metadata

language: Chinese
widget:
  - text: 小王在哪上学？
    context: 小王在北京上学，他今年二十岁。

Chinese RoBERTa Base Model for QA

Model description

The model is used for extractive question answering. You can download the model from the link roberta-base-chinese-extractive-qa.

How to use

You can use the model directly with a pipeline for extractive question answering:

>>> from transformers import pipeline
>>> path = 'uer/roberta-base-chinese-extractive-qa'
>>> nlp = pipeline('question-answering', model=path, tokenizer=path)
>>> QA_input = {'question': "小王在哪上学？",'context': "小王在北京上学，他今年二十岁。"}
>>> nlp(QA_input)
    {'score': 0.7618623375892639, 'start': 3, 'end': 5, 'answer': '北京'}

Training data

Training data contains three datasets ,including cmrc2018, webqa and 莱斯杯.

Training procedure

The model is fine-tuned by UER-py on Tencent Cloud TI-ONE. We fine-tune three epochs with a sequence length of 512 on the basis of the pre-trained model chinese_roberta_L-12_H-768.

python3 run_cmrc.py --dataset_path lyric_dataset.pt \
                    --pretrained_model_path models/cluecorpussmall_roberta_base_seq512_model.bin-250000 \
                    --vocab_path models/google_zh_vocab.txt \
                    --train_path extractive_qa.json \
                    --dev_path datasets/cmrc2018/dev.json \
                    --output_model_path models/extractive_qa_model.bin \
                    --learning_rate 3e-5 --batch_size 32 --epochs_num 3 \
                    --embedding word_pos_seg --encoder transformer --mask fully_visible

Finally, we convert the pre-trained model into Huggingface's format:

python3 scripts/convert_roberta_extractive_qa_from_uer_to_huggingface.py --input_model_path extractive_qa_model.bin \
                                                                         --output_model_path pytorch_model.bin \
                                                                         --layers_num 12

BibTeX entry and citation info

@article{zhao2019uer,
  title={UER: An Open-Source Toolkit for Pre-training Models},
  author={Zhao, Zhe and Chen, Hui and Zhang, Jinbin and Zhao, Xin and Liu, Tao and Lu, Wei and Chen, Xi and Deng, Haotang and Ju, Qi and Du, Xiaoyong},
  journal={EMNLP-IJCNLP 2019},
  pages={241},
  year={2019}
}