Show HN: Comments by Top HN Posters Analysed by IBM's Watson User Modelling API

We humans each build models kind of like these when we interact with one another, so when I ask somebody, "do you think is an agreeable person?" and they reply "sure I think he is!" they're consulting that model to provide me an answer. Humans can even do a kind of pairwise sorting on that model and tell you if person1 is more or less agreeable than person2.

However, even if our individual models may differ a bit, and the results of these kinds of questions to each other might differ a bit, there's an inherent "humanness" to the results because people generally have a pretty similar semantic understanding of what "agreeableness" means.

However, what does Watson think agreeableness means? I have no idea, nobody really knows. Watson can't really explain it. All we know is that there's a model that produces a scored (and thus rankable output) when asked to score a corpus on that model and somebody somewhere labelled that model as the "agreeableness" model, perhaps based on some heuristics or parameters that were intended to define that notion.

It's thus very hard for humans to trust scoring like this because when it doesn't make sense, it doesn't make sense for reasons that no human would have about the matter. For example, I would personally say pg is far more agreeable than I am, yet Watson scores our respective collection of comments exactly the same. I can't explain it, Watson can't explain it, and thus it feels "wrong" and now I can't trust the scores that Watson provides me.

Sorry for the confusion.