Use masks to build doc-topic matrix for each year interval, and compute word similarity by doc co-occurrence and topic distribution.