Learning from similarity and information extraction from structured documents

The automation of document processing has recently gained attention owing to its great potential to reduce manual work. Any improvement in information extraction systems or reduction in their error rates aids companies working with business documents because lowering reliance on cost-heavy and error-prone human work significantly improves the revenue. Neural networks have been applied to this area before, but they have been trained only on relatively small datasets with hundreds of documents so far. To successfully explore deep learning techniques and improve information extraction, we compiled a dataset with more than 25,000 documents. We expand on our previous work in which we proved that convolutions, graph convolutions, and self-attention can work together and exploit all the information within a structured document. Taking the fully trainable method one step further, we 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 F-1 of per-word classification in the huge real-world document dataset. The results verify that trainable access to a similar (yet still different) page, together with its already known target information, improves the information extraction. The experiments confirm that all proposed architecture parts (Siamese networks, employing class information, query-answer attention module and skip connections to a similar page) are all required to beat the previous results. The best model yields an 8.25% gain in the F-1 score over the previous state-of-the-art results. Qualitative analysis verifies 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, since all the techniques used in this work are not problem-specific and can be generalized for other tasks and contexts.