Improving Word Classification with Graph Structures on Semi-Structured Documents Using Neural Networks

The idea of the word classification is an extremely important and popular problem which is likely to appear between people involved in automatic document processing. It occurs at the stage of extracting various information fields from the document, while processing it. In this study we are investigating the area which relates to the semi-structured documents – in particular, cash receipts – characterized by high information density. The described method is a way of improvement to the existing classifying words method, considered as a basic model, for such documents’ information extraction. Background of this study is supposed to develop competitive method for extracting a graph structure from documents and a apply modified techniques of graph convolutional networks to work with extracted data. This will give us additional features at the stage of classifiers’ final output formation. We provide a separate, auxiliary block to help the neural architectures be more effective in spatial networks’ perception of described complex document structure and increase their final quality of word classification. The work contains 17 pages, 8 chapters, 6 drawings, 1 table, 7 sources. Keywords – word classification, graph convolutional networks.