How does Finmatics' artificial intelligence learn?

Finmatics uses Deep Learning and Auto Machine Learning to analyze and read out invoices.

How does Finmatics read information from documents?

Finmatics uses a combination of Deep Learning and Auto Machine Learning to capture information from invoices (invoice number, invoice date, amounts...).

Finmatics' Machine Learning technology recognizes the invoice layout (via master data, UID) and automatically "learns" a model per invoice issuer.
Corrections in the ERP systems or in the accountant cockpit automatically train the models and the system continuously improves (Auto Machine Learning). Typically, a business partner has to be trained 3-5 times to fully train the model.

The models of how to read invoice layouts are shared across all firms. More than 1,200,000 invoice layouts for multiple countries have already been learned by the system (as of April 2021)

The learning behavior of Finmatics' Auto Machine Learning is shown here as an example:

Figure 1: Learning behavior of Finmatics - Auto Machine Learning model.

The Auto Machine Learning technology has one disadvantage: If the invoice layout has not yet been trained in the Finmatics system, no values can be read out by this model.

Therefore, the Auto Machine Learning technology was combined with another model, a Deep Learning model.
The Deep Learning model is trained on all invoices in the Finmatics system and is further trained on a regular basis. The Deep Learning model of Finmatics learns general contexts similar to a human, i.e. it can generalize (also read out new, unknown invoices).

The combination of Auto Machine Learning and Deep Learning is ideal:
If the invoice creator is unknown to Finmatics, the Deep Learning model provides data.
If the invoice creator is known and learned, the Auto Machine Learning model "takes over".

 

What errors can occur when reading information from documents?

Business partner cannot be recognized

If the business partner is unknown or cannot be assigned via master data, the "Auto Machine Learning Model" cannot draw the correct model. Only the results of the Deep Learning model are displayed and the hit rate is lower.

Tip: Business partner recognition can be perfected by master data maintenance: Use the master data automation report for this and 50% higher automation by maintaining the top 100 business partners.

OCR errors

Poor scans can lead to OCR errors (e.g. confusion between the number zero (0) and the letter O. OCR errors can be reduced by a well defined scanning process (More information on proper scanning can be found in our Scanguide)

Labelling errors

Finmatics learns by searching for read-out values on the document and "marking" them automatically. Date formats, for example, are converted to 70 common formats and then searched for on the document. Likewise blanks etc. are reduced. Sometimes, however, it can happen that read values cannot be found on the document. If this is the case, no training takes place and the readout results do not improve.

Tip: Always capture the fields as they are on the document, otherwise Finmatics cannot learn.

Model is not (yet) perfectly trained.

For a small subset of documents, e.g. with special layouts or complex tax splits, neither Finmatics' Deep Learning, nor the Auto Machine Learning model provides correct results. The Finmatics Data Science team is always striving to further improve the models.