Classification of Fake and Real Articles Based on Support Vector Machines

Fake or real? That is the question, even in the context of languages. In this course project, we are given the task of distinguishing real Broadcast News articles from fake “articles” generated by a trigram model trained from the 100 million word corpus of Broadcast News articles from 1992–1996. This task is clearly not difficult for humans, while machines are not as smart as us to tell whether the articles make sense at higher levels such as semantics. In this report we presented a machine-oriented solution based on Support Vector Machines [1] to classify between fake and real articles. Features deployed in our task covered various aspects of languages. For example, we tried n-gram language models and parsers to generate features that capture the local dependencies and syntactical information of the texts in the data sets. The topic model based on Latent Dirichlet Allocation (LDA) [2] was used in an attempt to distinguish the articles at the semantic level. With all these features and a good classifier (namely SVM), we hope to design a program that is capable of discovering the deficiencies in the conventional trigram language model and therefore discriminating between fake and real articles.