zcgzcgzcg1 / mediasum Goto Github PK

What is the input format of BART training？
Thank you for your work！

Hello,
could you please tell in which order the predictions are generated?

according to train_val_test_split.json,the first prediction id should be CNN-45061 whose summary is
Report Out of Afghanistan that Taliban Supreme Leader Getting Ready to Hand Over Kandahar; Discussion with Brother of Special Forces Soldier Killed in Afghanistan,

but the first predictions you gave of BartModel is:
the summer job market is shaping up to be the weakest in more than 50 years. so how do you get your teenage kids not to spend the entire summer glued to the couch? day to day personal finance contributor michelle singletary offers some tips. on the designated day, dozens of eighth graders wore big curly wigs and chambray shirts. then they painted along to an episode of ross' pbs show the joy of painting.

Hello, I used you new released train_valid_test_split.json and output.txt. But I can not get the claimed rouge score, could you please help me out ? Thanks so much !

Hi, thanks for the great dataset!
Could you tell me the rouge version you are using? Is it files2rouge(https://github.com/pltrdy/files2rouge) which fairseq uses?

Hi Chenguang,

Where can I find the PG-Net output for the dataset?

Thanks a lot.

Is it convenient to share the source code?Thank you！！
thanks！！！！

Thank you for sharing your paper, I would like to know more about your pre-training model of UniLm+MEDIASUM.

Is it convenient to share?

Hi, I noticed that utterances are truncated after 1024 tokens as described in (Part D, Implementation Details).
I want to know do I need to add role information when con-cating utterances into one sentence? or do the 1024 tokens includes the role information? Please tell me more details about the preprocessing.
Thanks so much!

Hi there,

Looking at the MediaSum dataset, the identifiers for the speakers in the "speaker" list seems to be non-unique. E.g., NPR-7 has the speakers ['PROFITT', 'STEVE PROFFITT', 'Ms. SUSAN STURGILL', 'MADELEINE BRAND, host', 'Ms. BARBARA LEBEY', 'Mr. RANDY HALL'] where I understand 'PROFITT' and 'STEVE PROFFITT' should refer to the same person. I think this probably happens quite a few times (filtering naively for "2-person" dialogs leads to only 22020 interviews (with roughly 14,500 NPR interviews) although this dataset is meant to encompass the only NPR-based INTERVIEW dataset which includes 23,714 2-person dialogs.

Just wondering if you/anyone encountered this and developed a workaround?