Typically, AIML “training” begins with monitoring the chat logs, which consist of question-answer pairs. Whenever you see inappropriate answers, or null answer types, you start entering correct, or better ones manually. For this reason, I call AIML “artisanal’, or in other words “handmade”. I am not aware of anyone who has successfully attached machine learning to AIML chat log question-answer pairs; however, theoretically it should be possible. Attaching machine learning to AIML chat logs would not improve the content quality, or add domain expertise.
Domain expertise , also known as “”, is basically the automation of human expert knowledge, which is somewhat related to the now largely antiquated “ ” of yore. is perhaps a subset of . Applying domain expertise to AIML is more art than science. There is not a good way to apply to AIML (see also ). If you are not careful with AIML, you can add a lot of knowledge that will be inaccessible due to looping at higher levels, in effect trapping your knowledge below more simple concepts.