CloudScan - A configuration-free invoice analysis system using recurrent neural networks

摘要

We present CloudScan; an invoice analysis system that requires zeroconfiguration or upfront annotation. In contrast to previous work, CloudScandoes not rely on templates of invoice layout, instead it learns a single globalmodel of invoices that naturally generalizes to unseen invoice layouts. Themodel is trained using data automatically extracted from end-user providedfeedback. This automatic training data extraction removes the requirement forusers to annotate the data precisely. We describe a recurrent neural networkmodel that can capture long range context and compare it to a baseline logisticregression model corresponding to the current CloudScan production system. Wetrain and evaluate the system on 8 important fields using a dataset of 326,471invoices. The recurrent neural network and baseline model achieve 0.891 and0.887 average F1 scores respectively on seen invoice layouts. For the hardertask of unseen invoice layouts, the recurrent neural network model outperformsthe baseline with 0.840 average F1 compared to 0.788.