Recognizing text and using it to increase compression ratios is possible. I believe that's what this 1974 paper is about:

https://www.semanticscholar.org/paper/A-Means-for-Achieving-...