Information extraction from scanned invoice documents using deep learning methods

In this thesis, a comprehensive comparison was conducted between a Graph Convolutional Network (GCN) model and two transformer based models for the tasks of node classification and sequence labeling respectively. These deep learning models were trained to extract information from invoice documents, with the aim of improving the automated processing of such data. A dataset of 1000 invoices was utilized for training the models, while a separate dataset consisting of 250 invoices was employed for evaluating their performance. The results showed that the one of the transformer based models achieved better results than GCN model, by achieving an F1-score of 0.65 for LayoutLMv1 model and 0.72 for LayoutLMv3 model, compared to the GCN model's 0.32. The results of this study shows the effectiveness of transformer based neural network models in information extraction tasks from scanned invoice documents and provide information about success of transformers and graph convolutional networks for this task.