News Environment Perception

This is the official repository of the paper:

Zoom Out and Observe: News Environment Perception for Fake News Detection

Qiang Sheng, Juan Cao, Xueyao Zhang, Rundong Li, Danding Wang, and Yongchun Zhu

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (ACL 2022)

PDF / Poster / Code / Chinese Video / Chinese Blog

Datasets

The experimental datasets where can be seen in dataset folder, including the Chinese Dataset, and the English Dataset. Note that you can download the datasets only after an "Application to Use the Datasets for News Environment Perceived Fake News Detection" has been submitted.

Code

Key Requirements

python==3.6.10
torch==1.6.0
transformers==4.0.0

Preparation

Step 1: Obtain the representations of posts and news environment items

Step1.1: Prepare the SimCSE model

Due to the space limit of the GitHub, we upload the SimCSE's training data by Google Drive. You need to download the dataset file (i.e., [dataset]_train.txt), and move it into the preprocess/SimCSE/train_SimCSE/data of this repo. Then,

cd preprocess/SimCSE/train_SimCSE

# Configure the dataset.
sh train.sh

Of course, you can also prepare the SimCSE model by your custom dataset.

Step1.2: Obtain the texts' representations

cd preprocess/SimCSE

# Configure the dataset.
sh run.sh

Step 2: Construct the macro & micro environment

Get the macro environment and rank its internal items by similarites:

cd preprocess/NewsEnv

# Configure the specific T days of the macro environment.
sh run.sh

Step 3: Prepare for the specific detectors

This step is for the preparation of the specific detectors. There are six base models in our paper, and the preparation dependencies of them are as follows:

Model		Input (Tokenization)	Special Preparation
Post-Only	Bi-LSTM	Word Embeddings	-
	EANN	Word Embeddings	Event Adversarial Training
	BERT	BERT's Tokens	-
	BERT-Emo	BERT's Tokens	Emotion Features
"Zoom-In"	DeClarE	Word Embeddings	Fact-checking Articles
	MAC	Word Embeddings

In the table above, there are five preprocess in total: (1) Tokenization by Word Embeddings, (2) Tokenization by BERT, (3) Event Adversarial Training, (4) Emotion Features, and (5) Fact-checking Articles. We will describe the five respectively.

Tokenization by Word Embeddings

This tokenization is dependent on the external pretrained word embeddings. In our paper, we use the sgns.weibo.bigram-char (Downloading URL) for Chinese and glove.840B.300d (Downloading URL) for English.

cd preprocess/WordEmbeddings

# Configure the dataset and your local word-embeddings filepath. 
sh run.sh

Tokenization by BERT

cd preprocess/BERT

# Configure the dataset and the pretrained model
sh run.sh

Event Adversarial Training

cd preprocess/EANN

# Configure the dataset and the event number
sh run.sh

Emotion Features

cd preprocess/Emotion/code/preprocess

# Configure the dataset
sh run.sh

Fact-checking Articles

There are two preparation for fact-checking articles:

1. Retrieve the most relevant articles for every post. Specifically, we have retrieved every post's Top10 relevant articles that should be published BEFORE the post, whose results are saved in the preprocess/BM25/data folder. If you want to learn about more implementation details, just refer to preprocess/BM25/[dataset].ipynb.
2. Tokenize the fact-checking articles by word embeddings:

cd preprocess/WordEmbeddings

# Configure the dataset and your local word-embeddings filepath. Set the data_type as 'article'.
sh run.sh

Training and Inferring

cd model

# Configure the dataset and the parameters of the model
sh run.sh

After that, the results and classification reports will be saved in ckpts/[dataset]/[model].

Citation

@inproceedings{NEP,
    title = "Zoom Out and Observe: News Environment Perception for Fake News Detection",
    author = "Sheng, Qiang  and
      Cao, Juan  and
      Zhang, Xueyao  and
      Li, Rundong and
      Wang, Danding  and
      Zhu, Yongchun",
    booktitle = "Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics",
    month = may,
    year = "2022",
    publisher = "Association for Computational Linguistics"
}