This might not be a task for ML, especially assuming the only option would be unsupervised ML.

I would suggest using an ontology, or rolling your own from the English Wikipedia database dump, as a basis for tokenization of the menu text and go from there. What structured content exactly are you trying to extract?