UserSim icon indicating copy to clipboard operation
UserSim copied to clipboard

Possibility to take into account the upvotes/downvotes of the bot's responses?

Open gmox opened this issue 9 years ago • 1 comments

I know this is getting into AI territory, but still.

The NLT's ability to generate coherent responses is still kinda bad. There should be some sort of way to determine based on other people's responses the way in which sentences are structured/generated.

I'm definitely willing to help do some work on this, since AI and Language are both of my interests.

gmox avatar Jul 19 '15 22:07 gmox

Color me interested.

There are two issues. The first is that upvotes take time; an hour or a day from when a comment is created would be the best time to check its score. The way the program is structured right now, there's no easy way to do that: either I'd have to have the processes stay resident in memory (which would murder my memory) or I'd have to put some sort of scheduling system into effect, and have yet another active process that does nothing but crawl through old comments and grab their scores.

The second is that I don't know how the score data should be incorporated into the models. Should it just go through every trigram in the comment, find the corresponding weights in the Markov state model, and adjust them up or down depending on how the comment's score deviates from the average? Should it be the average of all comment scores? The average of that user's comment scores? The average of scores in that sub? In that thread?

It would be difficult to do right, I think.

trambelus avatar Jul 21 '15 14:07 trambelus