| With the penetration of the Internet,the network has become an important source for people to obtain information.The arbitrariness and freedom of modern network language expression make the expression more diverse.This also leads to the wide use of various variants of words on the Web page,which brings great challenges to the detection of information security on Web pages.The traditional method of detecting sensitive words through precise matching has poor performance in identifying variants.Illegal elements generate sensitive word variants by inserting special characters,phonetic substitution,glyph disassembly,similar glyph substitution,similar pronunciation substitution in sensitive words,to avoid the review of sensitive word filtering system,to pass bad information to users,and to threaten the security of the network environment.This paper investigates the detection scheme based on extended Trie tree and the detection algorithm based on deep learning model for the issue of Chinese sensitive word variant detection.The main work completed in this article is as follows:(1)Collect and organize variants of Chinese sensitive words on the internet,summarize the generation rules of variants,divide them into two categories: structurally separable and unstructured separable,and annotate the data with BIO.At the same time,on the basis of summarizing the rules of character variant generation,a method for calculating the similarity of Chinese characters based on the comprehensive sound and shape of characters is proposed to implement a sensitive word variant generator.Expand the Chinese sensitive word variant dataset through a variant generator to provide data support for subsequent experiments.(2)A variant detection scheme based on extended Trie tree is proposed for structured split variants.The generation rules of structured split variants are integrated into the construction and retrieval process of Trie tree.The accuracy of this scheme can reach 98.23% in the experiment,and it only takes about 748 milliseconds for text retrieval of 20000 words,which can efficiently identify structured split variants.(3)For unstructured split variants,a Chinese BERT Bi LSTM-CRF model is proposed.Compared to BERT,which adds pronunciation and glyph features of Chinese characters in the embedding layer,Chinese BERT is more in line with the recognition task of Chinese variants.Bi LSTM can effectively learn contextual features and CRF constrains the final label output.The experimental results show that this model outperforms other models on the Chinese sensitive word variant dataset,with accuracy,recall,and F1 values reaching 90.21%,87.75%,and 88.96%,respectively.(4)Applying the above variant recognition scheme to a sensitive word variant detection system,it can be developed by combining techniques such as web crawling,message queuing,and front-end and back-end separation,enabling real-time detection of web content.Users can customize monitoring tasks.Once sensitive words are found,the system will promptly send alarm information to users.Users can also view the statistical data of monitoring results through a visual interface.This system can provide convenient,efficient,and low-cost real-time detection services for sensitive words on web pages for small and medium-sized enterprises. |
