I really like the notion of "word perplexity reduction". That phrase cracks me up, particularly since the mathematical foundation upon which it is built is sound. :)
In fact, the maximum likelihood measurement seems to be a modified version of the term-frequency inverse document frequency (TF-IDF) measure. The author repeatedly uses new names for established concepts (such as "Folding In" = "holdout set", "Vector space model" = "context vectors"), making the paper difficult to read.

However, I like the novelty of the approach: minimizing perplexity for automated indexing. Yahoo is still primarily indexed by hand, and this provides a possible technique to streamline that entire process.