| In the recent years,with the rapid development and popularization of the Internet,China has accelerated the informatization process and entered the era of big data.The Internet econ-omy formed by various economic activities based on the Internet has become an indispensable part of the market economy,and the Internet advertising,which has grown rapidly in market size,has gradually become the pillar industry of the Internet economy.Nowadys,the Internet is full of various forms of ads,including a large number of illegal ads that will infringe on the rights and interests of users and affect the user experience.China’s Internet advertising administration is committed to building a sound regulatory system to ensure the healthy development of online advertising.However,due to various factors,the current regulatory system can not effectively su-pervise the rapidly expanding Internet advertising industry,among which the lagging of intelligent supervision technology for illegal advertising is the key technical reason.Meanwhile,benchmark datasets and methods for illegal advertising are scarce,which leads to the inability of related in-depth research.Therefore,this thesis collects a large-scale Internet ads,with the participation of legal professionals,combined with machine learning algorithms,to build a high-quality dataset for illegal advertising.Based on the dataset,this thesis studies the topic model used to mine the hot content and topic distribution of illegal ads,and the classification models used to accurately classify the violated provisions of illegal ads,which provides technical support for the realization of the intelligent process of the regulatory system from collection to classification to supervision.Our major contributions are as follows:Firstly,this thesis collects a large-scale Internet ads in various ways,and the legal profes-sionals participate in the construction process of dataset.The dataset is built in the way of small sample labeling training+large sample classification+manual review to guarantee the correctness and rigorousness of the dataset.Besides,detailed analysis of the characteristics of illegal ads are demonstrated based on datasets.Secondly,in order to mine the hot content and topic distribution of illegal ad,so as to track the tendency and analyze the public opinion of illegal ads,this thesis proposes a semi-supervised topic model named Lead LDA.Lead LDA extracts keywords of illegal ads from multiple perspectives as topic guide words,guides the topic distribution in a semi-supervised manner,and further improves the discrimination between topics and the quality of topics by promoting named entities.Finally,in order to accurately classify the violated provisions of illegal ads and improve the efficiency and accuracy of illegal advertising supervision,this thesis proposes two classification models from different perspectives:semantic structure feature fusion and abstract semantic asso-ciation,which are multi-feature integration classification model based on Lead LDA and IAD-Net model.The multi-feature integration classification model fully enriches the semantics and features of illegal ads and it’s accuracy reaches the advanced level.Based on deep learning methods,IAD-Net introduces an auxiliary embedding layer to enhance the semantics of lexicons in short ads,and an interactive attention mechanism to capture the relationship between lexicons in ads and its legality.Experimental results demonstrate the accuracy of IAD-Net exceeds the state-of-the-air methods. |
