Fake News Challenge

In this work, we have explored the task of stance detection between news article headline and body, which was hosted as part of the Fake News Challenge (FNC). This work focuses on developing machine learning and deep learning models for stance detection with a special focus on feature engineering techniques. We have experimented with word embeddings like Word2vec, Glove and fasttext in our models. We have also used Term Frequency–Inverse Document Frequency (TF-IDF) features and handcrafted features that include polarity and n-gram counts which are useful to this problem. In addition to word embeddings, skip-thought sentence embeddings have also been used to capture the context in the text at sentence level. We have devised various deep learning models like conditional and shared Bidirectional Long Short-Term Memory (BiLSTM) and utilized the effectiveness of fully connected neural networks with dropout. Motivated by the effectiveness of Gradient Boosting-based models for shared tasks, we trained a classifier using LightGBM technique which boosted the performance of our models. Extensive efforts were invested into experiments and evaluations for finding the correct set of features for training different learning models. Our ensemble model, which was a combination of the Skip thought and Gradient Boosting based models was able to achieve a FNC-score of 84.15% which, based on the rankings provided by the challenge, and to the best of our knowledge is the best score on this dataset reported till date.