基于谐音干扰词替换的中文仇恨言论检测方法  

Chinese hate speech detection method based on the replacement of homophonic noise words

作　　者：王琰慧 王小龙 张顺香[1,2] 周渝皓 汪才钦 WANG Yanhui;WANG Xiaolong;ZHANG Shunxiang;ZHOU Yuhao;WANG Caiqin(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan 232001,China;Artificial Intelligence Research Institute of Hefei Comprehensive National Science Center,Hefei 230026,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001 [2]合肥综合性国家科学中心人工智能研究院,安徽合肥230026

出　　处：《应用科技》2024年第3期72-81,共10页Applied Science and Technology

基　　金：国家自然科学基金面上项目(62076006);安徽高校协同创新项目(GXXT-2021-008).

摘　　要：社交网络中的仇恨言论常含有形式多变的谐音干扰词,使得现有方法难以适应此现象,不能满足即时检测的要求。针对此问题,提出一种基于谐音干扰词替换的中文仇恨言论检测方法,提取原义词替换谐音干扰词,解决原有方法处理相对滞后问题。首先,对文本预处理,通过N-gram提取干扰词候选项,并利用点间互信息和邻接熵进行过滤;然后,计算拼音相似度筛选出谐音干扰词及其对应的候选原义词,通过语法结构和上下文语义相似确定原义词并对相应谐音干扰词进行替换,将替换后的文本作为分类层输入;最后,使用RoBERTa-wmm-ext得到语义特征,并通过Softmax计算仇恨情感倾向以实现检测任务。在数据集上进行实验,结果表明提出的模型有效地提升中文仇恨言论的检测效果。Hate speech in social media often includes creatively disguised homophonic noise words,making it difficult for existing methods to adapt to this phenomenon and to meet the requirements of real-time detection.To locate this issue,a Chinese hate speech detection method is proposed to resolve the lag in processing by mining the original words to replace homophonic noise words,thereby to settle the lag problem in the solution of former method.Firstly,the texts were preprocessed,the candidate items of noise words were extracted through N-gram,and filtered by using the pointwise mutual information and branch entropy.Then,homophonic noise words and their corresponding candidate items of the original words were recognized by calculating phonetic similarity.The original words were determined through syntactic structure and contextual semantic similarity,and the homophonic noise words were replaced by them accordingly.The replaced texts were subsequently inputted into the classification layer for further processing.Finally,RoBERTa-wmm-ext was employed to extract semantic features and Softmax was used to calculate the hate sentiment tendency,achieving the detection task.Experimental results on the public COLDataset demonstrate that the proposed model can effectively improve the performance of Chinese hate speech detection.

关 键 词：仇恨言论检测 谐音干扰词 拼音相似 语法结构 上下文语义 RoBERTa-wmm-ext CNN N-GRAM 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]