Labeling instructions provision
Goal of the step: To enable SMEs tag documents unambiguously by providing tagging instructions.
Input: Data set split into batches.
Output: Precise, brief and demonstrative instructions for SMEs.
Tagging instructions should include only specific logic for fields in the particular batch — for example, how to tag the documents of supplier "Taco Inc." or secretary of state "Georgia". Specific instructions should be precise, brief, unambiguous and clear. Special logic for tagging the fields must be provided there.
It is strongly recommended to instruct SMEs separately on each batch and to assign a batch of one template to one person, for example, all the batches of "Columbia" template should be tagged only by SME A, and all the documents of "Nevada" template, by SME B. If each person can concentrate on one set of fields and similar layouts, that will result in fewer mistakes and higher speed of tagging.
First, provide the screenshot of correctly tagged document with comments, like this:
Also some important information should be specified, for example:
entity_typefield should be tagged from the row "1. Exact name of the".- Federal Employer Identification Number row title indicates
corporate_numberfield, don't confuse with "Filing Number". - Address information (street, city, state, zip, country) should be tagged from the block "5. Street address of the corporation's principal office:".
- Anything else you think an SME is going to have some doubts around.
If it's not clear from the business logic how the field must be tagged, or there are some corner cases, which require special logic, such things should be described as well:
- It is not clear where the boundaries of the field are, for example:
- It is not clear how to correct the value.
- There is some special logic when something must be selected from the list, etc.
More information about tagging requirements and high-quality data set here.