This question might be too general, or opinion-based, but I was looking for advice on how to extract rare words from a very large corpus. The rare words wouldn't necessarily be consistent from document to document, so traditional tf-idf wouldn't be quite right.
Specifically, I'm looking at restaurant reviews, and want to highlight reviews that use very specific language, e.g. "The steak tartare tasted like cotton candy" vs "The food was good". The fact that the sentiment was negative or positive is not as important.
Using a binary tf-idf seems to help, as it doesn't overweight the fact that a word was used multiple times in a single document. Is there any other advice as to how this can be detected?