Hate speech detection systems are usually trained on explicit language, but harm can also be delivered indirectly. A new arXiv preprint, "SarcasTraps: Effect of Sarcasm on Turkish Hate Speech Detection," argues that sarcasm is a particularly tricky form of indirect expression.

The abstract points out that standard datasets emphasize clearly hateful wording, which means models may miss messages where sarcasm carries the aggression. The paper focuses on Turkish, a language with rich sarcastic expression, and appears to introduce a resource or evaluation approach to expose this blind spot.

Since only the abstract is available, the details of the dataset or method are not described here. The work underscores the broader challenge of building hate speech detectors that understand tone and context, not just surface-level language.