nianlonggu / memsum Goto Github PK

I don't see where PGM used in model ? Thanks guys! @nianlonggu

oracle result of pubmed_trunc in your paper is 45.12/20.33/40.19. but I download the dataset you provided ,
but I got oracle result:
1 ROUGE-1 Average_R: 0.31187 (95%-conf.int. 0.30779 - 0.31588)
1 ROUGE-1 Average_P: 0.56634 (95%-conf.int. 0.56157 - 0.57099)
1 ROUGE-1 Average_F: 0.37910 (95%-conf.int. 0.37539 - 0.38300)

1 ROUGE-2 Average_R: 0.13605 (95%-conf.int. 0.13329 - 0.13893)
1 ROUGE-2 Average_P: 0.25674 (95%-conf.int. 0.25167 - 0.26153)
1 ROUGE-2 Average_F: 0.16707 (95%-conf.int. 0.16387 - 0.17027)

1 ROUGE-L Average_R: 0.27617 (95%-conf.int. 0.27245 - 0.27985)
1 ROUGE-L Average_P: 0.50572 (95%-conf.int. 0.50079 - 0.51033)
1 ROUGE-L Average_F: 0.33679 (95%-conf.int. 0.33319 - 0.34051)

I would appreciate it if you could provide the calculation code.Thanks.

Hi, I am implementing a use case where an algorithm can sequentially generate the sentence that is most relevant to the summary of the text. It's pretty much similar to what the greedy_summary is designed for.

However, when I am looking through your code implementation of greedy_summary (I haven't had time to look at paper yet), I noticed that it needs ground truth summary as input to pick the next most important sentence. I am thinking what if I use memsum to generate summary of a text first and feed that output as the parameter for greedy_summary, would that generate decent result?

作者您好，这种格式high_rouge_indices_and_scores.jsonl的数据是如何处理得到的,没有找到indices和score的来源

does ExtractionContextDecoder actually mean ExtractionContextEncoder？

How do you choose a model checkpoint to use for testing or simply the final checkpoint of that model?

Hi @nianlonggu ,

Thanks for providing the code.

I wanted to try your method but found some files are missing.

For example, based on your README file, it should be a file called summarizers.py but I could not find it.

In addition, is there a file called download_and_load_word_embedding.py? I am not able to find it as well.

I really appreciate any help you can provide.

• Hey guys, can u share to instruction for custom data from preprocessing data -> train -> inference
  Thanks guys !

Could you kindly provide the preprocessing code for the GovReport dataset ? Or, some descriptions about how to do the preprocessing. Thanks

Hello! I'd like to give MemSum a try, however I don't have access to machine that I can use for training yet. By any chance, are there any checkpoints that I can use for inference? Thank you!

作者您好，我从nianlong/memsum-word-embedding下载到了您预先训练好的模型，现在的问题是：如何在自己的中文数据集上训练vocabulary_200dim.pkl 和 unigram_embeddings_200dim.pkl。 希望您能回复我的消息

Hi @nianlonggu ,

Thanks a lot for your previous reply.

I just have some questions regarding your data format for PubMed.

For example, for train_PUBMED.jsonl, the keys are ['text', 'summary', 'indices', 'score'], but for the others like test_PUBMED.jsonl, we only have ['summary', 'text', 'sorted_indices'].

What are these keys mean? for example, sorted_indices and indices and score?

Also I can see test_PUBMED_ORACLE.jsonl, what is the difference between it and test_PUBMED.jsonl.

How did you generate these files?

If I understand correctly, during your training, you need to extract the gold extractive summaries and then during validation and test, your scores are reported by comparing with the real summaries.

Sorry for so many questions.

Thanks in advance.

Hey guys @nianlonggu , I want to show rouge1/2/L score a summary from a docs !

I guess sentence_score_history in file summarizers.py in model MemSum , I guess right ?

Thanks guys

Looking at the paper and data is not very clear how you have obtained the oracle summaries.

In the paper you say that B=2 and at each step you sample 1 greedy selection and 1 sub optimal selection. But that means the number of summaries should be a power of 2? But I see that's not the case?

I would appreciate it if you could clarify this selection process or provide some code?