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: xiaowenbin@dmetasoul.com
- Downloads last month
- 1,023
Spaces using DMetaSoul/sbert-chinese-general-v1 3
Evaluation results
- cos_sim_pearson on MTEB AFQMCvalidation set self-reported22.294
- cos_sim_spearman on MTEB AFQMCvalidation set self-reported22.567
- euclidean_pearson on MTEB AFQMCvalidation set self-reported22.526
- euclidean_spearman on MTEB AFQMCvalidation set self-reported22.567
- manhattan_pearson on MTEB AFQMCvalidation set self-reported22.502
- manhattan_spearman on MTEB AFQMCvalidation set self-reported22.537
- cos_sim_pearson on MTEB ATECtest set self-reported30.336
- cos_sim_spearman on MTEB ATECtest set self-reported30.299
- euclidean_pearson on MTEB ATECtest set self-reported32.331
- euclidean_spearman on MTEB ATECtest set self-reported30.299
- manhattan_pearson on MTEB ATECtest set self-reported32.312
- manhattan_spearman on MTEB ATECtest set self-reported30.268
- accuracy on MTEB AmazonReviewsClassification (zh)test set self-reported37.508
- f1 on MTEB AmazonReviewsClassification (zh)test set self-reported36.437
- cos_sim_pearson on MTEB BQtest set self-reported41.493
- cos_sim_spearman on MTEB BQtest set self-reported40.984
- euclidean_pearson on MTEB BQtest set self-reported41.123
- euclidean_spearman on MTEB BQtest set self-reported40.984
- manhattan_pearson on MTEB BQtest set self-reported41.026
- manhattan_spearman on MTEB BQtest set self-reported40.875
- accuracy on MTEB BUCC (zh-en)test set self-reported9.795
- f1 on MTEB BUCC (zh-en)test set self-reported9.351
- precision on MTEB BUCC (zh-en)test set self-reported9.179
- recall on MTEB BUCC (zh-en)test set self-reported9.795
- v_measure on MTEB CLSClusteringP2Ptest set self-reported34.985
- v_measure on MTEB CLSClusteringS2Stest set self-reported27.819
- map on MTEB CMedQAv1test set self-reported53.066
- mrr on MTEB CMedQAv1test set self-reported59.588
- map on MTEB CMedQAv2test set self-reported52.836
- mrr on MTEB CMedQAv2test set self-reported59.315
- map_at_1 on MTEB CmedqaRetrievalself-reported5.721
- map_at_10 on MTEB CmedqaRetrievalself-reported8.645
- map_at_100 on MTEB CmedqaRetrievalself-reported9.434
- map_at_1000 on MTEB CmedqaRetrievalself-reported9.586
- map_at_3 on MTEB CmedqaRetrievalself-reported7.413
- map_at_5 on MTEB CmedqaRetrievalself-reported8.050
- mrr_at_1 on MTEB CmedqaRetrievalself-reported9.627
- mrr_at_10 on MTEB CmedqaRetrievalself-reported13.094
- mrr_at_100 on MTEB CmedqaRetrievalself-reported13.854
- mrr_at_1000 on MTEB CmedqaRetrievalself-reported13.958
- mrr_at_3 on MTEB CmedqaRetrievalself-reported11.724
- mrr_at_5 on MTEB CmedqaRetrievalself-reported12.409
- ndcg_at_1 on MTEB CmedqaRetrievalself-reported9.627
- ndcg_at_10 on MTEB CmedqaRetrievalself-reported11.350
- ndcg_at_100 on MTEB CmedqaRetrievalself-reported15.593
- ndcg_at_1000 on MTEB CmedqaRetrievalself-reported19.619
- ndcg_at_3 on MTEB CmedqaRetrievalself-reported9.317
- ndcg_at_5 on MTEB CmedqaRetrievalself-reported10.049
- precision_at_1 on MTEB CmedqaRetrievalself-reported9.627
- precision_at_10 on MTEB CmedqaRetrievalself-reported2.796
- precision_at_100 on MTEB CmedqaRetrievalself-reported0.629
- precision_at_1000 on MTEB CmedqaRetrievalself-reported0.118
- precision_at_3 on MTEB CmedqaRetrievalself-reported5.476
- precision_at_5 on MTEB CmedqaRetrievalself-reported4.121
- recall_at_1 on MTEB CmedqaRetrievalself-reported5.721
- recall_at_10 on MTEB CmedqaRetrievalself-reported15.190
- recall_at_100 on MTEB CmedqaRetrievalself-reported33.633
- recall_at_1000 on MTEB CmedqaRetrievalself-reported62.020
- recall_at_3 on MTEB CmedqaRetrievalself-reported9.099
- recall_at_5 on MTEB CmedqaRetrievalself-reported11.423