Hi, thanks for your great work. I am wondering have you ever tried this general idea to other NLG tasks like dialogue or NMT? Hoping to get some insights from you guys !

Hi, yixinL7 Very beautiful work. Do you mind sharing your code about candidate generation?

I try to use the BartForConditionalGeneration model from huggingface to reproduce your results but it always generates repeat sentences in a beam.

Thanks!

Thanks for your insightful work. However, I am confused by some details in the Table 1. Is the model which derives the 'Max' result also trained by contrastive learning ? or simply sampled from different beam search process?

Thanks for your excellent work. I have a question about loss computation. Is there any difference between the loss function in the paper and the code part? The loss function in the pape:

But the code part:

It seems the code part just computes +ilamda instead of +(j-i) @lamda. Did I miss something?

Hello there, first of all thank you so much for giving your code as open source so others like me can learn from it. I saw that the preprocess.py script requires many file inputs including the candidate summary. But those are generated by the model right? I couldn't find those in the data. Also in the example json file, I noticed that article untokenized and tokenized both seem to be sentence tokenized. So what is the difference?

Thank you for the nice research and the paper is pretty readable and very intuitive.

But when I read the code some question came to my mind.

https://github.com/yixinL7/SimCLS/blob/1f08d260dce0668241e9d2fb9eed57cc6b0e60f2/model.py#L7-L10

I think the variable TotalLoss at line number 10 will always be 0. Ofcourse it would be change from the next line by the candi and summary score. Is it for just referencing before for loop?

Thanks

Hi,I am a newbee on NLP, and I don't know what's the raw data of cnn/dm and xsum dataset.I looked up those two links and have no idea about that,I'am glad if you can tell me more,thanks!

Follow your instruction, I've reproduced your results in dataset. So can you give me some instruction to use this SimCLS to summary document . Now, it has a version in pytorch and huggingface but I really want to use directly your repo (the official repo)

Hi, since the model used for CNNDM is Facebook/bart-large-cnn, which means the model actually got fine-tuned on the CNNDM training set. Considering the Neural model's amazing capacity for memorization, the candidate generation of training set for evaluation model should be nearly perfect. Do I understand this correctly ? How do you avoid this to generate useful data for ranking ? And does Pegasus also fine tuned on the CNNDM before generating summary candidate ? Thanks .

Hi, thank you for your nice work. I'm trying to rebuild your model. But I have a question, After I run the train part, "python main.py --cuda --gpuid [list of gpuid] -l", I got some config.txt files. Then when I run Evaluate part, "python main.py --cuda --gpuid [single gpu] -e --model_pt [model path]", It needs the model path, but I have no idea where it comes from, should It comes from the training part? But I only got some config and log files in the cache folder. Appreciate it if you could give me some help.

Hello, I'm trying to train and evaluate this model on a new dataset (SAMSum). To do that I need to first generate and score candidate summaries using the generative model. Where is the code that was used to do this? The README for this repo seems like it starts with a dataset that already includes candidate summaries and their rogue scores. I just want to make sure I do that step correctly. Thanks!

Hi， I have a question about BS and MS metrics. In BS metrics, I get a score of 0.88 when I use ‘rescale_WITH_baseline =True’. With ‘rescale_WITH_baseline =False’, the score drops to around 0.46. Both results are different from yours. I'am glad if you can tell me more,thanks!