DMetaSoul/sbert-chinese-general-v1

此模型基于 bert-base-chinese 版本 BERT 模型，在 NLI、PAWS-X、PKU-Paraphrase-Bank、STS 等语义相似数据集上进行训练，适用于通用语义匹配场景（此模型在 Chinese-STS 任务上效果较好，但在其它任务上效果并非最优，存在一定过拟合风险），比如文本特征抽取、文本向量聚类、文本语义搜索等业务场景。

注：此模型的轻量化版本，也已经开源啦！

Usage

1. Sentence-Transformers

通过 sentence-transformers 框架来使用该模型，首先进行安装：

pip install -U sentence-transformers

然后使用下面的代码来载入该模型并进行文本表征向量的提取：

from sentence_transformers import SentenceTransformer
sentences = ["我的儿子！他猛然间喊道，我的儿子在哪儿？", "我的儿子呢！他突然喊道，我的儿子在哪里？"]

model = SentenceTransformer('DMetaSoul/sbert-chinese-general-v1')
embeddings = model.encode(sentences)
print(embeddings)

2. HuggingFace Transformers

如果不想使用 sentence-transformers 的话，也可以通过 HuggingFace Transformers 来载入该模型并进行文本向量抽取：

from transformers import AutoTokenizer, AutoModel
import torch


#Mean Pooling - Take attention mask into account for correct averaging
def mean_pooling(model_output, attention_mask):
    token_embeddings = model_output[0] #First element of model_output contains all token embeddings
    input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
    return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)


# Sentences we want sentence embeddings for
sentences = ["我的儿子！他猛然间喊道，我的儿子在哪儿？", "我的儿子呢！他突然喊道，我的儿子在哪里？"]

# Load model from HuggingFace Hub
tokenizer = AutoTokenizer.from_pretrained('DMetaSoul/sbert-chinese-general-v1')
model = AutoModel.from_pretrained('DMetaSoul/sbert-chinese-general-v1')

# Tokenize sentences
encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')

# Compute token embeddings
with torch.no_grad():
    model_output = model(**encoded_input)

# Perform pooling. In this case, mean pooling.
sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])

print("Sentence embeddings:")
print(sentence_embeddings)

Evaluation

该模型在公开的几个语义匹配数据集上进行了评测，计算了向量相似度跟真实标签之间的相关性系数：

	csts_dev	csts_test	afqmc	lcqmc	bqcorpus	pawsx	xiaobu
spearman	84.54%	82.17%	23.80%	65.94%	45.52%	11.52%	48.51%

Citing & Authors

E-mail: [email protected]