Hi! When will the codes be released?

[Elenore1997]

Hi, I have read your paper and really intresting in your work! I want to ask that if the codes will be released in the future here? Because I am curious about how to represent the "word distribution" mentioned in your paper. Another question is that after calculating three scalar values of KL divergence, do we just concat them to the vector of network outputs directly? I think maybe the dimension of them is very different (3 vs. 768 if pretrained models). Thanks in advance!

[sdeva14]

Thanks for your interest! Yes, of course, codes should be shared. The problem is that our server is not available for the hardware problems :( It is an unexpected situation for us to lose our main server for more than two months. As soon as it comes back, we "will definitely" share the codes as we did for all our previous work.

Meanwhile, I would like to respond to your question.

1. Word distribution: the word distributions are simply frequency distribution, thus we implemented it using Counter in Python. Then, KL divergence values can be calculated between the word distribution of an input essay and of the essays in different qualities.
2. Concatenating: yes, just concatenating scalar values to the vector representation on the same dimension, then it will be 768+3=771. It is a very simple technique that but works somehow as previous work in other domains shows a few years ago. This work also has done two years ago.

The key message of our work is, we should focus on considering content, not just applying a new machine learning technique without intuitions on how it works. We prove that this simple method, following this direction, even outperforms more complicated previous work.

I will let you know once we share the codes. Please let me know if you have any questions :)

[sdeva14]

Hi, I found codes in the backup server, and upload them now. If you have any problems, please let me know so that I can check the consistency of the codes in the main server later!

[Elenore1997]

Thanks for your interest! Yes, of course, codes should be shared. The problem is that our server is not available for the hardware problems :( It is an unexpected situation for us to lose our main server for more than two months. As soon as it comes back, we "will definitely" share the codes as we did for all our previous work.

Meanwhile, I would like to respond to your question.

1. Word distribution: the word distributions are simply frequency distribution, thus we implemented it using Counter in Python. Then, KL divergence values can be calculated between the word distribution of an input essay and of the essays in different qualities.
2. Concatenating: yes, just concatenating scalar values to the vector representation on the same dimension, then it will be 768+3=771. It is a very simple technique that but works somehow as previous work in other domains shows a few years ago. This work also has done two years ago.

The key message of our work is, we should focus on considering content, not just applying a new machine learning technique without intuitions on how it works. We prove that this simple method, following this direction, even outperforms more complicated previous work.

I will let you know once we share the codes. Please let me know if you have any questions :)

Thanks for your quick reply! I agree with you that in AES task, we should try to model the text quality from different aspects, such as coherence(SKIPFLOW, HierCoh), structures(CNN-LSTM-ATT), semantics(like BERT but not work so good in AES)... Well, it is quite challenging to design networks/methods to represent these implicitly, and harder in some special settings (like other domain settings or languages, because ASAP dataset is quite limited). And thanks again for sharing the codes! But we are working on another project right now, maybe I will try to run the codes later this month when our gpu server is free!