A Hierarchical Neural Extractive Summarizer for Academic Papers

Recent neural network-based models have proven successful in summarization tasks. However, previous studies mostly focus on comparatively short texts and it is still challenging for neural models to summarize long documents such as academic papers. Because of their large size, summarization for academic papers has two obstacles: it is hard for a recurrent neural network (RNN) to squash all the information on the source document into a latent vector, and it is simply difficult to pinpoint a few correct sentences among a large number of sentences. In this paper, we present an extractive summarizer for academic papers. The idea is converting a paper into a tree structure composed of nodes corresponding to sections, paragraphs, and sentences. First, we build a hierarchical encoder-decoder model based on the tree. This design eases the load on the RNNs and enables us to effectively obtain vectors that represent paragraphs and sections. Second, we propose a tree structure-based scoring method to steer our model toward correct sentences, which also helps the model to avoid selecting irrelevant sentences. We collect academic papers available from PubMed Central, and build the training data suited for supervised machine learning-based extractive summarization. Our experimental results show that the proposed model outperforms several baselines and reduces high-impact errors.