Offensive language detection and target identification are essential for maintaining respectful online environments. While these tasks have been widely studied for English, comparatively less attention has been given to other language, including Persian and Pashto, and the effectiveness of recent large language models for these languages remains underexplored. To address this gap, we created a comprehensive benchmark of diverse modeling approaches in Persian and Pashto. Our evaluation covers zeroshot, fine-tuned, and cross-lingual transfer settings, analyzing when detection succeeds or fails across different model approaches. This study provides one of the first systematic analyses of offensive language detection and crosslingual transfer between these languages.

This paper describes the SMASH submission to SemEval-2026 Task~9 on multilingual, multicultural, and multi-event polarisation detection. The task comprises (i) binary polarisation detection, (ii) multi-label classification of polarisation types, and (iii) multi-label identification of polarisation manifestations across all available languages. We propose a language-adaptive ensemble framework combining monolingual and multilingual encoder-only transformers, together with a principled out-of-fold (OOF) threshold tuning strategy. Instead of relying on fixed probability thresholds, we jointly tune ensemble weights and class-wise decision thresholds to directly optimise macro-F1 under the official evaluation metric. Our experiments show that (1) monolingual encoders dominate in several high-resource languages but benefit from complementary multilingual signals, (2) no single multilingual backbone universally outperforms others across languages and subtasks, and (3) language-specific class threshold tuning substantially improves performance due to large cross-lingual variation in class distributions. Our results demonstrate that careful logit-level ensembling and threshold tuning provide strong performance for multilingual, imbalanced, multi-label polarisation detection. Across 22 evaluation languages, SMASH ranks among the top three systems in a substantial number of language–subtask pairs. Specifically, it ranks in the top three for 5 languages in Subtask 1, 14 languages in Subtask 2, and 16 languages in Subtask 3, demonstrating strong and consistent performance across diverse languages and tasks. Our system achieves average macro-F1 scores of 0.81, 0.62, and 0.53 for Subtasks 1, 2, and 3, respectively.

Effective detection of hate content is essential for maintaining healthy online communities, yet performance can vary substantially across languages. This work investigates a core practical question: Can training data from one language support hate-speech detection in another language with comparable performance? To answer this question, we conduct a large-scale study across 14 languages. Our analysis shows that transfer effectiveness could depend on several factors—including, but not limited to, dataset size, resources and temporal overlap. At the same time, we observe consistent patterns in which certain language pairs exhibit stronger transfer performance. Upon further analysis, we find that these patterns can be attributed to cultural cues or shared societal characteristics. Our analysis includes low-resource languages, offering practical insights into when cross-lingual transfer can support them and where its limitations emerge. These findings provide guidance for designing more reliable and generalizable hate detection systems.

Warning: This paper includes examples of hate speech for research purposes.

We introduce ParsHate, a manually annotated dataset of 10,000 Persian tweets spanning 2013–2022, representing the first decade-long benchmark for hate speech detection in Persian. The dataset contains 31% hateful content and supports both hate detection and multi-label fine-grained target identification across seven structured target categories. ParsHate also distinguishes explicit and implicit hate, marks explicit and implicit targets, and provides span-level rationales. Data collection combines random and score-stratified temporal sampling to reduce keyword-driven bias while preserving natural label distributions. Applying SOTA models for Persian hate-speech detection on ParsHate shows moderate performance (79% F1), especially with samples from earlier years, and low performance with target identification (25.5% macro-F1). This emphasizes the diverse sampling of hate speech in ParsHate and its challenging nature that requires more advanced methods for better performance. Dataset is made publicly available.

Toxic language detection is crucial for creating safer online environments and limiting the spread of harmful content. While toxic language detection has been under-explored in Persian, the current work compares different methods for this task, including fine-tuning, data enrichment, zero-shot and few-shot learning, and cross-lingual transfer learning. What is especially compelling is the impact of cultural context on transfer learning for this task: We show that the language of a country with cultural similarities to Persian yields better results in transfer learning. Conversely, the improvement is lower when the language comes from a culturally distinct country.