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:

Dataset	Prompt template 1
TNEWS	下面是一则关于__的新闻：
CSLDCP	这一句描述__的内容如下：
IFLYTEK	这一句描述__的内容如下：

第一个prompt模板的微调实验结果：

The fine-tuning results for prompt template 1:

Model	TNEWS	CLSDCP	IFLYTEK
Macbert-base	55.02	57.37	51.34
Macbert-large	55.77	58.99	50.31
Erlangshen-1.3B	57.36	62.35	53.23
TCBert-base_{110M-Classification-Chinese}	55.57	58.60	49.63
TCBert-large_{330M-Classification-Chinese}	56.17	60.06	51.34
TCBert-1.3B_{1.3B-Classification-Chinese}	57.41	65.10	53.75
TCBert-base_{110M-Sentence-Embedding-Chinese}	54.68	59.78	49.40
TCBert-large_{330M-Sentence-Embedding-Chinese}	55.32	62.07	51.11
TCBert-1.3B_{1.3B-Sentence-Embedding-Chinese}	57.46	65.04	53.06

第一个prompt模板的句子相似度结果：

The sentence similarity results for prompt template 1:

	TNEWS		CSLDCP		IFLYTEK
Model	referece	whitening	reference	whitening	reference	whitening
Macbert-base	43.53	47.16	33.50	36.53	28.99	33.85
Macbert-large	46.17	49.35	37.65	39.38	32.36	35.33
Erlangshen-1.3B	45.72	49.60	40.56	44.26	29.33	36.48
TCBert-base_{110M-Classification-Chinese}	48.61	51.99	43.31	45.15	33.45	37.28
TCBert-large_{330M-Classification-Chinese}	50.50	52.79	52.89	53.89	34.93	38.31
TCBert-1.3B_{1.3B-Classification-Chinese}	50.80	51.59	51.93	54.12	33.96	38.08
TCBert-base_{110M-Sentence-Embedding-Chinese}	45.82	47.06	42.91	43.87	33.28	34.76
TCBert-large_{330M-Sentence-Embedding-Chinese}	50.10	50.90	53.78	53.33	37.62	36.94
TCBert-1.3B_{1.3B-Sentence-Embedding-Chinese}	50.70	53.48	52.66	54.40	36.88	38.48

第二个prompt模板：

For prompt template 2:

Dataset	Prompt template 2
TNEWS	接下来的新闻，是跟__相关的内容：
CSLDCP	接下来的学科，是跟__相关：
IFLYTEK	接下来的生活内容，是跟__相关：

第二个prompt模板的微调结果：

The fine-tuning results for prompt template 2:

Model	TNEWS	CLSDCP	IFLYTEK
Macbert-base	54.78	58.38	50.83
Macbert-large	56.77	60.22	51.63
Erlangshen-1.3B	57.81	62.80	52.77
TCBert-base_{110M-Classification-Chinese}	54.58	59.16	49.80
TCBert-large_{330M-Classification-Chinese}	56.22	61.23	50.77
TCBert-1.3B_{1.3B-Classification-Chinese}	57.41	64.82	53.34
TCBert-base_{110M-Sentence-Embedding-Chinese}	54.68	59.78	49.40
TCBert-large_{330M-Sentence-Embedding-Chinese}	55.32	62.07	51.11
TCBert-1.3B_{1.3B-Sentence-Embedding-Chinese}	56.87	65.83	52.94

第二个prompt模板的句子相似度结果：

The sentence similarity results for prompt template 2:

	TNEWS		CSLDCP		IFLYTEK
Model	referece	whitening	reference	whitening	reference	whitening
Macbert-base	42.29	45.22	34.23	37.48	29.62	34.13
Macbert-large	46.22	49.60	40.11	44.26	32.36	35.16
Erlangshen-1.3B	46.17	49.10	40.45	45.88	30.36	36.88
TCBert-base_{110M-Classification-Chinese}	48.31	51.34	43.42	45.27	33.10	36.19
TCBert-large_{330M-Classification-Chinese}	51.19	51.69	52.55	53.28	34.31	37.45
TCBert-1.3B_{1.3B-Classification-Chinese}	52.14	52.39	51.71	53.89	33.62	38.14
TCBert-base_{110M-Sentence-Embedding-Chinese}	46.72	48.86	43.19	43.53	34.08	35.79
TCBert-large_{330M-Sentence-Embedding-Chinese}	50.65	51.94	53.84	53.67	37.74	36.65
TCBert-1.3B_{1.3B-Sentence-Embedding-Chinese}	50.75	54.78	51.43	54.34	36.48	38.36

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

python

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

# Loading models
tokenizer=BertTokenizer.from_pretrained("IDEA-CCNL/Erlangshen-TCBert-330M-Sentence-Embedding-Chinese")
model=BertForMaskedLM.from_pretrained("IDEA-CCNL/Erlangshen-TCBert-330M-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

python

# 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-330M-Sentence-Embedding-Chinese")
model=BertForMaskedLM.from_pretrained("IDEA-CCNL/Erlangshen-TCBert-330M-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:

text

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