LayoutLM for Invoices

This is a fine-tuned version of the multi-modal LayoutLM model for the task of question answering on invoices and other documents. It has been fine-tuned on a proprietary dataset of invoices as well as both SQuAD2.0 and DocVQA for general comprehension.

Non-consecutive tokens

Unlike other QA models, which can only extract consecutive tokens (because they predict the start and end of a sequence), this model can predict longer-range, non-consecutive sequences with an additional classifier head. For example, QA models often encounter this failure mode:

Before

After

However this model is able to predict non-consecutive tokens and therefore the address correctly:

Getting started with the model

The best way to use this model is via DocQuery.

About us

This model was created by the team at Impira.