thu-bpm / mr2

Obtain the Dataset

• Our dataset is currently being used for an international big data competition (https://www.ikcest.org/bigdata2023). You can find the complete dataset, along with guidance, at the following link: https://aistudio.baidu.com/datasetdetail/230144. The competition provides both training and validation sets which can aid in training your models. Please note that the full test set will be made available after the competition concludes.
• Update: We have release our full dataset in https://drive.google.com/file/d/14NNqLKSW1FzLGuGkqwlzyIPXnKDzEFX4/view?usp=sharing .

Usage

You can find the baselines and guidance at https://aistudio.baidu.com/datasetdetail/230144.

Notebook baseline: https://aistudio.baidu.com/projectdetail/6371316?sUid=2739660&shared=1&ts=1689818284921 baseline: https://aistudio.baidu.com/clusterprojectdetail/6549711

Data

Organization Structure

The organization structure of the dataset is as follows

.
├── dataset_items_test.json
├── dataset_items_train.json
├── dataset_items_val.json
├── test
├── train
└── val

We have merged the English and Chinese datasets together and you can split them if necessary

Labels 0, 1, 2 represent non-rumor, rumor, and unverified, respectively.

File Format

Images for each claim are stored in the img folder.

The img_html_news folder contains web pages and images retrieved based on the caption of each claim. The folder includes a direct_annotation.json file containing the following information:

{
      "img_link": "Link to the retrieved related image",
      "page_link": "Link to the retrieved web page",
      "domain": "Domain of the retrieved web page",
      "snippet": "Brief summary of the retrieved web page",
      "image_path": "Path to the retrieved image",
      "html_path": "Path to the retrieved web page",
      "page_title": "Title of the retrieved web page"
}

The inverse_search folder contains web pages found based on the images in the claims. The folder includes an inverse_annotation.json file containing the following information:

{
"entities": "Entities in the image of the claim", 
"entities_scores": "Scores of the entities in the image of the claim", 
"best_guess_lbl": "The most likely description of the image in the claim", 
"all_fully_matched_captions": "", 
"all_partially_matched_captions": "",
"fully_matched_no_text": "",
//The above three fields are values of the found web pages, which are a list. Each element in the list is a dictionary, formatted as follows:
	{
	"page_link": "Link to the retrieved web page", 
	"image_link": "Link to the retrieved image", 
	"html_path": "Path to the retrieved web page", 
	"title": "Title of the retrieved web page"
	}
}

Dataset Copyright

Our datasets are open sourced for commercial and academic use and licensed under CC-BY-SA 4.0 license.

mr2's Issues

Propagation based model source code

How can I reproduce the results reported against the propagation based baseline models? I cannot find the necessary source code for this part please help out asap. Additionally please provode the script to collect and store conversational threads for the english/chinese dataset. Thanks in advance.

Miss the JSON files in the downloaded datasets.

The full JSON files are not provided in the project.

Dataset is difficult to use

The dataset contains a large number of html and npz files, which are not clearly described as well as corresponding claim textual content, pictures, and dissemination structures. Would it be possible to provide a more detailed description of the dataset and how to use it? Thanks!

Miss the claim sentences of the source posts in the downloaded dataset.

Miss the claim sentences of the source posts (only the images and searching results). The train/test/val split of the dataset is missing.