mumin 1.10.0

Seamlessly build the MuMiN dataset.

MuMiN-Build

This repository contains the package used to build the MuMiN dataset from the paper Nielsen and McConville: MuMiN: A Large-Scale Multilingual Multimodal Fact-Checked Misinformation Social Network Dataset (2021).

See the MuMiN website for more information, including a leaderboard of the top performing models.

---

Installation

The mumin package can be installed using pip:

$ pip install mumin

To be able to build the dataset, Twitter data needs to be downloaded, which requires a Twitter API key. You can get one for free here. You will need the Bearer Token.

Quickstart

The main class of the package is the MuminDataset class:

>>> from mumin import MuminDataset
>>> dataset = MuminDataset(twitter_bearer_token=XXXXX)
>>> dataset
MuminDataset(size='small', compiled=False)

By default, this loads the small version of the dataset. This can be changed by setting the size argument of MuminDataset to one of 'small', 'medium' or 'large'. To begin using the dataset, it first needs to be compiled. This will download the dataset, rehydrate the tweets and users, and download all the associated news articles, images and videos. This usually takes a while.

>>> dataset.compile()
MuminDataset(num_nodes=388,149, num_relations=475,490, size='small', compiled=True)

Note that this dataset does not contain all the nodes and relations in MuMiN-small, as that would take way longer to compile. The data left out are timelines, profile pictures and article images. These can be included by specifying include_extra_images=True and/or include_timelines=True in the constructor of MuminDataset.

After compilation, the dataset can also be found in the mumin-<size>.zip file. This file name can be changed using the dataset_path argument when initialising the MuminDataset class. If you need embeddings of the nodes, for instance for use in machine learning models, then you can simply call the add_embeddings method:

>>> dataset.add_embeddings()
MuminDataset(num_nodes=388,149, num_relations=475,490, size='small', compiled=True)

Note: If you need to use the add_embeddings method, you need to install the mumin package as either pip install mumin[embeddings] or pip install mumin[all], which will install the transformers and torch libraries. This is to ensure that such large libraries are only downloaded if needed.

It is possible to export the dataset to the Deep Graph Library, using the to_dgl method:

>>> dgl_graph = dataset.to_dgl()
>>> type(dgl_graph)
dgl.heterograph.DGLHeteroGraph

Note: If you need to use the to_dgl method, you need to install the mumin package as pip install mumin[dgl] or pip install mumin[all], which will install the dgl and torch libraries.

For a more in-depth tutorial of how to work with the dataset, including training multiple different misinformation classifiers, see the tutorial.

Dataset Statistics

Dataset	#Claims	#Threads	#Tweets	#Users	#Articles	#Images	#Languages	%Misinfo
MuMiN-large	12,914	26,048	21,565,018	1,986,354	10,920	6,573	41	94.57%
MuMiN-medium	5,565	10,832	12,650,371	1,150,259	4,212	2,510	37	94.07%
MuMiN-small	2,183	4,344	7,202,506	639,559	1,497	1,036	35	92.87%

Related Repositories

• MuMiN website, the central place for the MuMiN ecosystem, containing tutorials, leaderboards and links to the paper and related repositories.
• MuMiN, containing the paper in PDF and LaTeX form.
• MuMiN-trawl, containing the source code used to construct the dataset from scratch.
• MuMiN-baseline, containing the source code for the baselines.