The sentence retrival codes.

• The fact verification shared task contains three steps: Document Retrieval, Sentence Retrival and Fact Verification.
• We utilize BERT based model for the sentence retrival part.
• Our paper main compare KGAT with ESIM based sentence retrieval, which is same as GEAR because the BERT model overfit on the development set. The data can also be found in data folder.
• We use pairwise loss for sentence retrieval, which can achieve best performance.

• Data process:
  • Go to the data folder.
  • Run bash process.sh to generate pairs for training and development sets.
• Train the retrieval model:
  • Run bash train.sh to train the BERT based sentence retrieval model.

• Run bash test.sh to get the data for claim verification and top5 evidence will be reserved.
• The process_data.py aims to include the golden data for claim verification to avoid the data bias.
• Note that if no golden evidence is provided the prediction should be NOT ENOUGH INFO (Different from the golden label). To avoid this senario, we add golden evidence for training and development sets to avoid the label bias.

We have tested the performance of BERT based retrieval model with different setting. We do not write them into the paper because of the page limitation. We hope these results can help you for further studies. The codes of BERT + Prediction are same with the pretrain.

• Development set.

• Testing set.