NLIDB

I think instead of manual scoring, I will feed the ft, dt, and pt into a classifier. maybe or maybe not. It will definitely reduce the coding required. pretrained embeddings may work. maybe also silence the ml guys that my system is old school. I can actually convert the sequence modeling problem into a multi class classifier problem. Not the best idea because the sequence(ngram) component will be lost. But it will be weakly supervised. It will only be column name. Of course, I will need to clean the data first and get the column values right first.