Live data from Hacker News

Analyze Your HN Posts with Watson User Modeling

hn.mybluemix.net

41–50 of 56 posts

Re: Analyze Your HN Posts with Watson User Modeling

#45

Is there any insight as to how this works? I got an Openness rating of 97% and a Harmony rating of 100%, both of which I know are not true. (I also received a Love rating of 1% under my Needs, although that's pretty accurate.)

A problem I've observed with these kind of blackbox systems is that the process from input to output really is a mystery.

When the results are right, they're just "right" so you should accept them, when they're wrong they're actually also right by whatever magical hamster wheel is operating inside of the thing and you just don't "get it".

The problem is that humans like to have some clue as to how the results were derived, something easy to explain that gets the gist across. Something like "Watson counted all the words you use and compared them to different reference lexicons to arrive at the score". This provides a little bit of context so we understand the semantics of the result and how to consider them and reason with them.

But for all we know the results we're seeing are from some arbitrary stochastic method:

openness=rand(90,99) harmony=rand(90,100)

etc.

For things like this to be accepted by the users (humans) there needs to be a quick explanation for how this works otherwise we get head scratchers.

Re: Analyze Your HN Posts with Watson User Modeling

#46

Is there any insight as to how this works? I got an Openness rating of 97% and a Harmony rating of 100%, both of which I know are not true. (I also received a Love rating of 1% under my Needs, although that's pretty accurate.)

We plan to add more information to our docs soon about the service, including a description of each of the traits, and possibly reference some of the many data sources used.

Meanwhile, you can do a search on "IBM System U" (the project's not-so-internal code name.) This particular slideshare.net prez has some great info on the methodology, validation tests and references: http://slidesha.re/1ri0vPV

Re: Analyze Your HN Posts with Watson User Modeling

#47

Not a lot of documentation on the IBM page about what the characteristics mean or how they learn a mapping from text to these categories. Curious to know what the average hacker news scores look like? I'm imagining it is a pretty small segment of "normal" society.

See my response to minimaxir above about additional documentation coming soon, and a link with good background info on the technology. Meanwhile, here are brief descriptions of the Big 5 Personality traits: Big 5 Personality: - Openness - associated with curiosity, intellect, and an appreciation for art and adventure - Conscientiousness - associated with organization and industriousness - Extraversion - associated with positive and outgoing attitudes toward other people - Agreeableness - associated with compassion and cooperation toward other people - Emotional Range - associated with a sensitivity to negative emotions

For more information on systematic associations between personality and individual differences in word use, please refer to studies like Tal Yarkoni, "Personality in 100,000 words: A Large scale analysis of personality and word use among bloggers", 2010

Re: Analyze Your HN Posts with Watson User Modeling

#49
post #20

Did this with Twitter: twurl "/1.1/statuses/user_timeline.json?count=200" | jq -r ".[] | .text" | pbcopy paste it into http://watson-um-demo.mybluemix.net/demo 87% Openness, 5% Agreeableness, that's funny.

This inspired me to do the same with reddit, so I threw together this function in zsh (should work with bash too):

        function redcom(){ ! [ -z "$2" ] && i="&after=$2" || i=""; data=$(curl -s "https://www.reddit.com/user/${1}/comments.json?count=100${i}"); j=$(jq -e -r '.data["after"]' >${1}-redcom.txt; echo $(echo -e $data|wc -l) lines; ( [ "$j" = "null" ] || [ -z "$j" ] ) || redcom $1 $j ;} ;
It's a bit more complicated due to the reddit api, but if you run that, then run redcom and it should throw all your comments into -redcom.txt

Never user/heard of jq before, it's a pretty nice tool.

Post reply on HN