About releasing the implementation of DESEL

[liangyingshao]

Hello! Thank You so much for sharing your awesome work!
I've been trying to find the implementation code for your proposed decoding strategy DESEL in "Rethinking Model Selection and Decoding for Keyphrase Generation with Pre-trained Sequence-to-Sequence Models". Would you please release the code for DESEL and the checkpoint of the one2one model you used for DESEL? Alternatively, the code to finetune a one2one model would also be helpful.
Thanks!

[xiaowu0162]

Hi @liangyingshao ,

Thank you for the interest. Unfortunately we have lost the checkpoint. You can directly train a new one by processing the data for one2one and run https://github.com/uclanlp/DeepKPG/blob/main/sequence_generation/seq2seq/run_bart_large.sh. We will update the implementation for the decoding experiments later due to limited bandwidth. However, the implementation of DESEL can be reduced to directly calling BartScore (https://github.com/neulab/BARTScore) with the one2one model for scoring the likelihood of each of the phrases, which is straightforward.

[liangyingshao]

Thank you! Looking forward to your updates.