中文RoBERTa-Base文本分类模型集

模型说明

本系列包含5个基于UER-py微调的中文RoBERTa-Base分类模型，该框架在此论文中提出。此外，这些模型也可通过腾讯预训练框架TencentPretrain进行微调，该框架继承UER-py以支持十亿级参数模型，并扩展为多模态预训练框架（详见此论文）。

您可以通过以下两种方式下载这5个中文RoBERTa-Base分类模型：

• 访问UER-py模型库页面
• 通过HuggingFace获取（链接如下）：

使用方式

您可以直接使用pipeline进行文本分类（以roberta-base-finetuned-chinanews-chinese为例）：

>>> from transformers import AutoModelForSequenceClassification,AutoTokenizer,pipeline
>>> model = AutoModelForSequenceClassification.from_pretrained('uer/roberta-base-finetuned-chinanews-chinese')
>>> tokenizer = AutoTokenizer.from_pretrained('uer/roberta-base-finetuned-chinanews-chinese')
>>> text_classification = pipeline('sentiment-analysis', model=model, tokenizer=tokenizer)
>>> text_classification("北京上个月召开了两会")
    [{'label': 'mainland China politics', 'score': 0.7211663722991943}]

训练数据

使用了5个中文文本分类数据集：

• 京东全量、京东二分类和大众点评数据集包含不同情感极性的用户评论
• 凤凰新闻和中国新闻数据集包含不同主题类别的新闻首段文本

这些数据由Glyph项目收集，更多细节详见对应论文。

训练流程

模型在腾讯云上通过UER-py微调完成。我们在预训练模型chinese_roberta_L-12_H-768基础上，以512的序列长度微调了3个epoch。每个epoch结束时，当在开发集上取得最佳性能时保存模型。不同模型采用相同的超参数。

以roberta-base-finetuned-chinanews-chinese为例：

python3 finetune/run_classifier.py --pretrained_model_path models/cluecorpussmall_roberta_base_seq512_model.bin-250000 \
                                   --vocab_path models/google_zh_vocab.txt \
                                   --train_path datasets/glyph/chinanews/train.tsv \
                                   --dev_path datasets/glyph/chinanews/dev.tsv \
                                   --output_model_path models/chinanews_classifier_model.bin \
                                   --learning_rate 3e-5 --epochs_num 3 --batch_size 32 --seq_length 512

最后将预训练模型转换为Huggingface格式：

python3 scripts/convert_bert_text_classification_from_uer_to_huggingface.py --input_model_path models/chinanews_classifier_model.bin \
                                                                            --output_model_path pytorch_model.bin \
                                                                            --layers_num 12

引用信息

@article{liu2019roberta,
  title={Roberta: A robustly optimized bert pretraining approach},
  author={Liu, Yinhan and Ott, Myle and Goyal, Naman and Du, Jingfei and Joshi, Mandar and Chen, Danqi and Levy, Omer and Lewis, Mike and Zettlemoyer, Luke and Stoyanov, Veselin},
  journal={arXiv preprint arXiv:1907.11692},
  year={2019}
}

@article{zhang2017encoding,
  title={Which encoding is the best for text classification in chinese, english, japanese and korean?},
  author={Zhang, Xiang and LeCun, Yann},
  journal={arXiv preprint arXiv:1708.02657},
  year={2017}
}

@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}
}

@article{zhao2023tencentpretrain,
  title={TencentPretrain: A Scalable and Flexible Toolkit for Pre-training Models of Different Modalities},
  author={Zhao, Zhe and Li, Yudong and Hou, Cheng and Zhao, Jing and others},
  journal={ACL 2023},
  pages={217},
  year={2023}