网络暴力言论检测的技术和实践

张冬瑜,卢俊宇,闵昶榮,林鸿飞

语言战略研究 ›› 2024, Vol. 9 ›› Issue (1) : 64-75.

欢迎访问《语言战略研究》官方网站!今天是
语言战略研究 ›› 2024, Vol. 9 ›› Issue (1) : 64-75. DOI: 10.19689/j.cnki.cn10-1361/h.20240106
语言战略研究

网络暴力言论检测的技术和实践

  • 张冬瑜1,卢俊宇2,闵昶榮2,林鸿飞2
作者信息 +

The Technologies and Practices of Detecting Online Violent Speech

  • Zhang Dongyu, Lu Junyu, Min Changrong and Lin Hongfei
Author information +
文章历史 +

摘要

暴力言论是网络语言治理的重点,由于社交媒体信息量巨大,运用自然语言处理等人工智能技术对网络进行自动检测是及时阻止网络暴力言论传播扩散的必然要求。由于网络暴力言论表达具有隐匿性和多样性,目前存在缺乏细粒度的中文暴力言论语料库,以及对于采取隐式表达的暴力言论检测效果不理想的问题。因此,本文构建了一个大规模的中文暴力言论语料库,并研制了基于双重对比学习和基于情绪辅助的多标签自训练暴力言论检测方法,并在多种数据集上验证了其有效性和先进性。在大语言模型时代,将其与背景知识相结合可以更准确地检测暴力言论,但其生成内容需要与人类价值观对齐以符合当今的社会道德标准。暴力言论检测研究在未来发展包括多模态暴力言论检测研究、基于大模型的暴力言论检测方法研究、基于大模型的暴力言论监管机制研究、暴力言论检测的可解释性研究等4个主要方向。

Abstract

With the rapid popularity of social media, user-generated content has shown an explosive growth. Against this background, online violent speech has become rampant, causing much concern and posing many challenges to internet language governance. Due to the implicitness and diversity of online violent speech, there are some problems in the automatic detection of those speeches. Firstly, there is a lack of a fine-grained and rich corpus of Chinese online violent speech for the training of automatic detection methods. Secondly, online violent speech adopts implicit expressions, thus weakening the capacity of automatic detection methods. This paper reviews the previous studies on internet language governance and points out the challenging issues. At the same time, a large-scale Chinese online violent speech corpus was constructed to lay the foundation for the detection of Chinese online violent speech. The online violent speech detection methods were proposed on the basis of Dual Contrastive Learning and Emotion-enriched Multi-label and verified in a variety of data sets. This paper argues that Large Language Models should be applied together with background information to detect violent speech more accurately. In addition, the content generated by these models should comply with moral standards of our society. This paper  ends with discussions of the impact of Large Language Models on online violent speech detection and suggestions for future research directions.

关键词

网络语言治理 / 暴力言论检测 / 大语言模型 / 语料库建设

引用本文

导出引用
张冬瑜,卢俊宇,闵昶榮,林鸿飞. 网络暴力言论检测的技术和实践[J]. 语言战略研究. 2024, 9(1): 64-75 https://doi.org/10.19689/j.cnki.cn10-1361/h.20240106
Zhang Dongyu, Lu Junyu, Min Changrong and Lin Hongfei. The Technologies and Practices of Detecting Online Violent Speech[J]. Chinese Journal of Language Policy and Planning. 2024, 9(1): 64-75 https://doi.org/10.19689/j.cnki.cn10-1361/h.20240106

基金

国家自然科学基金“基于认知视角和语义表示的隐喻识别与应用研究”(62076051)、国家自然科学基金“细粒度仇恨言论检测关键技术研究”(62376051)、国家自然科学基金“面向社交媒体的中文幽默计算研究”(62076046)、2023年度国家语委重点项目“汉语自然口语语料库建设及应用研究”(ZDI145–80)。

Accesses

Citation

Detail

段落导航
相关文章

/