Generating Fake but Realistic Headlines Using Deep Neural Networks

Social media platforms such as Twitter and Facebook implement filters to detect fake news as they foresee their transition from social media platform to primary sources of news. The robustness of such filters lies in the variety and the quality of the data used to train them. There is, therefore, a need for a tool that automatically generates fake but realistic news. In this paper, we propose a deep learning model that automatically generates news headlines. The model is trained with a corpus of existing headlines from different topics. Once trained, the model generates a fake but realistic headline given a seed and a topic. For example, given the seed “Kim Jong Un” and the topic “Business”, the model generates the headline “kim jong un says climate change is already making money”. In order to better capture and leverage the syntactic structure of the headlines for the task of synthetic headline generation, we extend the architecture - Contextual Long Short Term Memory, proposed by Ghosh et al. - to also learn a part-of-speech model. We empirically and comparatively evaluate the performance of the proposed model on a real corpora of headlines. We compare our proposed approach and its variants using Long Short Term Memory and Gated Recurrent Units as the building blocks. We evaluate and compare the topical coherence of the generated headlines using a state-of-the-art classifier. We, also, evaluate the quality of the generated headline using a machine translation quality metric and its novelty using a metric we propose for this purpose. We show that the proposed model is practical and competitively efficient and effective.