Skip to content

Latest commit

 

History

History
146 lines (108 loc) · 8.65 KB

readme.md

File metadata and controls

146 lines (108 loc) · 8.65 KB

FewRel Dataset, Toolkits and Baseline Models

Our benchmark website: https://thunlp.github.io/fewrel.html

FewRel is a large-scale few-shot relation extraction dataset, which contains more than one hundred relations and tens of thousands of annotated instances cross different domains. Our dataset is presented in our EMNLP 2018 paper FewRel: A Large-Scale Few-Shot Relation Classification Dataset with State-of-the-Art Evaluation and a following-up version is presented in our EMNLP 2019 paper FewRel 2.0: Towards More Challenging Few-Shot Relation Classification.

Based on our dataset and designed few-shot settings, we have two different benchmarks:

  • FewRel 1.0: This is the first one to incorporate few-shot learning with relation extraction, where your model need to handle both the few-shot challenge and extracting entity relations from plain text. This benchmark provides a training dataset with 64 relations and a validation set with 16 relations. Once you submit your code to our benchmark website, it will be evaluated on a hidden test set with 20 relations. Each relation has 100 human-annotated instances.

  • FewRel 2.0: We found out that there are two long-neglected aspects in previous few-shot research: (1) How well models can transfer across different domains. (2) Can few-shot models detect instances belonging to none of the given few-shot classes. To dig deeper in these two aspects, we propose the 2.0 version of our dataset, with newly-added domain adaptation (DA) and none-of-the-above (NOTA) detection challenges. Find our more in our paper and evaluation websites FewRel 2.0 domain adaptation / FewRel 2.0 none-of-the-above detection

Citing

If you used our data, toolkits or baseline models, please kindly cite our paper:

@inproceedings{han-etal-2018-fewrel,
    title = "{F}ew{R}el: A Large-Scale Supervised Few-Shot Relation Classification Dataset with State-of-the-Art Evaluation",
    author = "Han, Xu and Zhu, Hao and Yu, Pengfei and Wang, Ziyun and Yao, Yuan and Liu, Zhiyuan and Sun, Maosong",
    booktitle = "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing",
    month = oct # "-" # nov,
    year = "2018",
    address = "Brussels, Belgium",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/D18-1514",
    doi = "10.18653/v1/D18-1514",
    pages = "4803--4809"
}

@inproceedings{gao-etal-2019-fewrel,
    title = "{F}ew{R}el 2.0: Towards More Challenging Few-Shot Relation Classification",
    author = "Gao, Tianyu and Han, Xu and Zhu, Hao and Liu, Zhiyuan and Li, Peng and Sun, Maosong and Zhou, Jie",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)",
    month = nov,
    year = "2019",
    address = "Hong Kong, China",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/D19-1649",
    doi = "10.18653/v1/D19-1649",
    pages = "6251--6256"
}

If you have questions about any part of the paper, submission, leaderboard, codes, data, please e-mail [email protected].

Contributions

For FewRel 1.0, Hao Zhu first proposed this problem and proposed the way to build the dataset and the baseline system; Ziyuan Wang built and maintained the crowdsourcing website; Yuan Yao helped download the original data and conducted preprocess; Xu Han, Hao Zhu, Pengfei Yu and Ziyun Wang implemented baselines and wrote the paper together; Zhiyuan Liu provided thoughtful advice and funds through the whole project. The order of the first four authors are determined by dice rolling.

Dataset and Pretrain files

The dataset has already be contained in the github repo. However, due to the large size, glove files (pre-trained word embeddings) and BERT pretrain checkpoint are not included. Please use the script download_pretrain.sh to download these pretrain files.

We also provide pid2name.json to show the Wikidata PID, name and description for each relation.

Note: We did not release the test dataset for both FewRel 1.0 and 2.0 for fair comparison. We recommend you to evaluate your models on the validation set first, and then submit it to our evaluation websites (which you can find above).

Training a Model

To run our baseline models, use command

python train_demo.py

This will start the training and evaluating process of Prototypical Networks in a 5-way 5-shot setting. You can also use different args to start different process. Some of them are here:

  • train / val / test: Specify the training / validation / test set. For example, if you use train_wiki for train, the program will load data/train_wiki.json for training. You should always use train_wiki for training and val_wiki (FewRel 1.0 and FewRel 2.0 NOTA challenge) or val_pubmed (FewRel 2.0 DA challenge) for validation.
  • trainN: N in N-way K-shot. trainN is the specific N in training process.
  • N: N in N-way K-shot.
  • K: K in N-way K-shot.
  • Q: Sample Q query instances for each relation.
  • model: Which model to use. The default one is proto, standing for Prototypical Networks. Note that if you use the PAIR model from our paper FewRel 2.0, you should also use --encoder bert --pair.
  • encoder: Which encoder to use. You can choose cnn or bert.
  • na_rate: NA rate for FewRel 2.0 none-of-the-above (NOTA) detection. Note that here na_rate specifies the rate between Q for NOTA and Q for positive. For example, na_rate=0 means the normal setting, na_rate=1,2,5 corresponds to NA rate = 15%, 30% and 50% in 5-way settings.

There are also many args for training (like batch_size and lr) and you can find more details in our codes.

Inference

You can evaluate an existing checkpoint by

python train_demo.py --only_test --load_ckpt {CHECKPOINT_PATH} {OTHER_ARGS}

Here we provide a BERT-PAIR checkpoint (trained on FewRel 1.0 dataset, 5 way 1 shot).

Reproduction

BERT-PAIR for FewRel 1.0

python train_demo.py \
    --trainN 5 --N 5 --K 1 --Q 1 \
    --model pair --encoder bert --pair --hidden_size 768 --val_step 1000 \
    --batch_size 4  --fp16 \

Note that --fp16 requires Nvidia's apex.

5 way 1 shot 5 way 5 shot 10 way 1 shot 10 way 5 shot
Val 85.66 89.48 76.84 81.76
Test 88.32 93.22 80.63 87.02

BERT-PAIR for Domain Adaptation (FewRel 2.0)

python train_demo.py \
    --trainN 5 --N 5 --K 1 --Q 1 \
    --model pair --encoder bert --pair --hidden_size 768 --val_step 1000 \
    --batch_size 4  --fp16 --val val_pubmed --test val_pubmed \
5 way 1 shot 5 way 5 shot 10 way 1 shot 10 way 5 shot
Val 70.70 80.59 59.52 70.30
Test 67.41 78.57 54.89 66.85

BERT-PAIR for None-of-the-Above (FewRel 2.0)

python train_demo.py \
    --trainN 5 --N 5 --K 1 --Q 1 \
    --model pair --encoder bert --pair --hidden_size 768 --val_step 1000 \
    --batch_size 4  --fp16 --na_rate 5 \
5 way 1 shot (0% NOTA) 5 way 1 shot (50% NOTA) 5 way 5 shot (0% NOTA) 5 way 5 shot (50% NOTA)
Val 74.56 73.09 75.01 75.38
Test 76.73 80.31 83.32 84.64

Proto-CNN + Adversarial Training for Domain Adaptation (FewRel 2.0)

python train_demo.py \
    --val val_pubmed --adv pubmed_unsupervised --trainN 10 --N {} --K {} \ 
    --model proto --encoder cnn --val_step 1000 \
5 way 1 shot 5 way 5 shot 10 way 1 shot 10 way 5 shot
Val 48.73 64.38 34.82 50.39
Test 42.21 58.71 28.91 44.35