A Hierarchical Multi-label Classification Algorithm for Scientific Papers Based on Graph Attention Networks

Abstract

Scientific paper classification refers to assigning one or more subject categories to papers. This task requires a lot of domain knowledge and heavy manual annotation. With the gradual increase in interdisciplinary research, a paper often has multiple categories. For instance, both Chinese Library Classification (http://www.ztflh.com/) and Engineering Village (EI) have a complete classification system, and there is a hierarchical relationship between the categories. The category of the paper has a hierarchical structure, so the paper classification can be converted into a hierarchical classification problem. However, the existing methods cannot effectively classify papers due to the following two reasons: First, these methods cannot well capture the semantic relationship between papers. Second, they neglect to model the hierarchical structure of labels. In this paper, we propose a hierarchical label attention model based on graph attention network, which utilizes word co-occurrence to model the semantic relationship of papers. We use multiple linear layers to model the category hierarchy and combine every hierarchy of labels through an attention mechanism. The experiments are conducted on CNKI (https://www.cnki.net/) and RCV1 datasets. The experimental results demonstrate that our method is superior to the other methods in the task of scientific paper classification.