Knowledge-Graph- and GCN-Based Domain Chinese Long Text Classification Method

In order to solve the current problems in domain long text classification tasks, namely, length of a document, which makes it difficult for model capture key information, and lack expert knowledge, leads insufficient accuracy, based on knowledge graph convolutional neural network is proposed. BERT used encode text, each word’s corresponding vector as node so that initialized contains rich semantic information. Using trained entity–relationship extraction model, entity-to-entity–relationships document are extracted edges network, together with syntactic dependency The structure mask learn about edge relationships types further enhance learning ability dependencies between words. method improves accuracy by fusing features data features. Experiments three datasets—IFLYTEK, THUCNews, Chinese corpus Fudan University—show improvements 8.8%, 3.6%, 2.6%, respectively, relative model.