Subgraph-Induced Extraction Technique for Information (SETI) from Administrative Documents

Information Extraction plays a key role in the automation of auditing processes in administrative documents. However, variety in layout and language is always a challenging task. On the other hand, large volumes of public training datasets related to administrative documents such as invoices are rare to find. In this work, we use Graph Attention Network model for information extraction. This type of model makes it easier to understand the mechanism as compared to classical neural networks due to the visualization of link between entities in the graph. Moreover, it maximizes the layout and structure retrieval which is a crucial advantage in administrative documents. From the same graph, our model learns at different graph levels to encapsulate dynamic and more enriched knowledge in each batch, thus maximizing the generalization on smaller dataset. We present how the model learns in each graph level and compare the results with baselines on private as well as public datasets. Our model succeeds in improving recall and precision scores for some classes in our private dataset and produces comparable results for public datasets designed for Form Understanding and Information Extraction.