Extracting structured data from invoices

Extracting structured data from invoices Xavier Holt * Sypht xavier@sypht.com Andrew Chisholm * Sypht andy@sypht.com Abstract Business documents encode a wealth of information in a format tailored to human consumption – i.e. aesthetically disbursed natural language text, graphics and tables. We address the task of extracting key fields (e.g. the amount due on an invoice) from a wide-variety of potentially unseen document formats. In contrast to tradi- tional template driven extraction systems, we introduce a content-driven machine- learning approach which is both robust to noise and generalises to unseen docu- ment formats. In a comparison of our ap- proach with alternative invoice extraction systems, we observe an absolute accuracy gain of 20% across compared fields, and a 25%–94% reduction in extraction latency. 1 Introduction To unlock the potential of data in documents we must first interpret, extract and structure their con- tent. For bills and invoices, data extraction enables a wide variety of downstream applications. Ex- traction of fields such as the amount due and biller information enable the automation of in- voice payment for businesses. Moreover, extrac- tion of information such as the daily usage or supply charge as found on an electricity bill (e.g. Figure 1) enables the aggregation of usage statistics over time and automated supplier switch- ing advice. Manual annotation of document con- tent is a time-consuming, costly and error-prone process (Klein et al., 2004). For many organi- sations, processing accounts payable or expense claims requires ongoing manual transcription for * Authors contributed equally to this work Figure 1: Energy bill with extracted fields. verification of payment, supplier and pricing in- formation. Template and RegEx driven extrac- tion systems address this problem in part by shift- ing the burden of annotation from individual doc- uments into the curation of extraction templates which cover a known document format. These ap- proaches still necessitate ongoing human effort to produce reliable extraction templates as new sup- plier formats are observed and old formats change over time. This presents a significant challenge – Australia bill payments provider BPAY covers 26,000 different registered billers alone1. We introduce SYPHT – a scaleable machine- learning solution to document field extraction. SYPHT combines OCR, heuristic filtering and a su- pervised ranking model conditioned on the con- tent of document to make field-level predictions that are robust to variations in image quality, skew, orientation and content layout. We evaluate sys- tem performance on unseen document formats and compare 3 alternative invoice extraction systems on a common subset of key fields. Our system achieves the best results with an average accuracy of 92% across field types on unseen documents and the fastest median prediction latency of 3.8 seconds. We make our system available as an API2 – enabling low latency key-field extraction scal- able to hundreds of document per second. 1www.bpay.com.au 2www.sypht.com Xavier Holt and Andrew Chisholm. 2018. Extracting structured data from invoices. In Proceedings of Australasian Language Technology Association Workshop, pages 53−59. 2 Background Information Extraction (IE) deals broadly with the problem of extracting structured …

Content truncated to 500 words