While detecting offensive language in online spaces remains an important societal issue, there is still a significant gap in existing research and practial datasets specific to chatbots. Furthermore, many of the current efforts by service providers to automatically filter offensive language are vulnerable to users’ deliberate text manipulation tactics, such as misspelling words. In this study, we analyze offensive language patterns in real logs of 6,254,261 chat utterance pairs from the commercial chat service Simsimi, which cover a variety of conversation topics. Based on the observed patterns, we introduce a novel offensive language detection method—a contrastive learning model that embeds chat content with a random masking strategy. We show that this model outperforms existing models in detecting offensive language in open-domain chat conversations while also demonstrating robustness against users’ deliberate text manipulation tactics when using offensive language. We release our curated chatbot dataset to foster research on offensive language detection in open-domain conversations and share lessons learned from mitigating offensive language on a live platform.

***Warning**: this paper contains content that may be offensive or upsetting.*Most hate speech datasets neglect the cultural diversity within a single language, resulting in a critical shortcoming in hate speech detection. To address this, we introduce **CREHate**, a **CR**oss-cultural **E**nglish **Hate** speech dataset.To construct CREHate, we follow a two-step procedure: 1) cultural post collection and 2) cross-cultural annotation.We sample posts from the SBIC dataset, which predominantly represents North America, and collect posts from four geographically diverse English-speaking countries (Australia, United Kingdom, Singapore, and South Africa) using culturally hateful keywords we retrieve from our survey.Annotations are collected from the four countries plus the United States to establish representative labels for each country.Our analysis highlights statistically significant disparities across countries in hate speech annotations.Only 56.2% of the posts in CREHate achieve consensus among all countries, with the highest pairwise label difference rate of 26%.Qualitative analysis shows that label disagreement occurs mostly due to different interpretations of sarcasm and the personal bias of annotators on divisive topics.Lastly, we evaluate large language models (LLMs) under a zero-shot setting and show that current LLMs tend to show higher accuracies on Anglosphere country labels in CREHate.Our dataset and codes are available at: https://github.com/nlee0212/CREHate

Increasingly, language models and machine translation are becoming valuable tools to help people communicate with others from diverse cultural backgrounds. However, current language models lack cultural awareness because they are trained on data representing only the culture within the dataset. This presents a problem in the context of hate speech classification, where cultural awareness is especially critical. This study aims to quantify the cultural insensitivity of three monolingual (Korean, English, Arabic) hate speech classifiers by evaluating their performance on translated datasets from the other two languages. Our research has revealed that hate speech classifiers evaluated on datasets from other cultures yield significantly lower F1 scores, up to almost 50%. In addition, they produce considerably higher false negative rates, with a magnitude up to five times greater, demonstrating the extent of the cultural gap. The study highlights the severity of cultural insensitivity of language models in hate speech classification.