ChineseHarm-Bench: A Chinese Harmful Content Detection Benchmark

TL;DR

This paper introduces ChineseHarm-Bench, a comprehensive benchmark for detecting harmful content in Chinese, including a knowledge rule base and a knowledge-augmented baseline to improve detection performance.

Contribution

It provides the first large-scale, annotated Chinese harmful content dataset with explicit knowledge rules and a novel baseline integrating human knowledge and LLMs.

Findings

The benchmark covers six harmful content categories.

The knowledge-augmented baseline improves detection accuracy.

Code and data are publicly available.

Abstract

Large language models (LLMs) have been increasingly applied to automated harmful content detection tasks, assisting moderators in identifying policy violations and improving the overall efficiency and accuracy of content review. However, existing resources for harmful content detection are predominantly focused on English, with Chinese datasets remaining scarce and often limited in scope. We present a comprehensive, professionally annotated benchmark for Chinese content harm detection, which covers six representative categories and is constructed entirely from real-world data. Our annotation process further yields a knowledge rule base that provides explicit expert knowledge to assist LLMs in Chinese harmful content detection. In addition, we propose a knowledge-augmented baseline that integrates both human-annotated knowledge rules and implicit knowledge from large language models, enabling smaller models to achieve performance comparable to state-of-the-art LLMs. Code and data are available at https://github.com/zjunlp/ChineseHarm-bench.