WikiNLI

The dataset and code for the paper Mining Knowledge for Natural Language Inference from Wikipedia Categories.

Dataset

The training and development dataset are under data/WikiNLI

In the paper we sampled 100k instances for our experiments, and the 100k version of training set can be found under data/WikiNLI/100k

Other languages

• WikiNLI constructed from Wikipedia of other languages are under data/mWikiNLI, we provide four versions, Chinese(zh), French(fr), Arabic(ar) and Urdu(ur)
• WikiNLI constructed by translating English WikiNLI to other languages are under data/trWikiNLI, we provide four versions, Chinese(zh), French(fr), Arabic(ar) and Urdu(ur)

WordNet and WikiData

• The WordNet and WikiData we used in the experiments described in the WikiNLI paper

Code

To pretrain a transformer based model on WikiNLI with the Huggingface transformers framework, use the following scripts.

python code/run_wikinli.py \
    --model_type bert \
    --model_name_or_path bert-large-uncased \
    --task_name wikinli \
    --num_train_examples 500000 \
    --do_train \
    --do_eval \
    --do_lower_case \
    --data_dir ./data \
    --max_seq_length 40 \
    --per_gpu_eval_batch_size=32   \
    --per_gpu_train_batch_size=32   \
    --learning_rate 2e-5 \
    --num_train_epochs 1.0 \
    --save_steps 3000 \
    --logging_steps 3000 \
    --eval_all_checkpoints \
    --output_dir ./saved_outputs/bert-large

python code/run_wikinli.py \
    --model_type roberta \
    --model_name_or_path roberta-large \
    --task_name wikinli \
    --num_train_examples 500000 \
    --do_train \
    --do_eval \
    --do_lower_case \
    --data_dir ./data \
    --max_seq_length 40 \
    --per_gpu_eval_batch_size=32   \
    --per_gpu_train_batch_size=32   \
    --learning_rate 1e-5 \
    --warmup_steps 1000 \
    --num_train_epochs 3.0 \
    --save_steps 3000 \
    --logging_steps 3000 \
    --eval_all_checkpoints \
    --output_dir ./saved_outputs/roberta-large 

after the model is finished pretraining, modify the saved model by removing the top linear layer.

mv [PATH]/pytorch_model.bin [PATH]/raw_pytorch_model.bin
python code/modify_saved_model.py [PATH]/raw_pytorch_model.bin [PATH]/pytorch_model.bin

A WikiNLI pretrained roberta-large model can be downloaded from https://drive.google.com/file/d/1RJgewj2TPXI2lDNuxEO1gq9aSTkdRxiZ/view?usp=sharing

To evaluate on NLI related tasks after pretraining on WikiNLI, follow the instructions of evaluating GLUE tasks by Huggingface.

Dependency

• PyTorch 1.4.0
• transformers 3.1.0

Citation

@inproceedings{chen2020mining,
      title={Mining Knowledge for Natural Language Inference from Wikipedia Categories}, 
      author={Mingda Chen and Zewei Chu and Karl Stratos and Kevin Gimpel},
      booktitle = {Findings of {EMNLP}},
      year={2020},
}