A system is involved in dialog with a human partner. The system has a model of a problem domain and knows mappings between words an concepts. The human has its own model of the problem domain and mappings between words an concept. The goal of the system is to learn this user specific model.
Given a user utterance with unknown words, the system can use a probabilistic model to map these words to likely concepts. It then can query the user to see if these hypothetical labels are correct or wrong. Finally the system uses this feedback to build a user specific domain model that extends its own model, which is used to optimize future interaction with the same user.
I first thought that this resembles an active learning scenario, because the system can interactively query the user to label data. But it seems that most definitions of active learning require that the system can actively choose the unlabeled data that it wants to label. In this scenario the system is reduced to a passive role because the interaction initiative will come from the human user and the system cannot change the subject of discussion on its own.
My question is how should i describe this approach? Is it a form of simplified active learning or is there a more fitting machine learning scenario available?