OCR technology combined with the artificial intelligence developed by Notilus staff makes users’ everyday lives easier by automatically integrating data shown on the receipts in their business expenses claims. With an emphasis on continuous improvement of the user experience, our mobile solutions integrate these technological advances as new versions are released.
Simplifying the everyday life of our 2 million users so that they can truly dedicate themselves to their job is the Notilus range’s mission. At the heart of developments, this project automating expenses claims using OCR (Optical Character Recognition) technology is already enabling employees to automatically retrieve information shown on their receipts such as date, amount, type and currency. In the future, new data will be recognised with an emphasis on continuous improvement of the user experience. Laurence Aucordier, Notilus Mobility Product Manager at Dimo Software, shines a light on how this technology works.
‘Project OCR’: the desire to improve the user experience
Under the code name ‘Project OCR’, the objective is to automate the entry of employees’ expenses claims. The aim is to make the task easier by reducing as much as possible the entry of information relating to business expenses claims on their smartphone.
When a Notilus user takes a photo of their expenses receipt, the OCR system, combined with artificial intelligence, recognises a certain amount of data shown on the receipt. Notilus incorporates it automatically in the data entry form. ‘The user’s role will then be to check that the recognition is correct, instead of the tedious entry of figures. The data currently included relates to date, amount, type and currency – which are shown almost all the time – and we are in the optimisation phase for the recognition of other data such as country, VAT, service provider, etc.’, explains Laurence Aucordier. Dimo is also continuing to explore the possibilities offered by OCR, especially internationally.
Deep learning, neural network: AI as a driver of innovation
Technically, the first step in OCR involves transforming an image into text. In the second step, structured data is extracted and the AI interprets it. For the same expense entry field (for example, amount), several algorithms can be launched simultaneously.
Before initiating a development, Laurence explains that you need to launch a ‘POC’ (proof of concept) to test the model algorithms on data records. It may be that promising avenues fail, and others are better suited to the context for obtaining better results. Laurence explains: ‘There is no single solution to a problem, but several solutions to the same problem, with a cascade or parallel mode, and reliability indices’.
How does the OCR used by Notilus work?
‘Imagine a travelling sales rep fills up with fuel. He gets his petrol receipt. Recognition of ‘type’ is based on algorithms developed by engineers, and on what we call ‘deep learning’ in artificial intelligence (AI). If the system does not recognise the type of the expense, by entering the value ‘fuel’, the system will be able to learn and use this for another similar occasion. Furthermore, if the values on a receipt from a foreign country include particular semantics, but relate to fuel, this will also enrich the system in the relevant language’, explains Laurence.
Our mobile business expenses management solutions using this technology are already in production, offering increasingly simplified entry, with more and more data recognised. We set the bar very high to offer our customers the best solution on the market: only values recognised with at least 80% success are included in new versions of Notilus.
No effective recognition without quality data entry
Quality of recognition is an issue. Laurence Aucordier explains: ‘the system interprets a receipt and, according to this interpretation, it knows whether it is more or less accurate because the process is based on several algorithms in parallel. If our algorithms converge towards the same value, the reliability index will be closer to 1’. Laurence adds: ‘When you take a photo, there can be issues with blurring or illegibility of the receipt. The user receives an alert asking them to retake their photo, for example, if they moved when the photo was taken’.
Continuous quality improvement at Notilus
Developers test the recognition rate on a representative sample – a database of around 50,000 receipts – analysing their recognised values (predictions) and comparing them to the ‘real’ values. Each new development is tested to make sure it does not affect the performance of recognition of data already in production.
Laurence shines a light on what happens behind the scenes of OCR: ‘Developers have monitoring tools to check the quality of AI predictions and to ensure continuous improvement of our recognition rates. Big data is already a reality for Notilus with our 2 million users who enrich our learning every day and allow deep learning to be increasingly efficient.’