Training a neural network model from small size of high quality labeled dataset for fake news detection[go to overview]
Fact-checking is the task of assessing the truthfulness of a factual claim made in written or spoken statements. Only small percentage of the set of stories/claims spreading over time could be covered by human fact-checkers. One reason is that fact-checking requires labour intensive research. For example, fact-checkers might contact with person/organizations mentioned in the claim, they might speak to experts if they lack background knowledge, investigate sources of facts, etc. To optimise fact-checking task, it is necessary to prioritise the documents which would be potentially false and have higher impacts in the society.
In this talk, I will focus on task of detecting potentially fake news articles, namely fake news prediction. First I will present the problems of popular datasets for a neural network based model for fake news detection. Then, I will briefly explain two deep semi supervised approaches  which are widely used in computer vision tasks, and the experimental settings for the task.
 Tarvainen, Antti et al. “Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results.” Advances in neural information processing systems. 2017. Lee, Dong-Hyun. “Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks.” Workshop on Challenges in Representation Learning, ICML. Vol. 3. 2013.
09.01.20 - 10:15