I want to teach the machine to specifically recognize the start date of the contract. The machine already knows how to recognize dates. Is there an easy answer for this?
thank you!
yishairasowsky@gmail.com
You probably want to do this as a sentence-level annotation, rather than annotate the date specifically. If you do it as an entity-level task and try to recognise only the date, you'll teach the algorithm to focus on the contents of the date, which will probably overfit to the training date. For instance, you might have a set of contracts where the starting dates are all between 2014 and 2018. When you run your trained model on other contracts, the model won't generalise well. What you want to teach it is what sort of sentential contexts indicate the start date --- so you're better off using sentence-level annotations. You could also consider a rule-based approach, which might perform well and let you control exactly what generalisation is being made.
Thank you Matt for this nice reply!
We are trying a rule-based approach now for starters. But how will the recall and precision be very high if we do not employ machine learning?
Thanks for your time and help!!
Well, that depends on the problem! Often you can get either high precision or high recall with a rule-based approach, and sometimes you can get both. High precision comes when the terms are unambiguous. You might not get full coverage from the rules, but they're at least not wrong. In other cases, the recall is easy, but the terms are somewhat ambiguous, so the precision suffers. In this situation it's sometimes harder to make use of the rules.