Feature Engineering Techniques for N L P Problems?

The idea was, we have a somantic space, the universe of concepts,. It has a probablistic structure behind it. When you see a new concept with partially observable properties, let's say that if in this instance we see that wonkeymunk collocated with hairy cute. So these contexts give you some information about some of his but you don't have a lot. The question is, how do you come up with that inference? Of wamkemock, it reiterated for the convenience of the reader. And then in terms of building out your model, are there any sort of feature engineering techniques or tokenization steps that you implement before really getting the m l pieces going