Erlangshen-TCBert-110M-Sentence-Embedding-Chinese

IDEA-CCNL/Erlangshen-TCBert-110M-Sentence-Embedding-Chinese

• Github: Fengshenbang-LM
• Docs: Fengshenbang-Docs

简介 Brief Introduction

110M参数的句子表征Topic Classification BERT (TCBert)。

The TCBert with 110M parameters is pre-trained for sentence representation for Chinese topic classification tasks.

模型分类 Model Taxonomy

模型信息 Model Information

为了提高模型在话题分类上句子表征效果，我们收集了大量话题分类数据进行基于prompts的对比学习预训练。

To improve the model performance on sentence representation for the topic classification task, we collected numerous topic classification datasets for contrastive pre-training based on general prompts.

下游效果 Performance

我们为每个数据集设计了两个prompt模板。

We customize two prompts templates for each dataset.

第一个prompt模板：

For prompt template 1:

第一个prompt模板的微调实验结果：

The fine-tuning results for prompt template 1:

第一个prompt模板的句子相似度结果：

The sentence similarity results for prompt template 1:

第二个prompt模板：

For prompt template 2:

第二个prompt模板的微调结果：

The fine-tuning results for prompt template 2:

第二个prompt模板的句子相似度结果：

The sentence similarity results for prompt template 2:

更多关于TCBERTs的细节，请参考我们的技术报告。基于新的数据，我们会更新TCBERTs，请留意我们仓库的更新。

For more details about TCBERTs, please refer to our paper. We may regularly update TCBERTs upon new coming data, please keep an eye on the repo!

使用 Usage

使用示例 Usage Examples

# Prompt-based MLM fine-tuning
from transformers import BertForMaskedLM, BertTokenizer
import torch

# Loading models
tokenizer=BertTokenizer.from_pretrained("IDEA-CCNL/Erlangshen-TCBert-110M-Sentence-Embedding-Chinese")
model=BertForMaskedLM.from_pretrained("IDEA-CCNL/Erlangshen-TCBert-110M-Sentence-Embedding-Chinese")

# Prepare the data
inputs = tokenizer("下面是一则关于[MASK][MASK]的新闻：怎样的房子才算户型方正？", return_tensors="pt")
labels = tokenizer("下面是一则关于房产的新闻：怎样的房子才算户型方正？", return_tensors="pt")["input_ids"]
labels = torch.where(inputs.input_ids == tokenizer.mask_token_id, labels, -100)

# Output the loss
outputs = model(**inputs, labels=labels)
loss = outputs.loss

# Prompt-based Sentence Similarity
# To extract sentence representations.
from transformers import BertForMaskedLM, BertTokenizer
import torch

# Loading models
tokenizer=BertTokenizer.from_pretrained("IDEA-CCNL/Erlangshen-TCBert-110M-Sentence-Embedding-Chinese")
model=BertForMaskedLM.from_pretrained("IDEA-CCNL/Erlangshen-TCBert-110M-Sentence-Embedding-Chinese")

# Cosine similarity function
cos = torch.nn.CosineSimilarity(dim=0, eps=1e-8)

with torch.no_grad():

    # To extract sentence representations for training data
    training_input = tokenizer("怎样的房子才算户型方正？", return_tensors="pt")
    training_output = BertForMaskedLM(**token_text, output_hidden_states=True)
    training_representation = torch.mean(training_outputs.hidden_states[-1].squeeze(), dim=0)

    # To extract sentence representations for training data
    test_input = tokenizer("下面是一则关于[MASK][MASK]的新闻：股票放量下趺，大资金出逃谁在接盘？", return_tensors="pt")
    test_output = BertForMaskedLM(**token_text, output_hidden_states=True)
    test_representation = torch.mean(training_outputs.hidden_states[-1].squeeze(), dim=0)

# Calculate similarity scores
similarity_score = cos(training_representation, test_representation)

引用 Citation

如果您在您的工作中使用了我们的模型，可以引用我们的技术报告:

If you use for your work, please cite the following paper

@article{han2022tcbert,
  title={TCBERT: A Technical Report for Chinese Topic Classification BERT},
  author={Han, Ting and Pan, Kunhao and Chen, Xinyu and Song, Dingjie and Fan, Yuchen and Gao, Xinyu and Gan, Ruyi and Zhang, Jiaxing},
  journal={arXiv preprint arXiv:2211.11304},
  year={2022}
}

如果您在您的工作中使用了我们的模型，可以引用我们的网站:

You can also cite our website:

@misc{Fengshenbang-LM,
  title={Fengshenbang-LM},
  author={IDEA-CCNL},
  year={2021},
  howpublished={\url{https://github.com/IDEA-CCNL/Fengshenbang-LM}},
}