Why should we use neural networks on invoice documents?
Managing incoming documents in a company can be a very labor-intensive task — especially if you are working in accounting, where still most invoices arrive on paper via post mail. Recent studies have shown that German SMEs (Small and Medium-sized Enterprises) send out an average of 1.895 invoices per month, of which 79% are still on paper. Larger companies even deliver up to 90% of their invoices on paper. Even though some companies would like to use electronic formats or offer them in addition, the lack of standardization and system compatibility impedes all efforts.
Neural networks can handle an invoice’s complexity much better than traditional rule-based systems
So, what happens with all those printed invoices when they are delivered to the recipient? Someone has to sort and read them in order to extract relevant data pieces (also called entities) and sort them into categories. In total, Hypatos has identified up to 70 different types of entities in invoices. Here are a few examples:
- total price,
- unit price,
- tax amount,
- sender address,
- sender bank account’s IBAN and BIC,
- product name and description,
- amount of units bought,
- due date for payment.
On invoices, these pieces of information are usually available in a semi-structured format: combining unstructured text descriptions and various structured tabular forms for displaying monetary amounts. This characteristic makes it quite difficult to extract information from invoices automatically: until recently, hand-crafted rule-based systems would only be able to extract a limited amount of data from half-structured documents. Such models require many hours of manual feature engineering — while only being able to process document types that are known to the system already.
Neural networks developed in the research field of Natural Language Processing (NLP) are able to extract relevant entities from various document layouts — even if the model has not seen that type of document before. This is accomplished by training such a model with thousands of documents, teaching it where to look for which piece of information. During that process, a properly designed neural network learns to extract frequent patterns automatically. These are more or less the same patterns that a rule-based system would require you to define manually in many hours of research and work. Based on these patterns, the model is able to extract relevant entities automatically — simplifying and accelerating the accounting process. Hypatos offers pre-trained machine learning models for this purpose.
There is a significant financial impact
A study commissioned by the German Department of Commerce estimates that the German public administration is processing approximately 200 million invoices per year. An accountant usually has to perform four steps with an estimated manual processing time of 16 to 23 minutes per invoice before a payment can be authorized: sorting the papers, reading their contents to extract relevant information, inserting the data into a computer system, and validating the transaction. Using neural networks, the first three steps can be automated, leaving the accountant to do their actual work: validating the correctness of each transaction and authorizing the necessary payments — enabling them to process four times as many documents in the same time.
The same study published by the Department of Commerce suggests that reducing the processing time per invoice such a significant amount offers a savings potential of up to € 5 billion — only in the public sector and only by saving time. Additionally, successful automation reduces human-generated errors, which usually is another cost factor. This is only the tip of the iceberg, indicating the enormous potential of automation in accounting using neural networks.
Conclusion
We have seen that there is a significant potential in automating information extraction from invoices, which extends to many other semi-structured documents like medical receipts, income statements, or technical specification sheets. This is why Hypatos is dedicated to build pre-trained Deep Learning models for solving this task in various use cases, achieving the goals of accelerating processing time, decreasing cost and reducing errors. Even though invoices are probably the most complex type of semi-structured documents, similar model architectures can be applied to a multitude of (semi-)structured documents, boosting productivity in many relevant administrative tasks.
