Raise your level of abstraction
I would like if someone knows some papers, good reads, tools or anything else about practices about guessing models from unstructured data?
Even if it's tightly related to your metamodel, I identified some key features (IMHO):
But you maybe have some other point of view and thoughts?
Xavier, do you mean that data and models already exist, but relation between the two needs to be validated?
You're right, what I wrote is not clear.
What I mean is as input I only have a metamodel and some unstructured (or semi structured) data.
The goal is to create models that conforms my metamodel from the data.
To achieve that, I have some heuristics (business rules actually with drools) to propose the end user some deducted choices.
I'm working on machine learning stuff to analyse when and why the end user validate and/or invalidate (i.e propose something else than the deducted choice) the deducted choices.
In the case of my metamodel, it should work not that bad, even I know that is not appliable for every metamodels.
Hope I made my point a bit clearer.
Yes, it is clear now. However, I have little experience with the topic. If the data is more like knowledge, you may found some help in the AOP (aspect oriented programming) community. A part of AOP is to create a model of a source code (data). This model has to conform to a programming language definition (metamodel). This looks like a challenge very similar to yours.
Thanks for your answer, it seems that retromodeling can be applied too. Another initiative is the Eclipse MoDisco project (http://eclipse.org/MoDisco/).
But I think it's quite hard to apply it to unstructured data since you don't have any pattern to recognize at the beginning of the process of retromodeling in that case. That's why I try to investigate the area of machine learning.