Questions about the sequence length dynamic adjustment about nag-bert HOT 7 CLOSED

Hi, thanks for the coming soon source code.
I have two questions about the sequence length dynamic adjustment.

1. I got the point that you use two consecutive [eos]s to indicate the end of the sequence. But at the intermediate of the sequence, it is still possible to generate a single [eos], e.g., I ate an [eos] apple [eos] [eos], and you need to remove all these intermediate [eos]s, is this correct?

   1. If this is true, then why do you need two [eos]s instead of a single [eos]? You have mentioned "Once the decoded trajectory enters the [eos] state, the state transition term in S(X, Y_0) will be dominated by the transition score term t([eos], [eos])", so the point here is to make [eos] a black hole? Once decoding trajectory transits to [eos], it will not have a chance to get out? If this is correct, then why not simply set all [eos] -> non-[eos] transitions very negative weights and do not update them during training?

2. At the training stage, say the target sequence is I ate an apple and the length of the source sequence is 9, which of the following do you use to train the model as the target?

   1. I ate an apple [eos] [eos]
   2. I ate an apple [eos] [eos] [eos] [eos] [eos]

Hope I can get your reply, and thanks~

Friend, I also have this question. Have you figured out this question later? I need your help.
Recently, there was a paper in 2021acl that used this method to do a task of correcting Chinese grammar. He was even more outrageous. The length of the target sentence was known by default during the test, and I was completely confused.

from nag-bert.

@clearloveclearlove
No, I am still waiting for the reply from the authors.
By the way, which ACL2021 do you mean? I am curious about why you called it "outrageous"

from nag-bert.

@clearloveclearlove
No, I am still waiting for the reply from the authors.
By the way, which ACL2021 do you mean? I am curious about why you called it "outrageous"

i mean the paper <>. same architecture bert+crf for Grammatical Error Correction.
the code supply by the author when he test, he direct use the information of target length, so ...

from nag-bert.

@clearloveclearlove
No, I am still waiting for the reply from the authors.
By the way, which ACL2021 do you mean? I am curious about why you called it "outrageous"

i mean the paper <>. same architecture bert+crf for Grammatical Error Correction.
the code supply by the author when he test, he direct use the information of target length, so ...

Tail-to-Tail Non-Autoregressive Sequence Prediction for ChineseGrammatical Error Correction

from nag-bert.

Hi, thanks for the coming soon source code. I have two questions about the sequence length dynamic adjustment.

1. I got the point that you use two consecutive [eos]s to indicate the end of the sequence. But at the intermediate of the sequence, it is still possible to generate a single [eos], e.g., I ate an [eos] apple [eos] [eos], and you need to remove all these intermediate [eos]s, is this correct?

   1. If this is true, then why do you need two [eos]s instead of a single [eos]? You have mentioned "Once the decoded trajectory enters the [eos] state, the state transition term in S(X, Y_0) will be dominated by the transition score term t([eos], [eos])", so the point here is to make [eos] a black hole? Once decoding trajectory transits to [eos], it will not have a chance to get out? If this is correct, then why not simply set all [eos] -> non-[eos] transitions very negative weights and do not update them during training?

2. At the training stage, say the target sequence is I ate an apple and the length of the source sequence is 9, which of the following do you use to train the model as the target?

   1. I ate an apple [eos] [eos]
   2. I ate an apple [eos] [eos] [eos] [eos] [eos]

Hope I can get your reply, and thanks~

Hello, sorry for my very late reply... During training, we use this configuration I ate an apple [eos] [eos]. Because we found that if we append many [eos] tokens as I ate an apple [eos] [eos] [eos] [eos] [eos], the model parameters will be overwhelmed by the occurance of [eos] token and it only learns to generate [eos] token as well. In practice, the generation of sequences like 'I ate an [eos] apple [eos] [eos]' are possible, but putting two [eos] tokens in training could reduce this phenomenon. Feel free to ask follow up questions! Sorry for my late reply again.

from nag-bert.

@clearloveclearlove
No, I am still waiting for the reply from the authors.
By the way, which ACL2021 do you mean? I am curious about why you called it "outrageous"

i mean the paper <>. same architecture bert+crf for Grammatical Error Correction. the code supply by the author when he test, he direct use the information of target length, so ...

I know this paper too. The paper is from one of co-author of NAG-BERT. I did not look into details of his paper. I can help you ask him about the details if you need.

from nag-bert.

Thanks~

from nag-bert.