and, where appropriate, supplying constructive Korean-context information alongside refusal.

دراسة حديثة بعنوان 'دمج الثقافة الكورية في توافق نماذج اللغة الكبيرة: نحو التماسك الثقافي' تؤكد على الحاجة إلى استجابات ثقافية متماسكة في نماذج اللغة الكبيرة (LLMs)، مع التركيز بشكل خاص على الثقافة الكورية. تقدم الدراسة مولد بذور LLM قائم على المطالبات يوسع تصنيف الأذى الكوري ويضع سياسة استجابة آمنة تتماشى مع الأطر القانونية والمعايير الاجتماعية الكورية.

Un estudio reciente titulado 'Cultura Coreana en la Alineación de LLM: Hacia la Coherencia Cultural' enfatiza la necesidad de respuestas culturalmente coherentes en los modelos de lenguaje de gran tamaño (LLM), enfocándose específicamente en la cultura coreana. La investigación introduce un generador de semillas LLM basado en indicaciones que expande una taxonomía de daño coreana y establece una política de respuesta segura alineada con los marcos legales y normas sociales coreanas.

Une étude récente intitulée 'L'Intégration de la Culture Coréenne dans l'Alignement des LLM : Vers une Cohérence Culturelle' souligne la nécessité de réponses culturellement cohérentes dans les modèles de langage de grande taille (LLM), en se concentrant spécifiquement sur la culture coréenne. La recherche introduit un générateur de semences LLM basé sur des invites qui élargit une taxonomie des préjudices coréens et établit une politique de réponse sécurisée alignée sur les cadres juridiques et les normes sociales coréennes.

A recent study titled 'Korean Culture into LLM Alignment: Toward Cultural Coherence' emphasizes the need for culturally coherent responses in large language models (LLMs), specifically focusing on Korean culture. The research introduces a prompt-based LLM seed generator that expands a Korean harm taxonomy and establishes a safe-response policy aligned with Korean legal frameworks and social norms.

Obfuscation Rules for Detecting and Detoxifying Korean Toxicity

Was this article worth reading? Share it

Ready to build your own newsroom?