Closed JaniceXiong closed 2 years ago
Hey @JaniceXiong ,
We append a special token (e.g. PURPOSE) at the beginning of the source text to indicate the facet of summary to generate. And we train a single model to generate all facets.
Best, Rui
Thanks a lot @memray But how about the "section tags" at the begining of each training input? I read the paper "Tldr: Extreme summarization of scientific documents" which describes CATTS. CATTS means a scaffold task to predict titles indicated by control codes. How to use CATTS in your model? I want to build the model described in your papar, and these details seem important :)
Oh I only mean that we use the control code to indicate the facet to output, as illustrated in the figure below. We didn't use the title generation as an auxiliary training task.
We are organizing our code for data processing and it will be available very soon. Also, our slides might be useful for clarification.
Best, Rui
Ok I got it! Thanks again for your work and reply !!! @memray
I have another question about "Full" in result tables. @memray
BART-Faceted summarize a specific facet at a time, but BART-CAT generate a long summary by concatenating all facets. When comparing these two model, did you concatenate all the facets summarized by BART-Faceted?
@JaniceXiong Yes, that's right.
@memray Thanks! By the way, do you have a plan to release the model code and pretrained model?
I am confused about "indicator" amd "section tags" described in the implementation of BART-Facet.
Could you give an example like if I want to predict the Purpose of the article, how to change the src (full or I+C) and tgt?
![image](https://user-images.githubusercontent.com/43976565/128354679-6eb59942-548e-4675-b170-0cc1a5ceef9b.png)
We train the model for purpose, method, findings, value separately, or just train a model? If only a model, how can I tell the model which facet to generate?