It seems to me one of the problems with machine learning in the nlp domain is that language concepts are mutable, but at varying degrees. In the dog/cat example

It seems to me one of the problems with machine learning in the nlp domain is that language concepts are mutable, but at varying degrees. In the dog/cat example used in the original post, the degree of mutability is very low, given the concept of a dog and a cat are rooted in the physical world. However, consider more abstract human concepts or language that is new and changing often. Ironically, much of the language used to describe AI falls into this category (and thus subject to confusion among humans). Any sort of machine learning algorithm would need to include some sort of 'adaptability' parameter that could tell the machine when to discard the current concept of the word and try forming a new one. This would need to be based on checks in both immediate context of the phrase, and related phrases. Disclaimer: My knowledge of machine learning is limited to passive reading, so this may already be a part of any nlp algorithm, or I'm just completely off base. So please consider my comments are coming from the perspective of an outsider!