Re:Natural language generation: The commercial state of the ...
by gxqoz on Monday January 11, 2021 @11:14AM
(#25730915)
I did some experiments last summer trying to get GPT-3 to author quizbowl questions, a type of trivia question. As expected, GPT-3's loose relationship with the truth made these results not acceptable. But a system that really could generate new questions from existing facts would be immensely useful in this domain.
[ Parent
| on HN ]
Re:Natural language generation: The commercial state of the ...
by polm23 on Monday January 11, 2021 @11:35AM
(#25731311)
I'm not surprised GPT is awful at that, but there are good ways to do it. Look for papers on "factoid question generation".
Re:Natural language generation: The commercial state of the ...
by gxqoz on Monday January 11, 2021 @01:36PM
(#25733673)
Quizbowl questions have a particular structure that makes them generally not work with these other approaches. For instance, it needs to generate paragraph-long questions with the clues that go from the most to least difficult. This is hard skill for a human to learn, much less a machine.