search

MarkWuNLP / MultiTurnResponseSelection

This repo contains our ACL 2017 paper data and source code
720 stars 189 forks source link

Tensorflow版本输入错误 #21

Closed yunjianyang closed 6 years ago

yunjianyang commented 6 years ago

在tensorflow版本中，actions从来没有进入到模型训练这一环节。总是把true_utt给response的placeholder。

MarkWuNLP commented 6 years ago

self.response_ph: np.concatenate([true_utt[low:low + n_sample]] + negs, axis=0)

negs就是action。。。

yunjianyang commented 6 years ago

true_utt[low:low + n_sample] 这个不是positive 的action吧，如果我没理解错的话。 On Tue, Apr 24, 2018 at 11:27 PM MarkWuNLP notifications@github.com wrote:

self.response_ph: np.concatenate([true_utt[low:low + n_sample]] + negs, axis=0)

negs就是action。。。

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/MarkWuNLP/MultiTurnResponseSelection/issues/21#issuecomment-384149736, or mute the thread https://github.com/notifications/unsubscribe-auth/AI0xmsL8joiIl2fF3iQcZv_P3u8vaT5Xks5tr-0ZgaJpZM4TW9-2 .

MarkWuNLP commented 6 years ago

true_utt 是正例，negs是负例

yunjianyang commented 6 years ago

那action和true_utt的区别是什么呢？谢谢！

MarkWuNLP commented 6 years ago

action是所有回复的全集，true_utt是原始message的回复，也就是m_i对应r_i，而negs是随机的一个r_j，让m_i对应r_j。也就是true。在数据上action和true_utt是一样的，如果不考虑顺序，但是用的时候由于action里面的utterance被随机抽取了，所以达到了负采样的效果

yunjianyang commented 6 years ago

I see. Thanks for the explanation!