As business grow, there is a commensurate growth in the number of documents that need to be analyzed. Traditionally, businesses had to manually review documents as they came in, but now, this process can be automated with help of Machine Learning and Artificial Intelligence technologies in Amazon Web Services.
This solution uses Amazon Textract, Amazon Comprehend and Amazon A2I to deploy an end-to-end document analysis solution. This solution takes documents in the for of scanned images as input performs the following steps on them:
- Amazon Textract to extract the text from the uploaded images
- Recreate the document (in text format) using an AWS Lambda Function
- Use Amazon Comprehend to analyse the text and perform Custom Named Entity Recognition on the extracted text
- Use Amazon Augmented AI (Amazon A2I) to send the extract entities and the text to a human reviewer for a human review for any missing entities
- Collate the results and feedback provided by the human reviewer and retrain the model every time new custom entities are detected.
In order to setup the required resources for this solution, you need to execute the Cloudformation Template that is provided with the solution. Once you have completed the prerequisites listed below, you should be able to deploy the Cloudformation Template in your account.
There are 2 prerequisite steps that you need to perform before you can deploy this solution to your AWS account and here are the steps to do them:
-
Create an S3 Bucket called S3BucketNamePlaceholder in the region that you intend to deploy this solution.
Your bucket name should be different than the one above and unique in the region but this tutorial will use S3BucketNamePlaceholder when it references to this bucket.
-
Create a private workforce from the Amazon Sagemaker console.
-
Create a Custom Worker Template from Amazon Augmented AI console. Use the file /code/ui/translate_template.html as the custom template.
-
Create a Flow Definition from the Amazon Augmented AI console.
- Use S3BucketNamePlaceholder for S3 Bucket.
- For IAM role, choose 'Create a new role'.
- For Task Type, choose 'Custom'
- For Template, choose the custom worker template you created in step 4 above.
- For Worker Types, choose 'Private'.
- For Private Teams, choose the team you created in step 3 above.
- After creation, copy the Flow Definition ARN. We will use it later.
-
Go to S3 console, and create a folder called comprehend-data in S3BucketNamePlaceholder Bucket.
-
Copy the contents of the folder
./comprehend-cer-resources/in the comprehend-data folder. -
Go to Amazon Comprehend and train a custom entity recognizer using the files you just copied in comprehend-data folder.
There are several parameters that you would need to deploy the Cloudformation template to your account. This section lists those parameters so that you understand them and get prepared for deploying the Cloudformation Template.
Use the same bucket that you have used so far and enter S3BucketNamePlaceholder in this box.
The dataset that will be used for training the Amazon Comprehend Custom Entity Recognizer. For this example, a single file contains all the training data, one document per line. An example document is available in `./comprehend-cer-resources/raw_text.csv'
This document contains the list of Custom Entities as shown here. An example file can be found in `./comprehend-cer-resources/entity_list.csv'
Train a Custom Entity Recognizer using the documents above as shown here. Once you have trained the Custom Entity Recognizer, you can use the ARN as a parameter to the Cloudformation template.
Use the task template available at ./ui/task-template.html to create a Custom Human Review workflow as shown here. Once the workflow becomes available to use, the ARN of this workflow can be used as FlowDefinitionARN parameter.
Provide any random string to be a bucket name (which needs to be unique in the region). This string will be used as the bucket name for storing the data generated by Amazon Comprehend temporarily. You don't need to create this bucket. The Cloudformation template will create it and manage it for you.
Create an S3 Bucket by the name CFN-RESOURCES-123456789 (replace CFN-RESOURCES-123456789 with your own bucket name) to host your Cloudformation resources.
Package the Cloudformation Template and code for the Lambdas using the following command:
aws cloudformation package --template-file ./source/Textract-Comprehend-A2I.yaml --output-template-file ./source/CFN-Output-Textract-Comprehend-A2I.yaml --s3-bucket CFN-RESOURCES-123456789 --region us-east-1 --force-upload
Now that you have packaged your code you can run the following command to deploy the Cloudformation Template.
aws cloudformation deploy --template-file ./source/CFN-Output-Textract-Comprehend-A2I.yaml --region us-east-1 --capabilities CAPABILITY_IAM --stack-name <INSERT STACK NAME HERE> --parameter-overrides S3BucketName=<INSERT PARAMETER HERE> FlowDefinitionARN=<INSERT PARAMETER HERE> CustomEntityRecognizerARN=<INSERT PARAMETER HERE> S3ComprehendBucketName=<INSERT PARAMETER HERE> CustomEntityTrainingListS3URI=<INSERT PARAMETER HERE> CustomEntityTrainingDatasetS3URI=<INSERT PARAMETER HERE>
Go to the S3 Bucket, S3BucketNamePlaceholder, and create a new folder titled input. Whenever you will upload a .jpg, or a .png file in this folder, the workflow will begin.
Log in to your A2I Review Console and make an desired changes.
At the end of each day, a Cloudwatch Event invokes a Lambda function automatically that looks for any new custom entities that the human reviewers may have identified using the Amazon A2I worker portal. If there are any identities that did not exist in the entity list file that was used to train the model, then it retrains the Amazon Comprehend Custom Entity model and uses the new model for inference thereafter.
This automated retraining process, allows the model to improve perpetually and requires lesser human intervention over time, hence saving time and cost for your business.
See CONTRIBUTING for more information.
This library is licensed under the MIT-0 License. See the LICENSE file.