Skip to content

nguyenvo09/EMNLP2020

Repository files navigation

EMNLP2020

This is the repository to reproduce results in the paper "Where Are the Facts? Searching for Fact-checked Information to Alleviate the Spread of Fake News", EMNLP 2020.

Multimodal Attention Network

alt text

Datasets

Snopes

PolitiFact

Images data

Structure of dataset folders

After downloading and extracting data, the expected structure of formatted_data is as follows:

EMNLP2020/
├── formatted_data
│   ├── Politifact
│   │   ├── 50_candidates_bm25_extended_reranking
│   │   │   ├── Politifact.dev.tsv
│   │   │   ├── Politifact.test.tsv
│   │   │   ├── Politifact.test2_hard.tsv
│   │   │   └── Politifact.train.tsv
│   │   ├── 50_candidates_bm25_extended_reranking_and_text_in_img
│   │   │   ├── Politifact.dev.tsv
│   │   │   ├── Politifact.test.tsv
│   │   │   ├── Politifact.test2_hard.tsv
│   │   │   └── Politifact.train.tsv
│   │   ├── 50_candidates_bm25_extended_reranking_and_text_in_img_avoid_bias
│   │   │   ├── Politifact.dev.tsv
│   │   │   ├── Politifact.test.tsv
│   │   │   ├── Politifact.test2_hard.tsv
│   │   │   └── Politifact.train.tsv
│   │   ├── article_mapped.json
│   │   ├── articles_content.json
│   │   ├── elmo_features_avoid_bias
│   │   │   ├── articles_feats.pth
│   │   │   └── queries_feats.pth
│   │   ├── elmo_features_only_text_in_tweets
│   │   │   ├── articles_feats.pth
│   │   │   └── queries_feats.pth
│   │   ├── elmo_features_use_text_in_img
│   │   │   ├── articles_feats.pth
│   │   │   └── queries_feats.pth
│   │   ├── queries_content.json
│   │   ├── query.negatives
│   │   ├── query_article_interaction.csv
│   │   └── query_mapped.json
│   └── Snopes
│       ├── 50_candidates_bm25_extended_reranking
│       │   ├── Snopes.dev.tsv
│       │   ├── Snopes.test.tsv
│       │   ├── Snopes.test2_hard.tsv
│       │   └── Snopes.train.tsv
│       ├── 50_candidates_bm25_extended_reranking_and_text_in_img
│       │   ├── Snopes.dev.tsv
│       │   ├── Snopes.test.tsv
│       │   ├── Snopes.test2_hard.tsv
│       │   └── Snopes.train.tsv
│       ├── 50_candidates_bm25_extended_reranking_and_text_in_img_avoid_bias
│       │   ├── Snopes.dev.tsv
│       │   ├── Snopes.test.tsv
│       │   ├── Snopes.test2_hard.tsv
│       │   └── Snopes.train.tsv
│       ├── article_mapped.json
│       ├── articles_content.json
│       ├── elmo_features_avoid_bias
│       │   ├── articles_feats.pth
│       │   └── queries_feats.pth
│       ├── elmo_features_only_text_in_tweets
│       │   ├── articles_feats.pth
│       │   └── queries_feats.pth
│       ├── elmo_features_use_text_in_img
│       │   ├── articles_feats.pth
│       │   └── queries_feats.pth
│       ├── queries_content.json
│       ├── query.negatives
│       ├── query_article_interaction.csv
│       └── query_mapped.json
├── images_data
│   ├── full_Snopes_extracted_features.pth
│   ├── full_images_otweet_DataC_extracted_features.pth
│   ├── resnet50_Politifact_documents_extracted_features.pth
│   └── resnet50_Polititact_queries_extracted_features.pth

Usage

1. Install required packages

We use Pytorch 0.4.1 and python 3.5.

pip install -r requirements.txt

2. Download and extract images data

pip install gdown
cd EMNLP2020
gdown https://drive.google.com/uc?id=17clyyiWyMDMUl6KqrDGGZCi2ZUeNSimh
unzip images_data.zip
rm images_data.zip

If you want to see raw images, you can download it as follows:

gdown https://drive.google.com/u/0/uc?id=11sxoTJx49TBOde_xFY-fgWcG-aHNFhAp
unzip raw_images.zip

3.1 Running SC1 (Table 2 in our paper)

For Snopes

gdown https://drive.google.com/uc?id=1S_WWvU1Q1bKElJ04E3MI7z_bLzPIPw5C
unzip SC1_snopes.zip -d formatted_data/Snopes
mkdir logs
python Masters/master_man.py --attention_type=4 \
                             --conv_layers=2 \
                             --cuda=1 \
                             --use_elmo=1 --use_visual=1 \
                             --filters=256 \
                             --filters_count_pacrr=16 \
                             --fixed_length_left=50 \
                             --fixed_length_right=1000 \
                             --log="logs/man" \
                             --loss_type="hinge" \
                             --max_ngram=1 \
                             --n_s=48 \
                             --path="formatted_data/Snopes/50_candidates_bm25_extended_reranking" \
                             --query_mapped="formatted_data/Snopes/query_mapped.json" \
                             --article_mapped="formatted_data/Snopes/article_mapped.json" \
                             --left_images_features="images_data/full_images_otweet_DataC_extracted_features.pth" \
                             --right_images_features="images_data/full_Snopes_extracted_features.pth" \
                             --elmo_feats="formatted_data/Snopes/elmo_features_only_text_in_tweets"

For PolitiFact

gdown https://drive.google.com/uc?id=1zeqlv3JeBn-ygn0juTO4SWBucZXIMKZi
unzip SC1_politifact.zip -d formatted_data/Politifact
python Masters/master_man.py --attention_type=4 \
                             --conv_layers=2 \
                             --cuda=1 \
                             --use_elmo=1 --use_visual=1 \
                             --filters=256 \
                             --filters_count_pacrr=16 \
                             --fixed_length_left=50 \
                             --fixed_length_right=1000 \
                             --log="logs/man" \
                             --loss_type="hinge" \
                             --max_ngram=1 \
                             --n_s=48 \
                             --path="formatted_data/Politifact/50_candidates_bm25_extended_reranking" \
                             --query_mapped="formatted_data/Politifact/query_mapped.json" \
                             --article_mapped="formatted_data/Politifact/article_mapped.json" \
                             --left_images_features="images_data/resnet50_Polititact_queries_extracted_features.pth" \
                             --right_images_features="images_data/resnet50_Politifact_documents_extracted_features.pth" \
                             --elmo_feats="formatted_data/Politifact/elmo_features_only_text_in_tweets"

3.2 Running SC2 (MAN in Table 3 in our paper)

For Snopes dataset

gdown https://drive.google.com/uc?id=1VDtJk_C-pZtBQXon2jvp4NTyxUnDv-gY
unzip SC2_snopes.zip -d formatted_data/Snopes
python Masters/master_man.py --attention_type=2 \
                             --conv_layers=2 \
                             --cuda=1 \
                             --use_elmo=1 --use_visual=1 \
                             --filters=256 \
                             --filters_count_pacrr=16 \
                             --fixed_length_left=100 \
                             --fixed_length_right=1000 \
                             --log="logs/man" \
                             --loss_type="hinge" \
                             --max_ngram=1 \
                             --n_s=32 \
                             --path="formatted_data/Snopes/50_candidates_bm25_extended_reranking_and_text_in_img" \
                             --query_mapped="formatted_data/Snopes/query_mapped.json" \
                             --article_mapped="formatted_data/Snopes/article_mapped.json" \
                             --left_images_features="images_data/full_images_otweet_DataC_extracted_features.pth" \
                             --right_images_features="images_data/full_Snopes_extracted_features.pth" \
                             --elmo_feats="formatted_data/Snopes/elmo_features_use_text_in_img"

For Politifact dataset

gdown https://drive.google.com/uc?id=1UDPJdnawYZiicx02shywYGQ3c091Q8xW
unzip SC2_politifact.zip -d formatted_data/Politifact
python Masters/master_man.py --attention_type=2 \
                             --conv_layers=3 \
                             --cuda=1 \
                             --use_elmo=1 --use_visual=1 \
                             --filters=256 \
                             --filters_count_pacrr=16 \
                             --fixed_length_left=100 \
                             --fixed_length_right=1000 \
                             --log="logs/man" \
                             --loss_type="hinge" \
                             --max_ngram=1 \
                             --n_s=32 \
                             --path="formatted_data/Politifact/50_candidates_bm25_extended_reranking_and_text_in_img" \
                             --query_mapped="formatted_data/Politifact/query_mapped.json" \
                             --article_mapped="formatted_data/Politifact/article_mapped.json" \
                             --left_images_features="images_data/resnet50_Polititact_queries_extracted_features.pth" \
                             --right_images_features="images_data/resnet50_Politifact_documents_extracted_features.pth" \
                             --elmo_feats="formatted_data/Politifact/elmo_features_use_text_in_img"

3.3 Running SC2 with augmented data (MAN-A in Table 3 in our paper)

This test is memory-intensive so we recommend to run this test on a server with 64Gb RAM.

For Snopes dataset

gdown https://drive.google.com/u/0/uc?id=1GDONqAZ5lllmF-_XMgk4gVnJNyLP079v
unzip augment_snopes.zip -d formatted_data/Snopes
python Masters/master_man.py --attention_type=2 \
                             --conv_layers=2 \
                             --cuda=1 \
                             --use_elmo=1 --use_visual=1 \
                             --filters=256 \
                             --filters_count_pacrr=16 \
                             --fixed_length_left=100 \
                             --fixed_length_right=1000 \
                             --log="logs/man" \
                             --loss_type="hinge" \
                             --max_ngram=2 \
                             --n_s=32 \
                             --path="formatted_data/Snopes/50_candidates_bm25_extended_reranking_and_text_in_img_avoid_bias" \
                             --query_mapped="formatted_data/Snopes/query_mapped.json" \
                             --article_mapped="formatted_data/Snopes/article_mapped.json" \
                             --left_images_features="images_data/full_images_otweet_DataC_extracted_features.pth" \
                             --right_images_features="images_data/full_Snopes_extracted_features.pth" \
                             --elmo_feats="formatted_data/Snopes/elmo_features_avoid_bias"

For PolitiFact dataset

gdown https://drive.google.com/u/0/uc?id=10e1JhhbfQWYILkovaeopGuhD1VQ_ZPYc
unzip augment_politifact.zip -d formatted_data/Politifact
python Masters/master_man.py --attention_type=4 \
                             --conv_layers=2 \
                             --cuda=1 \
                             --use_elmo=1 --use_visual=1 \
                             --filters=256 \
                             --filters_count_pacrr=16 \
                             --fixed_length_left=100 \
                             --fixed_length_right=1000 \
                             --log="logs/man" \
                             --loss_type="hinge" \
                             --max_ngram=3 \
                             --n_s=48 \
                             --path="formatted_data/Politifact/50_candidates_bm25_extended_reranking_and_text_in_img_avoid_bias" \
                             --query_mapped="formatted_data/Politifact/query_mapped.json" \
                             --article_mapped="formatted_data/Politifact/article_mapped.json" \
                             --left_images_features="images_data/resnet50_Polititact_queries_extracted_features.pth" \
                             --right_images_features="images_data/resnet50_Politifact_documents_extracted_features.pth" \
                             --elmo_feats="formatted_data/Politifact/elmo_features_avoid_bias"

Citation

If you feel our paper and resources are useful, please consider citing our work as follows:

@inproceedings{vo2020facts,
	title={Where Are the Facts? Searching for Fact-checked Information to Alleviate the Spread of Fake News},
	author={Vo, Nguyen and Lee, Kyumin},
	booktitle={Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP 2020)},
	year={2020}
}

Slides of our talk at EMNLP 2020

https://slideslive.com/38938793/where-are-the-facts-searching-for-factchecked-information-to-alleviate-the-spread-of-fake-news

About

This is official Pytorch code and datasets of the paper "Where Are the Facts? Searching for Fact-checked Information to Alleviate the Spread of Fake News", EMNLP 2020.

Topics

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages