boxxfish/hover

Fork of the HoVER repo, with a focus on updated dependencies.

Hover

This repo contains the source code of the baseline models described in the following paper

• "HoVer: A Dataset for Many-Hop Fact Extraction And Claim Verification" in Findings of EMNLP, 2020. (paper).

The basic code structure was adapted from Transformers.

0. Preparation

Dependencies

• PyTorch 1.4.0/1.6.0
• See requirements.txt.

Data

• Run download_data.sh to download the HoVer dataset.

1. Document Retrieval

TF-IDF Pre-retrieval

We provide the top-100 Wikipedia articles retrieved by running DRQA on the HoVer dataset. It was already downloaded in data/hover/tfidf_retrieved.

Training Neural-based Document Retrieval Model

• Prepare the data by running:

python prepare_data_for_doc_retrieval.py --data_split=dev --doc_retrieve_range=20
python prepare_data_for_doc_retrieval.py --data_split=train --doc_retrieve_range=20

This will add the top-20 TF-IDF retrieved documents to the data as candidates of the following neural document retrieval stage.

• Run ./train_scripts/train_doc_retrieval.sh. The model checkpoints are saved in out/hover/exp1.0/doc_retrieval.

Evaluating Neural-based Document Retrieval Model

• Run the evaluation:

./eval_scripts/eval_doc_retrieval_on_dev.sh
./eval_scripts/eval_doc_retrieval_on_train.sh

This will evaluate the model on both the training set and dev set because we need both predictions to construct the training/dev set for the sentence selection.

2. Sentence Selection

Training Sentence-selection Model

• First, start the Stanford Corernlp in the background. We use Corenlp to split the sentences:

java -mx4g -cp "*" edu.stanford.nlp.pipeline.StanfordCoreNLPServer -annotators "tokenize,ssplit,pos,lemma,parse,sentiment" -port 9000 -timeout 30000

• Prepare the data by running:

python prepare_data_for_sent_retrieval.py --data_split=dev --sent_retrieve_range=5
python prepare_data_for_sent_retrieval.py --data_split=train --sent_retrieve_range=5

This will add the sentences from the top-5 retrieved documents as candidates of the following sentence selection stage.

• Run ./train_scripts/train_sent_retrieval.sh. The model checkpoints are saved in out/hover/exp1.0/sent_retrieval.

Evaluating Sentence-selection Model

• Run the evaluation:

./eval_scripts/eval_sent_retrieval_on_dev.sh
./eval_scripts/eval_sent_retrieval_on_train.sh

This will evaluate the model on both the training set and dev set because we need both predictions to construct the training/dev set for the claim verification.

3. Claim Verification

Training Claim-verification Model

• Prepare the data by running:

python prepare_data_for_claim_verification.py --data_split=dev
python prepare_data_for_claim_verification.py --data_split=train

• Run ./train_scripts/train_claim_verification.sh. The model checkpoints are saved in out/hover/exp1.0/claim_verification.

Evaluating Claim-verification Model

• Run the evaluation:

./eval_scripts/eval_claim_verification_on_dev.sh

Citation

@inproceedings{jiang2020hover,
  title={{HoVer}: A Dataset for Many-Hop Fact Extraction And Claim Verification},
  author={Yichen Jiang and Shikha Bordia and Zheng Zhong and Charles Dognin and Maneesh Singh and Mohit Bansal.},
  booktitle={Findings of the Conference on Empirical Methods in Natural Language Processing ({EMNLP})},
  year={2020}
}