Can I build a Generic model using a custom Neural Model. Example a generic Receipt Parser Model that can Analyse Receipts no matter the template of the document?

Question

Hello,

I need to create a general Receipt Parser model, and I'm exploring options at the Form Recognizer Studio. I have learned that there are two types of custom models available: Template models, which allow for one template per model, and Neural models, which allow for 15-20 templates per model. Additionally, there is a Composed model option, which merges models into a single endpoint, with a maximum limit of 100 models that can be combined. However, for my specific use case, there could be millions of templates, despite the document type being the same. I would appreciate your assistance with this matter.

Thanks in Advance.

Answer 1

Hari Hara Sudhan KrishnaKumar Custom neural model is used for scenarios where the visual template of a form is probably not the same, but the documents could have the same information, with different page structures. For example, United States W2 forms which have the same information but can vary in appearance across companies.

If you have a similar scenario, you can start with a custom neural model. The recommendation in general is to use the custom neural model first if you are not sure which type of custom model to choose from.

Finally, if you have a very huge number of templates then the best-case scenario is to have a good dataset that covers most of your templates and train your custom model to get the best model which works for you. You can then use other template datasets to train a new model with a different id and so on to create a composed model. The composed model limit is 100 and if you exceed this limit then you will have to create another composed model and pass your form through these composed models and use the best response in your application.

I hope this helps!!

If this answers your query, do click Accept Answer and Yes for was this answer helpful. And, if you have any further query do let us know.