Confirm or Refute?: A Comparative Study on Citation Sentiment Classification in Clinical Research Publications.

•We present a comparison of three approaches to citation sentiment analysis in clinical trial publications: a rule-based method, SVM and two neural network variants: a convolutional neural network model and a bidirectional long short-term memory model.•Convolution neural network model augmented with hand-crafted features yields the best performance both in 10-fold cross-validation and generalizes better to unseen data.•Hand-crafted phrase-based features and high-level contextual features are more predictive for this task than n-gram or embedding-based features.