A previous project of mine attempted to mine text corpuses (news articles, wikipedia) for what *wasn't* said

https://www.thewinnower.com/papers/3619-data-mining-for-the-taboo-searching-for-what-isn-t-there . Initial results were promising but more development is needed, which is where this proposal comes in. In particular I would suggest adding a topic model to the algorithm outlined in the link above.