Best practices for Amazon Textract Custom Queries

Amazon Textract lets you customize the output of its pretrained Queries feature by training and using an adapter for its base model. With Amazon Textract Custom Queries, you can use your own documents and train an adapter to customize the base model, while keeping control over your proprietary documents. When creating queries for your adapters, use the following best practices. For more information about adapters, see Custom Queries.

1. The sample data should contain layout variations and image variations that are representative of the documents to be used in production.

2. A minimum of five samples (per query level) are required for either training or testing. The more samples used for training, the better.

3. The annotated Queries should have a variety of answers. For example, if the answer to a query is 'Yes' or 'No,' the annotated samples should have occurrences of both 'Yes' and 'No,'

4. Composed queries should have logical meanings and structures. For example, to extract a payee name on the check, the query should be ‘Who is the payee?’ and not ‘What is the name?’.

5. If a value that needs to be extracted has an associated key, include the key in the query for better results. If there are keys that occur multiple times, use hierarchical questions such as ‘What is the first name under borrower?’ and ‘What is the first name under co-borrower?’

6. Maintain consistency in annotation style. If a field appears in multiple places on a document and you decide to annotate only one occurrence, follow the same guideline and annotate only one occurrence in all documents. Alternatively, if you choose to annotate all occurrences of a value, follow the same process for all documents.

7. Maintain consistency while annotating fields with spaces. If a field's value is ‘123 456 78’ and if you choose to annotate it as ‘12345678’ follow the same guideline for all documents. Do not change between annotating the value as ‘123 456 78’ and ‘12345678.’

8. Annotate the data across all pages within a document even though you might run inference on only a few of the pages. If annotating all the pages is not feasible, split the document and use only the pages that you are able to annotate in your dataset.

9. Custom Queries does not support questions related to signatures (such as, 'Is the document signed?'). To know if a document is signed, use the signatures feature in your API request.

10. When correcting responses from pre-labeling, you can make small modifications to the text presented in the documents, such as correcting a few character errors, or removing spaces and punctuation. But if text isn't shown in the documents, don't enter it as a response.

11. Use the exact query used in training for inference.

12. Provide pre-labeling results only for the source-ref page. If you have pre-labeling results with Blocks that aren't from the source-ref page, it might impact performance.