I'm not certain, but to do that optimally I think is an np-complete problem. If you put some restrictions on it: all files are the same format or only N fixed formats or context free grammars then it's manageable. Otherwise it boils down to natural language processing and imperfect, but useful algorithms in that domain will probably help.

Are you sure it's context sensitive grammar rather than a collection of N context free grammars? Many file format standards can be expressed as context free grammars.

If you are trying to figure out something like "what is the unknown XML structure" based on an example file or files then you are firmly inside natural language processing.