Advertising Legality Recognition

摘要

As online marketing and advertising keep growing on the Internet, a large amount of advertisements are presented to consumers. How consumers, advertisers and the authorities identify false and overstated advertisements becomes a critical issue. In this paper, we address this problem, and propose various classification models to detect illegal advertisements. Illegal advertisement lists announced by the government and legal advertising data crawled from an online shopping website are used for training and testing the classification models. Naieve Bayes and SVM classifiers with various feature settings are explored on food and cosmetic datasets to demonstrate their feasibility. The experimental results show that log relative frequency ratio can be used as weights for unigram features to achieve the best accuracy. The accuracies of SVM classifiers on food and cosmetic datasets are 93.433％ and 86.037％; the false alarm rates are 0.083 and 0.166; and the missing rates are 0.053 and 0.115, respectively. Log relative frequency ratio is further used to mine verb phrases consisting of a transitive verb and an object noun from the illegal datasets. The mined verb phrases, which form an illegal advertising statement list, can be used as a reference for both the advertisers and the authority.