This repository is for our NAACL2022 paper: Cross-document Misinformation Detection based on Event Graph Reasoning

Requirements

1. Install requirements using pip install -r requirements.txt
2. Install dgl==0.6.1 following the instructions

Dataset

Please download data from Google Drive. The file contains 3 directories (IED, TL17, Crisis), each for one dataset. Each directory is organized as follows:

{IED,TL17,Crisis}
├── split.{train,valid,test}.json    # Split and labels for training / valid / test sets
├── event_labels.json                # Labels for event-level detection
├── ltf/                             # Tokenized documents in xml format
└── merged/                          # IE outputs for clusters
    ├── cluster-0/                     # IE outputs for cluster 0 
    │   ├── edl.json                     # Entity linking results
    │   └── {entity,relation,event}.cs   # IE outputs of entity / relation / event extraction in cold-start format
    └── ...

For both document-level and event-level detection, 0 stands for real and 1 stands for fake.

Training

Specify $DATA $OUTPUT $EPOCH before you continue. $EPOCH for IED, TL17 and Crisis datasets are 120, 120 and 10 respectively.

• Train event-level detector:

python train.py $DATA $OUTPUT --accum-step 16 --lr 5e-05 --num-epochs $EPOCH --warmup 100 \
  --grad-clip 1.0 --scheduler linear-warmup --model hetero --event-detection-lambda 1.0

• Train document-level detector:

python train.py $DATA $OUTPUT --accum-step 16 --lr 5e-05 --num-epochs $EPOCH --warmup 100 \
  --grad-clip 1.0 --scheduler linear-warmup --model hetero --event-detection-lambda 0.0

• Train document-level detector using event-level features: ($EVENT_CKPT is checkpoint for the event-level detector)

python train.py $DATA $OUTPUT --accum-step 16 --lr 5e-05 --num-epochs $EPOCH --warmup 100 \
  --grad-clip 1.0 --scheduler linear-warmup --model fuse --event-detection-lambda 0.0 \
  --feature-ckpt $EVENT_CKPT

Evaluation

Specify $DATA $OUTPUT before you continue.

• Evaluate document-level detector:

python eval.py $DATA $OUTPUT/checkpoint-best.doc.pt --model hetero --set test --event-detection

• Evaluate document-level detector using event-level features: ($EVENT_CKPT is checkpoint for the event-level detector)

python eval.py $DATA $OUTPUT/checkpoint-best.doc.pt --model fuse --feature-ckpt $EVENT_CKPT --set test --event-detection

• Evaluate event-level detector. Since event-level detection data is too imbalanced, we first select threshold based on the valid set, and then evalute on the test set:

python eval.py $DATA $OUTPUT/checkpoint-best.event.pt --model hetero --set valid --event-detection --dump-best-th
python eval.py $DATA $OUTPUT/checkpoint-best.event.pt --model hetero --set test --event-detection \
  --use-th $OUTPUT/checkpoint-best.event.pt.valid.best-th.pkl