Automation training set best practices
Automation training set based on historical data
Creation of Automation training set based on historical data takes big effort from development side. In this case rule-based approach of tagging will be implemented.
The tagging in this situation should be thoroughly verified. Historical data is usually normalized while the data in documents is not and can have OCR mistakes - that should be taken into account while making up the logic for rule-based approach.
The code for rule-based approach use case: gold-data-tagging-tool-package.zip
Machine Learning for training 300+ fields
In case of ML for training set 300+ fields, it's advisable to create several manual tasks up to 50 fields each (usually 20 is preferable, but all depends on the use-case and complexity of the fields).
Each field is trained during 12-13 hours on the standard infrastructure that corresponds to the platform requirements. It's advised to train fields one by one (not concurrently), therefore it will take considerable time to train 300+ fields.