-
Paper
-
Prior projects
Rank | Accuracy(%) | SS | RTE |
---|---|---|---|
1 | 60.36 | org | MPE(noised)-CLS |
2 | 58.72 | org | SPE-CLS |
3 | 56.16 | org | MPE(noised)-RGS |
4 | 54.06 | org | MPE-RGS |
5 | 52.58 | org | SPE-RGS |
6 | 50.88 | org | MPE-CLS |
7 | 48.82 | KNN (k=5) | SPE-CLS |
8 | 43.54 | KNN (k=5) | MPE(noised)-CLS |
9 | 40.01 | KNN (k=5) | MPE(noised)-RGS |
10 | 37.72 | KNN (k=5) | MPE-CLS |
git clone https://github.com/OnedayOneyeah/factcheck-ko-2023.git
cd factcheck-ko-2023
conda create -n factcheck-ko-2023 python=3.8
conda activate factcheck-ko-2023
pip install -r requirements.txt
Download the data for training here. Save the files in ./data/
data/wiki_claims.json
: Human-Annotated Dataset for the Factcheckdata/train_val_test_ids.json
: Lists of claim ids for train/validation/test split
-
Documents are available here.
-
Save SS checkpoint (download) in
ss/checkpoints
-
Save RTE checkpoint (download) in
rte/checkpoints
-
Save additional RTE checkpoints in
new_rte/checkpoints
: (SPE-RGS), (MPE-RGS), (MPE-CLS)[Descriptions]
- SPE: single premise entailment approach
- MPE: multiple premises entailment approach
- CLS: classifcation model
- RGS: regression model
-
Make sure you download the pretrained checkpoints for SS, RTE model.
-
Run demo.py, i.e.,
python demo.py
Make sure your gpu is available. -
Test the model with your own claim.
-
For evaluation, run eval_pipeline.py, i.e.,
eval_pipeline.py
python eval_pipeline.py --dr_pipeline <id> --ss_pipeline <id> --rte_pipeline <id>
-
Various combinations can be implemented as followed:
python eval_pipeline.py --dr_pipeline 2 --ss_pipeline 0 --rte_pipeline 2
1. DR
-
The pipelines are loaded from
dr/document_retrieval.py
oreval_pipeline.py
-
SimpleDR
andSimpleDR2
reduce time for document retrieval.Id Model Description 0 DocumentRetrieval Loading wiki document titles and texts using wiki API 1 SimpleDR Using pre-retrieved wiki document texts 2 SimpleDR2 Using pre-retrieved wiki document titles and texts
2. SS
-
The pipelines are loaded from
pipelines/ss_org.py
andpipelines/ss_knn.py
respectively.Id Model Description 0 org Unigram similarity approach 1 KNN K-nearest neighbors apporach
3. RTE
-
The pipelines are loaded from
pipelines/rte.py
Id Model Accuracy(%) 0 SPE-CLS 64.67 1 MPE(noised)-CLS 76.79 2 MPE(noised)-RGS 67.60 3 SPE-RGS 54.93 -
You can remove noise by
--remove_noise
option.