melalab / nela-gt Goto Github PK

First off, thank you so much for your efforts in making available NELA GT datasets.. I would like to extend the use case of the dataset to news classification. Does it make sense use source-level labels directly as news-level label? I understand that labelling wont be 100% accurate but would it be accurate enough to be used for fake news classification problems?

Hello,

I'm currently working with the news data from NELA-GT-2020. I'm using the labels.csv file to assign labels to the articles. However, I've encountered an issue where I can't perfectly match the sources' scores with the sources mentioned in the articles. To clarify, there are 519 sources in the dataset, but the labels.csv file, which can be found here, only contains 336 sources.

My question is, should I consider the missing sources as unlabeled?

Additionally, I've noticed that there are sources without labels that seem very similar to others that are labeled. For instance, there are sources without labels like "chicagosun-times" and "thehuffingtonpost," while there are labeled sources like "chicagosuntimes" and "huffingtonpost."

I would like to express my gratitude in advance for any help you can provide regarding this issue.
Alexandra Silva

Hello,

I have two questions regarding datasets.

1. In the labels.csv file, the labels are described in terms of news data, which includes lots of articles inside.
   Is it that all the articles included in the same news file have the same labels?
   (e.g., if the label of the abcnews in labels.csv is False, then all the articles (around 10k articles) have all the same label?)

2. In the fake news detection task, where the claim and evidence exist, what field of news data can be used as claim and evidence respectively? Is title for claim and content for evidence?

Could you give more details on how you preprocess the data? I noticed underscore characters are present instead of some special characters, for example.

It would be ideal if you could share the code you used to preprocess the data. I am comparing another dataset to NELA and I need to apply the same preprocessing steps to make sure the discriminators don't pick up preprocessing differences between the datasets.

Thank you for your help!

Great dataset, thanks for your work. I wonder if you can also provide outgoing links in the articles?It would be also interesting to see tweets that link those news besides the embedded ones in the articles.

Bests
ZP

Thanks for your work!
I'm trying to base my study on your valuable corpus and found your keywords provided to filter the two subsets in the 2021 corpus helpful.
I'm wondering if you could provide the keywords you used to filter the 2020 covid/election subsets, like the ones you provided in the 2021 version. I found the 2021 covid keywords all-encompassing while the 2021 capital riot keywords very limiting to that event.
If you still have that, could you please share the file? I'll really appreciate it:)

I have read your paper 'NELA-GT-2020: A Large Multi-Labelled News Dataset for The Study of Misinformation in News Articles' and downloaded the dataset from https://doi.org/10.7910/ DVN/CHMUYZ. I found that each article has ‘@@@@@@@’ popped up instead of normal words which brings great trouble for research. Can I have the true context of every news article? By the way, can I know what keyword dataset did you use to get the 2020 U.S. Presidential Election subset?

Hello, will you release the newer version of the dataset? This resource is very useful and having the pu to date version would be great.
Thank you!