Explainable importance ranking of research paper (NVIDIA)

Abstract

In this paper, we propose 2 parts of the work. First, we do research and evaluated 7 BERT (Bidirectional Encoder Representation from Transformers) models on citation intent classification task on scientific papers in medical and computer science fields, and proposed a method for data augmentation which is able to improve all 7 models' performances. Second, we build a citation network using a citation relationship dataset we extracted and propose a new Leaky PageRank algorithm, which can perform on nodes of higher dimensions, and has managed to rank papers with high explainability. We have also compared this new method with the original weighted PageRank method and have proved in mathematics that our new method can solve problems which original one cannot.