Pete suggested that rawdog's output could include "I like this" and "I don't like this" links, which would train a Bayesian filter. The output could then be sorted by how much you're likely to like an article based on the filter's output. This'd be easily doable using a rawdog plugin and a CGI script.
Pete also suggested using a Bayesian approach to detect duplicate articles, although I suspect this wouldn't be as effective as just doing edit distance calculation (like rawdog's state format converter already does).