FLIE: Form Labeling for Information Extraction

Information extraction (IE) from forms remains an unsolved problem, with some exceptions, like bills. Forms are complex and the templates are often unstable, due to the injection of advertising, extra conditions, or document merging. Our scenario deals with insurance forms used by brokers in Switzerland. Here, each combination of insurer, insurance type and language results in a new document layout, leading to a few hundred document types. To help brokers extract data from policies, we developed a new labeling method, called FLIE (form labeling for information extraction). FLIE first assigns a document to a cluster, grouping by language, insurer, and insurance type. It then labels the layout. To produce training data, the user annotates a sample document by hand, adding attribute names, i.e. provides a mapping. FLIE applies machine learning to propagate the mapping and extracts information. Our results are based on 24 Swiss policies in German: UVG (mandatory accident insurance), KTG (sick pay insurance), and UVGZ (optional accident insurance). Our solution has an accuracy of around 84–89%. It is currently being extended to other policy types and languages.