Automated anonymization of documents is one popular research subject and finds extremely important usage in areas like medical data where privacy is taken very seriously. It is in general a difficult task, and even more so for data outside of medical domain due to the lack of specialized tools. In this post, I explored the possibility of using the vector embedding [Mikolov 13] approach to anonymize general forms of documents. I arrived at some interesting results with hints from statistical physics, which I would like share about.