geek-ai / irgan

IRGAN SIGIR paper experimental code
623 stars 243 forks source link

abount QA #19

Closed facingwaller closed 6 years ago

facingwaller commented 6 years ago

请教一下

  1. 既然QA的G产生的NEG answer来自候选里面的根据CNN模型挑选的一部分。如果把全部的候选都给D做判别效果是否会更好?只是比较花时间。 比如 从示例代码中,G的过程中,从100个里面用CNN识别后选出了5个,再给D用CNN训练这个5个。为什么不直接把这100个都拿去给D训练?(花的时间更少?)
wnzhang commented 6 years ago

When you are optimizing top-k performance (instead of the global pairwise ranking performance like AUC) you normally need to focus on the ranking of top k-2k instead of the overall ranking. Thus you need to sample top-ranked negative samples instead of using all the negative samples.

facingwaller commented 6 years ago

谢谢回答,请问为什么会只需要关注前K-2K的相关的负例?有什么理论依据吗?直觉上来说,不应该是训练了更多的负例,效果更好吗?还是只是把NN模型当做一个黑盒子,从系统的外部上发现了这个规律并不能实证?

wnzhang commented 6 years ago

You can check out this paper to see why :) http://wnzhang.net/papers/lambdarankcf-sigir.pdf

facingwaller commented 6 years ago

谢谢回答。我去学习一下这篇论文再向您请教。