Discourse Parsing via Weighted Bag-of-Words, Coattention Neural Encoder, and Coattentative Convolutional Neural Network

Discourse parsing is an important task in NLP for understanding logical relations between adjacent sentences. In this report, we experiment with three models to improve classification task on discourse parsing: 1) weighted bag of words model using scoring DNN; 2) coattention neural encoder model using affinity matrix; 3) coattentative convolutional neural network model. We then show in our experiments that these three models all improve significantly over baseline model and give competitive accuracy on the task of discourse parsing over Penn Discourse Treebank .