MCKIE: Multi-class Key Information Extraction from Complex Documents Based on Graph Convolutional Network

The majority of key information extraction in document analysis work relies on simple layout scenes with few classes, such as the date and amount on invoices or receipts. However, many document applications entail sophisticated layouts with various types of elements. The mortgage contract agreement, for example, often has over ninety entities, and it is expensive to design template for each document. To this end, we propose an efficient multi-class key information extraction (MCKIE) method based on graph convolutional network. In detail, we design a graph construction strategy to generate a text box graph. Then, MCKIE utilizes the message-passing mechanism to learn node representation by aggregating information from neighborhoods. Besides, we compare the performance of various graph construction methods and verify the effectiveness of MCKIE on the realistic contract document dataset. Extensive experimental results show that the proposed model significantly outperforms other baseline models.