A plausible approach would be to instrument the editor to log 'edit -events'. Then one could fit/learn an appropriate probabilistic model/grammar. After that gi

A plausible approach would be to instrument the editor to log 'edit -events'. Then one could fit/learn an appropriate probabilistic model/grammar. After that given two pieces of text A and B it becomes a question of finding the most probable edits that takes you from A to B. This is far from a new idea. Age old Levenstein distance can be seen as a special case. Here the edit model is very simple. All letter addition, deletion, and substitutions are equally likely. More realistic a model one uses the better the results will look like provided one has enough data to train. One could train models for specific languages (natural as well as programming). If we have language specific syntax highlighting, why not language specific edit models. One could do personalized models, etc etc.