Document processing with DocAI

Introduction

The current deployment of OGRRE uses Google DocAI for digitizing documents.

This page describes the workflow used to process well records through DocAI.

Terminology

Following are terms used in this page:

Term	Synonyms	Definition
model	AI model	Trained (DocAI) model that can transform inputs to outputs
collated documents	Raw PDFs	Multiple-page PDFs containing well records
splitter		AI model that can extract the PDF pages from collated documents
category		A label that identifies a type of document
training document		Document used to train an AI model
testing document		Document used to assess the accuracy of the AI model as part of the model training
classifier		AI model that groups documents into categories for processing
schema		Set of attributes that are expected for a document in a given category
field	attribute	One named value in a document

Workflow

Below is a diagram showing an overview of the workflow:

Split

Typical input documents are multiple page PDFs, only some of which contain the data of interest. Different categories of report document have different boundaries, so they are labeled and identified.

Training

The training set is created with the following manual (human) steps, performed on the collated raw PDFs uaing the Google DocAI web interface:

1. Identify the document boundaries
2. Assign a category, based on titles or content, to separate different collated document types

Google DocAI needs at least 10 training documents and 2 testing documents for each category to automatically apply the categories and boundaries to the rest of the collated documents.

Application

This 'model' of document boundaries and categories can now be applied to the rest of the collated documents. Then the processing workflow produces files like: {original file name}_{Splitter label}_{Label count}.

• A useful original file name is the API number (e.g., if not inside the file itself)
• The "Label countitle: Usage Guide sidebar_position: 2 t" tracks the number of documents with the same “Splitter Label” for a given collated Document

Classify

Once the documents have been split so that only the pages with the desired data are present, they need to be grouped into categories based on document layout.

Training

Classifier category labels are based on the order of when each document type was identified, using letters as follows: {Splitter Label}_{A,B,C,..}. Each document receives one label.

The training set is created by labeling document types. As for the Split step, Google DocAI needs at least 10 training and 2 testing documents for each document type.

Application

The classifier model takes as input the splitter outputs from the previous step and categorizes them into different schema label "bins", which are mapped to different Google buckets for extraction.

Extract

The final step in the workflow is to extract the data from each category of documents. To do this, one must first train the processor, adding schema (field) labels for every field in the category of documents.

Training

Unlike the Split and Classify steps for a field to included in a model, Google DocAI needs the field to be included on at least 10 training and 10 testing documents to create a model. One model is trained for each category identified in the previous Classify step.

Some important things to keep in mind for this process:

• Field labels matching the document text reduces incorrect auto-labeling.
• Field labels may also be included into tables to facilitate grouping tabular data.
• Unless otherwise required, the occurrence setting for fields inside of tables should be once.
• Documents that have a field labels with the occurrence setting required will be excluded from training if that field cannot be detected.
• Field labels that may occur more than once should have the occurrence setting multiple.
• Keeping a record of the field labels, label type, occurrence, field groupings, and page order is best practice.
  • When training new models recording the fields that are disabled.

Below are details on the labeling process in Google DocAI:

• Upon opening a document, Google attempts to auto-label any unconfirmed field colored in purple.
• Confirm field bounding boxes that are in the correct position.
• Edit field bounding boxes that are in the wrong position.
• Add field bounding boxes that are missed.
• Bounding boxes must include detected “text”, Document with empty bounding boxes will be excluded from training.
• Extracted text in the labeling process should not be edited.
  • The text reader is not part of the model training process.
  • Bounding boxes with edited text will be treated as incorrect detections during model testing in correctly reducing the accuracy of the model.

Application

The completed models are applied, in batch mode, with a script that is tied to the OGRRE tool. This process involves 2 steps:

1. Before running the script, create a mapping between the trained model and the documents for which it is intended (which correspond to a Google bucket)
2. The script takes this mapping and runs the Google DocAI processor to generate digitized data for each, storing the results in the tool database.

The combination of the document image (PDF) and digitized data are used by OGRRE to provide the interface for reviewing and correcting the digitized data.