News for hackers. Stuff that matters.

A comment on We're Afraid Language Models Aren't Modeling Ambiguity. Parent comment

Re:We're Afraid Language Models Aren't Modeling Ambiguity by gwd on Friday April 28, 2023 @04:31AM (#35738903)
The problem here however is language. "I'm afraid the cat has been hit by a car" literally means that the person speaking is experiencing fear. However, colloquially in English, "I'm afraid X" is a form of telling someone something which will make them sad; e.g., "I'm afraid the shields will be quite operational" [1]. (GPT-4 thinks "hit by a car" might be meant figuratively too; that's a bit of a stretch for me, but in some contexts would certainly be possible.) It is literally impossible to determine, given only that one sentence, whether the cat was actually hit by a car or not.

[1]https://www.youtube.com/watch?v=zGwDwx10wB4

[ Parent | on HN ]
Re:We're Afraid Language Models Aren't Modeling Ambiguity by casmaxima on Friday April 28, 2023 @05:01AM (#35739043)
Perhaps the ambiguity arises because we are omitting words: I'm afraid to tell you that the cat has been hit by a car. Now the question: What are you afraid of? has a non ambiguous answer: To tell you that something awful has happened.
[ Parent | on HN ]