Martin Holeček

Martin Holeček

The automation of document processing is gaining recent attention due to the great potential to reduce manual work through improved methods and hardware. Neural networks have been successfully applied before - even though they have been trained only on relatively small datasets with hundreds of documents so far. To successfully explore deep learning techniques and improve the information extraction results, a dataset with more than twenty-five thousand documents has been compiled, anonymized and is published as a part of this work. We will expand our previous work where we proved that convolutions, graph convolutions and self-attention can work together and exploit all the information present in a structured document. Taking the fully trainable method one step further, we will now design and examine various approaches to using siamese networks, concepts of similarity, one-shot learning and context/memory awareness. The aim is to improve micro F1 of per-word classification on the huge real-world document dataset. The results verify the hypothesis that trainable access to a similar (yet still different) page together with its already known target information improves the information extraction. Furthermore, the experiments confirm that all proposed architecture parts are all required to beat the previous results. The best model improves the previous state-of-the-art results by an 8.25 gain in F1 score. Qualitative analysis is provided to verify that the new model performs better for all target classes. Additionally, multiple structural observations about the causes of the underperformance of some architectures are revealed. All the source codes, parameters and implementation details are published together with the dataset in the hope to push the research boundaries since all the techniques used in this work are not problem-specific and can be generalized for other tasks and contexts.

Martin Holeček, Antonín Hoskovec, Petr Baudiš et al.

Abstract--- Table detection and extraction has been studied in the context of documents like reports, where tables are clearly outlined and stand out from the document structure visually. We study this topic in a rather more challenging domain of layout-heavy business documents, particularly invoices. Invoices present the novel challenges of tables being often without outlines - either in the form of borders or surrounding text flow - with ragged columns and widely varying data content. We will also show, that we can extract specific information from structurally different tables or table-like structures with one model. We present a comprehensive representation of a page using graph over word boxes, positional embeddings, trainable textual features and rephrase the table detection as a text box labeling problem. We will work on our newly presented dataset of pro forma invoices, invoices and debit note documents using this representation and propose multiple baselines to solve this labeling problem. We then propose a novel neural network model that achieves strong, practical results on the presented dataset and analyze the model performance and effects of graph convolutions and self-attention in detail.