Multi-stage News-Stance Classification Based on Lexical and Neural Features

The amount of fake news present on the internet poses a great challenge for many online communities including manual fact-checkers who struggle to prevent the spread of misinformation and its negative impact. Detecting the stance of a news article involves classifying its perspective (e.g. agree, disagree, discuss, or unrelated) to a particular claim or headline which could support human fact-checkers to determine the veracity of the claims. Prior work on fake-news stance detection has proposed one-stage multi-class classification solutions which have limited success in detecting related pairs due to imbalanced class distributions in the data. This paper describes an improved approach to the stance detection of Fake News Challenge (FNC-1) based on multi-stage feature-assisted Deep Learning approaches. We break down the multi-class classification problem into two-stage and three-stage classifiers by combining the lexical-overlap features with Deep Learning techniques in an effort to mitigate the class imbalance problem. The experimental results demonstrate that the proposed models improve upon the state-of-the-art Accuracy and F1 score for stance detection. We also experimentally show that our models achieve solid results on minority classes i.e. agree and disagree without using fine-tuning approach or adding more training samples.