Hierarchical Classification with Convolutional Neural Networks for Biomedical Literature

Multi-label document classification is a challenge task in many real-world applications. Recently, hierarchical classification methods have been widely used in document classification. However, at each layer of the hierarchical architecture, a classifier is trained independently, ignoring the relations between the other layers. In addition, compared with general documents, the biomedical literature only consists of the title and abstract information instead of the whole context. To overcome this problem, in this paper, we propose a novel hierarchical indexing method with Convolutional Neural Networks (CNNs) to tackle with the biomedical abstract document collections. First, we construct a hierarchical CNN indexing architecture which adaptively groups word2vec categories into (coarse) subsets by clustering. Next, a suitable loss function is designed for CNN training, where multi-label classification is actually performed in a coarse-to-fine learning style. Thereafter, a high-dimensional space representation is generated with feature extension by word sequence embedding, which contains more semantic information than bag-of-words. Experimental results show that our CNN model achieves an impressed performance.