3 ms·
GPT-3 and similar approaches are diametrically opposed to how humor works. They try to find predictable continuations, while the it's unpredictability which is
by pmoriarty 5y ago
GPT-3 and similar approaches are diametrically opposed to how humor works. They try to find predictable continuations, while the it's unpredictability which is the hallmark of comedy.
- PartiallyTyped 5y agoSo then the objective function is the least predictable from the set of all possible logical continuations. So given sufficiently large corpus of logical and reasonable continuations, selecting something from the lower end of the probability distribution yields comedy, no?
- pmoriarty 5y ago"selecting something from the lower end of the probability distribution yields comedy, no?" Unfortunately, it's not so easy, or anything random would be funny, but we know that's not likely. Though unpredictability is key, comedy also has some other characteristics, like (depending on the style of comedy) like wit or a commentary on something familiar that is shown in a new light. Some types of comedy (like caricature) exaggerate characteristics, or play on someone's suffering (the old slip on a bannana peel gag), sometimes it breaks taboos or says uncomfortable truths, etc... There's a lot of literature out there analyzing what makes something funny. Maybe there's some way of integrating some of these insights in to a NLP system, or have it focus on some of those.
- PartiallyTyped 5y agoI have a hunch that it may be possible to learn those with an RL model or Contextual Bandit.
- kgeist 5y ago>it's unpredictability which is the hallmark of comedy. I think it's more about making connections the audience weren't aware could exist, but actually do exist, or can potentially exist. And GPT3 is pretty good at finding such connections. Complete nonsense isn't usually funny. It's always grounded in reality.