How To Classify Document Segments Using Graph Based Representation and Neural Networks

Bottom-up segmentation approaches, where words are combined into lines, lines into paragraphs, etc., are well-known nowadays. In this work, we use this principle for classification of segments. The features of the blocks that make up a segment and their relationships can be sufficient for classification. Our approach combines a specific choice of features that increases explainability by employing the generalizing ability of neural networks. This work presents an intermediate result, but, nevertheless, the 0.793 classification accuracy on the part of PubLayNet indicates the competitive potential of the approach. It is expected that, similar to the transition from MLP to CNN for MNIST, the transition from vector representation at our stage to GCN at the planned stage will allow us to achieve good results in document classification.