Hate Speech Detection Method Integrating Prefix Tuning and Prompt Learning

doi:10.3778/j.issn.1673-9418.2405026

Abstract

Abstract: With the development of online social platforms, the harm of cyberbullying has become increasingly prominent, with hate speech being one of its manifestations. Researching detection methods for hate speech is conducive to building a healthy Internet environment. However, on one hand, previous hate speech detection methods which relied heavily on manual supervision and keyword filtering, are not only costly in terms of human resources but also lack precision in classification; on the other hand, under the regulation of Internet platforms, the expression of hate speech has become more subtle and covert, posing higher demands on the text comprehension capabilities of detection models. Large models like ChatGPT have opened up new possibilities for hate speech detection tasks. Inspired by their exceptional performance in various downstream tasks, a hate speech detection method named P-Prompt, which integrates prefix tuning and prompt learning, is proposed. Specifically, it first uses the prefix tuning method to fine-tune the large model on relevant datasets, and incorporates the prompt learning method to enable the model to recognize and focus on hate speech vocabulary in the text to be detected, thereby further enhancing the model􀆳s ability to identify hate speech. Experimental results validate the effectiveness of large models in hate speech detection tasks, and compared with traditional methods, the P-Prompt method achieves significant improvements in both binary and multi-class classification tasks of hate speech detection across various evaluation metrics.

Key words: hate speech detection, large language models, prefix tuning, prompt learning

摘要： 随着网络社交平台的发展，网络暴力的危害性日益凸显，其中仇恨言论作为网络暴力的一种表现形式，对其检测方法的研究有助于构建一个健康的互联网环境。然而，一方面先前的仇恨言论检测方法主要依赖于人工监管和关键词过滤，这些方法存在人工成本高且分类效果不佳的问题；另一方面，在互联网平台的监管下，仇恨言论的表达也变得更加含蓄、隐蔽，这对检测模型的文本理解能力提出了更高的要求。以ChatGPT为代表的大模型为仇恨言论检测任务提供了新的可能性，受其在各项下游任务中出色表现的启发，提出了一种融合前缀调优和提示学习的仇恨言论检测方法P-Prompt。利用前缀调优方法将大模型在相关数据集上进行微调，同时结合提示学习方法使模型能够识别并关注待检测文本中的仇恨言论词汇，从而进一步提升模型对仇恨言论的识别能力。实验结果验证了大模型在仇恨言论检测任务的有效性，与传统方法相比P-Prompt方法在仇恨言论检测的二分类和多分类任务中各项评价指标都取得了较明显的提升。

关键词: 仇恨言论检测, 大语言模型, 前缀调优, 提示学习

XU Lei, HU Yahao, CHEN Man, CHEN Jun, PAN Zhisong. Hate Speech Detection Method Integrating Prefix Tuning and Prompt Learning[J]. Journal of Frontiers of Computer Science and Technology, 2025, 19(1): 97-106.

徐磊, 胡亚豪, 陈满, 陈军, 潘志松. 融合前缀调优和提示学习的仇恨言论检测方法[J]. 计算机科学与探索, 2025, 19(1): 97-106.

References

[1] KOWSARI K, MEIMANDI K J, HEIDARYSAFA M, et al. Text classification algorithms: a survey[J]. Information, 2019, 10(4): 150.
[2] SCHMIDT A, WIEGAND M. A survey on hate speech detection using natural language processing[C]//Proceedings of the 5th International Workshop on Natural Language Processing for Social Media. Stroudsburg: ACL, 2017: 1-10.
[3] PETERS M E, NEUMANN M, IYYER M, et al. Deep contextualized word representations[EB/OL]. [2023-10-29]. http:// arxiv.org/abs/1802.05365.
[4] DEVLIN J, CHANG M W, LEE K, et al. BERT: pre-training of deep bidirectional transformers for language understan-ding[EB/OL]. [2023-04-12]. http://arxiv.org/abs/1810.04805.
[5] STRUSS J M, SIEGEL M, RUPPENHOFER J, et al. Overview of GermEval Task 2, 2019 Shared Task on the identiﬁcation of offensive language[C]//Proceedings of the 15th Conference on Natural Language Processing, Friedrich-Alexander-Universität Erlangen-Nürnberg,Oct 9-11, 2019: 352-363.
[6] PARASCHIV A, CERCEL D C. UPB at GermEval-2019 Task 2: BERT-based offensive language classification of German Tweets[C]//Proceedings of the 15th Conference on Natural Language Processing, Friedrich-Alexander-Universität Erlangen-Nürnberg,Oct 9-11, 2019: 297-303.
[7] RISCH J, STOLL A, ZIEGELE M, et al. hpiDEDIS at Germ-Eval 2019: offensive language identiﬁcation using a German BERT model[C]//Proceedings of the 15th Conference on Natural Language Processing, Friedrich-Alexander-Universität Erlangen-Nürnberg, Oct 9-11, 2019: 231-237.
[8] DENG J, ZHOU J, SUN H, et al. COLD: a benchmark for Chinese offensive language detection[EB/OL]. [2024-01-24]. http://arxiv.org/abs/2201.06025.
[9] XIANG G, FAN B, WANG L, et al. Detecting offensive Tweets via topical feature discovery over a large scale Twitter corpus[C]//Proceedings of the 21st ACM International Conference on Information and Knowledge Management. New York: ACM, 2012: 1980-1984.
[10] CHEN Y, ZHOU Y, ZHU S, et al. Detecting offensive language in social media to protect adolescent online safety[C]//Proceedings of the 2012 International Conference on Privacy, Security, Risk and Trust and 2012 International Confernece on Social Computing. Piscataway: IEEE, 2012: 71-80.
[11] DINAKAR K, JONES B, HAVASI C, et al. Common sense reasoning for detection, prevention, and mitigation of cyberbullying[J]. ACM Transactions on Interactive Intelligent Systems, 2012, 2(3): 1-30.
[12] GITARI N D, ZHANG Z, DAMIEN H, et al. A lexicon-based approach for hate speech detection[J]. International Journal of Multimedia and Ubiquitous Engineering, 2015, 10(4): 215-230.
[13] VAN HEE C, LEFEVER E, VERHOEVEN B, et al. Detection and fine-grained classiﬁcation of cyberbullying events[C]//Recent Advances in Natural Language Processing, Hissar, Sep 7-9, 2015: 672-680.
[14] 李锦, 夏鸿斌, 刘渊. 基于BERT的双特征融合注意力的方面情感分析模型[J]. 计算机科学与探索, 2024, 18(1): 205-216.
LI J, XIA H B, LIU Y. Dual features local-global attention model with BERT for aspect sentiment analysis[J]. Journal of Frontiers of Computer Science and Technology, 2024, 18(1): 205-216.
[15] 韩坤, 潘宏鹏, 刘忠轶. 融合BERT多层次特征的短视频网络舆情情感分析研究[J]. 计算机科学与探索, 2024, 18(4): 1010-1020.
HAN K, PAN H P, LIU Z Y. Research on sentiment analysis of short video network public opinion by integrating BERT multi-level features[J]. Journal of Frontiers of Computer Science and Technology, 2024, 18(4): 1010-1020.
[16] DADVAR M, TRIESCHNIGG D, ORDELMAN R, et al. Improving cyberbullying detection with user context[C]//Advances in Information Retrieval-35th European Conference on IR Research. Berlin, Heidelberg: Springer, 2013: 693-696.
[17] BROWN T B, MANN B, RYDER N, et al. Language models are few-shot learners[C]//Advances in Neural Information Processing Systems 33, Dec 6-12, 2020: 1877-1901.
[18] WEI J, WANG X, SCHUURMANS D, et al. Chain-of-thought prompting elicits reasoning in large language models[EB/OL]. [2023-04-19]. http://arxiv.org/abs/2201.11903.
[19] ZHOU D, SCHÄRLI N, HOU L, et al. Least-to-most prom-pting enables complex reasoning in large language models[EB/OL]. [2024-03-18]. http://arxiv.org/abs/2205.10625.
[20] LI X L, LIANG P. Prefix-Tuning: optimizing continuous prompts for generation[EB/OL]. [2023-11-13]. http://arxiv.org/abs/2101.00190.
[21] LIU X, ZHENG Y, DU Z, et al. GPT understands, too[EB/OL]. [2024-04-09]. http://arxiv.org/abs/2103.10385.
[22] LIU X, JI K, FU Y, et al. P-Tuning v2: prompt tuning can be comparable to fine-tuning universally across scales and tasks[EB/OL]. [2024-04-09]. http://arxiv.org/abs/2110.07602.
[23] JAHAN M S, OUSSALAH M. A systematic review of hate speech automatic detection using natural language processing[EB/OL]. [2024-03-07]. http://arxiv.org/abs/2106. 00742.
[24] ZHOU J, DENG J, MI F, et al. Towards identifying social bias in dialog systems: frame, datasets, and benchmarks[EB/OL]. [2023-05-15]. http://arxiv.org/abs/2202.08011.
[25] RADFORD A, NARASIMHAN K, SALIMANS T, et al. Improving language understanding by generative pre-trai-ning[EB/OL]. [2023-05-15]. http://arxiv.org/abs/2010.03455.
[26] NIKOLOV A, RADIVCHEV V. Nikolov-Radivchev at SemEval-2019 Task 6: offensive Tweet classification with BERT and ensembles[C]//Proceedings of the 13th International Workshop on Semantic Evaluation. Stroudsburg: ACL, 2019: 691-695.
[27] RANASINGHE T, ZAMPIERI M, HETTIARACHCHI H. BRUMS at HASOC 2019: deep learning models for multilingual hate speech and offensive language identiﬁcation[EB/OL]. [2024-04-27]. http://arxiv.org/abs/2004.06465.
[28] DOWLAGAR S, MAMIDI R. HASOCOne@FIRE-HASOC 2020: using BERT and multilingual BERT models for hate speech detection[EB/OL]. [2024-04-27]. http://arxiv.org/abs/ 2101.09007.
[29] DU Z, QIAN Y, LIU X, et al. GLM: general language model pretraining with autoregressive blank infilling[C]//Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics. Stroudsburg: ACL, 2022: 320-335.