Amazon Augmented AI features

Amazon Augmented AI (Amazon A2I) is integrated with Amazon Textract for document processing and Amazon Rekognition for content moderation, so you can implement human review workflows for these use cases with just a few clicks in the Amazon A2I console or a few API parameters. The Amazon A2I API also allows you to integrate your workflows into custom models that you’ve built with Amazon SageMaker or other machine learning tools. For more information, see the Amazon A2I developer guide.

Amazon A2I supports multiple choices for human reviewers. You can use your private team of reviewers for in-house review jobs, especially when handling sensitive data that needs to stay within your organization. If you want to scale up to a large number of reviewers and your data does not contain confidential or personal data, you have access to an on-demand 24x7 workforce of over 500,000 independent contractors worldwide, through Amazon Mechanical Turk. Mechanical Turk is a crowdsourcing marketplace that connects your review jobs with a distributed workforce who can perform these tasks virtually. Alternatively, you can use a third-party workforce vendor through the AWS Marketplace. These vendors have been screened by AWS to provide high-quality reviews and follow security processes. AWS Marketplace provides all the relevant details including pricing and customer reviews to help you select the best vendor for your needs.

With Amazon A2I, you provide instructional guidance to human reviewers to help ensure consistency. These detailed instructions are available to reviewers within their review interface. You can update these instructions at any time, which makes it easy to add more detail to tasks where reviewers often commit mistakes or to adjust instructions based on evolving needs. A sample instruction is shown below.

Amazon A2I provides built-in workflows that route predictions to reviewers and take the reviewers step by step through their tasks. The conditions under which workflows route predictions to reviewers can be either a confidence threshold or a random sampling percentage. If you specify a confidence threshold, the workflow routes only those predictions that fall below the threshold for human review. You can adjust these thresholds at any time to achieve the right balance between accuracy and cost-effectiveness. Alternatively, if you specify a sampling percentage, the workflow routes a random sample of the predictions for human review. This can help you implement model audits to monitor the model’s accuracy regularly. Workflows also provide reviewers a web interface consisting of all the instructions and tools they need to complete their tasks. Amazon A2I provides built-in workflows for text extraction and image moderation use cases.

You can also build custom workflows by providing an AWS Lambda function that you write to tell Amazon A2I when to trigger human reviews, and a web interface that you create using one of the over 60 available HTML templates or from scratch. For more information on workflows, see the Amazon A2I developer guide.

The image below shows the reviewer interface for content moderation with an example task to identify potentially unsafe or inappropriate content on images. The interface allows you to specify clear instructions to help reviewers complete their tasks.

The image below shows the reviewer interface for form extraction, which enables you to extract key-value pairs from document images or online forms. The interface allows you to specify clear instructions to help reviewers complete their tasks.

The image below shows a custom interface through which reviewers are asked to identify the sport being played in the picture. The interface allows you to specify clear instructions to help reviewers complete their tasks.

You can use multiple workers in reviews to increase the confidence level of the results. When defining an Amazon A2I workflow, you can specify the number of workers per review, and Amazon A2I routes each review to that many reviewers.