There's a variant of LDA called "structured LDA" that allows you to introduce a linear combination of "structural variables" into the topic distribution prior. If you wanted to make this project "language-aware", might be a good use case for sLDA.