I've been flirting with different approaches to dealing with comments/weblogs. "Scoring" if you will.

There are several approaches that can be emulated, slashdot (so-so), osnews (broken), advogato (person-rating), and bayasian(non-human).

So far I've been writing simple code to see if I can decide which approach works best in practise. I like the idea of bayasian filtering - but I can't decide how to code it.

So far I've been using the AI::NaiveBayes1 module to good effect. But I cannot make it work on "live" data.

This sample shows training the model on some text:

good text: 'Debian GNU/Linux Rocks'
good text: 'Testing this filter works'
good text: 'Steve Kemp is sleepy'
bad text: 'F**kers'
bad text: 'Linux fanbois!'
bad text: 'Get a life'
bad text: 'you suck!'
bad text: 'Redhat Rules'
Now testing line 'Linux Sucks!'
        Good score: 0.279078013208562
         Bad score: 0.960268432545589
BAD

So here we train the model with three pieces of "good" text, then six pieces of "bad" text. Afterwards we give a pair of scores to a new input - and if the bad score is the greater of the two we mark a comment as "bad".

All seems good. But there is a problem. How do I seed the database with "bad" comments?

I could ask people to click on links beneath comments they feel are bad in some way - but that means I'd have to retrain the model and update the scores of all the comments.

It also doesn't avoid people corrupting the model by scoring reasonable comments badly - and bringing down the effectivenes.

I guess I still have a lot to think about - definitely not an easy thing to do. I like Advogatos model, as that is "provablye" hard to subvert - however I do not understand the mechanics well enough to implement it, even after trying several times.

I guess that University education is good for something after all ;)