Chinese Toxic Comment Detection: A Comparative Study of Traditional ML, Deep Learning, Encoder-Based and Decoder-Based Models

Youxi Tan; Mingjie  Fang; Da Shen; Jiayi  Xu; Baha Ihnaini

doi:10.47852/bonviewJCCE52026117

Authors

Youxi Tan Institute of Advanced Natural Language Processing, Wenzhou-Kean University, China
Mingjie Fang Institute of Advanced Natural Language Processing, Wenzhou-Kean University, China https://orcid.org/0009-0005-3256-8279
Da Shen Department of Mathematical Science, Wenzhou-Kean University, China
Jiayi Xu Department of Psychology, Wenzhou-Kean University, China
Baha Ihnaini Institute of Advanced Natural Language Processing, Wenzhou-Kean University, China https://orcid.org/0000-0002-2109-4793

DOI:

https://doi.org/10.47852/bonviewJCCE52026117

Keywords:

Chinese toxic comment detection, machine learning, deep learning, encoder-based model, decoder-based model, binary classification, multilingual NLP

Abstract

In recent years, with the fast growth of the internet and the continuous expansion of technological applications such as social media, the health and safety of the online environment has become a matter that requires serious attention. In the Chinese context, due to the complexity and diversity of syntactic expression, the detection of toxic language in Chinese faces unique challenges. This study focuses on the performance of traditional machine learning, deep learning, encoder-based transformation models, and decoder-based transformation models (LLMs) in the identification of toxic comments in Chinese, and compares the performance characteristics of different models. The study combines two main datasets, COLD and TOCAB, into a binary classification task, using accuracy, F1-score, precision, and recall as evaluation metrics to assess all the tested models. The final results show that among the tested models, the decoder-based Qwen1.5-7B (8-bit quantization) has the highest accuracy (94.71%), the traditional machine learning models and encoder-based transformation models perform moderately, while the deep learning models have lower accuracy (77%–80%) due to the limited context understanding, indicating that decoder-based large language models have advantages in the detection of toxic comments in Chinese.

Received: 8 May 2025 | Revised: 24 September 2025 | Accepted: 15 October 2025

Conflicts of Interest

The authors declare that they have no conflicts of interest to this work.

Data Availability Statement

The data that support the findings of this study are openly available in GitHub at https://github.com/TanYouxi/Chinese-Toxic-Comment-Detection.

Author Contribution Statement

Youxi Tan: Conceptualization, Software, Formal analysis, Datacuration, Writing – original draft, Writing – review & editing, Visualization. Mingjie Fang: Software, Validation, Data curation, Writing – original draft, Writing – review & editing, Visualization. Da Shen: Validation, Formal analysis, Writing – original draft, Writing – review & editing. Jiayi Xu: Writing – review & editing. Baha Ihnaini: Methodology, Investigation, Resources, Supervision, Project administration, Funding acquisition.