Hello All,
I am trainning a length controlled abstractive summarizer and I would like to add the length of the golden summary as an additional feature for the model to learn the desired length. I saw the answer here for classification in which a numeric feature is added to the hidden state of the [CLS] token. Since my problem is Seq2Seq language generation, does anyone have an idea on how to add this extra feature?
I understand it should be added to the decoder which is the one generating the output, but if you guys know a better way or if Im incorrect please let me know.
Best regards,