language: es thumbnail: https://imgur.com/a/G77ZqQN pipeline_tag: sentence-similarity datasets:

• stsb_multi_mt tags:
• sentence-transformers
• feature-extraction
• sentence-similarity
• transformers

基于stsb_multi_mt西班牙语数据集微调的Distiluse-m-v2模型（语义文本相似度任务）

这是一个sentence-transformers模型（distiluse-base-multilingual-cased-v2版本）：能够将句子和段落映射到768维稠密向量空间，适用于聚类或语义搜索等任务。

使用方式（Sentence-Transformers库）

安装sentence-transformers后即可便捷使用：

pip install -U sentence-transformers

调用示例：

from sentence_transformers import SentenceTransformer
sentences = ["Nerea用比特币购买画作", "可以用比特币购买艺术品"]

model = SentenceTransformer('mrm8488/distiluse-base-multilingual-cased-v2-finetuned-stsb_multi_mt-es')
embeddings = model.encode(sentences)
print(embeddings)

使用方式（HuggingFace Transformers库）

若不使用sentence-transformers，可通过以下方式调用：首先将输入传入transformer模型，然后对上下文相关的词嵌入执行正确的池化操作。

from transformers import AutoTokenizer, AutoModel
import torch

#均值池化 - 考虑注意力掩码进行精确平均
def mean_pooling(model_output, attention_mask):
    token_embeddings = model_output[0] #模型输出的首个元素包含所有词嵌入
    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)

#待编码句子
sentences = ["Nerea用比特币购买画作", "可以用比特币购买艺术品"]

#从HuggingFace Hub加载模型
tokenizer = AutoTokenizer.from_pretrained('mrm8488/distiluse-base-multilingual-cased-v2-finetuned-stsb_multi_mt-es')
model = AutoModel.from_pretrained('mrm8488/distiluse-base-multilingual-cased-v2-finetuned-stsb_multi_mt-es')

#句子分词
encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')

#计算词嵌入
with torch.no_grad():
    model_output = model(**encoded_input)

#执行均值池化
sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])

print("句子嵌入向量：")
print(sentence_embeddings)

评估方法

from datasets import load_dataset
from sentence_transformers import SentenceTransformer, InputExample
from sentence_transformers.evaluation import EmbeddingSimilarityEvaluator

#加载测试集
test_data = load_dataset('stsb_multi_mt', 'es', split='test')
test_data = test_data.rename_columns({'similarity_score': 'label'})
test_data = test_data.map(lambda x: {'label': x['label'] / 5.0})

#构建测试样本
samples = []
for sample in test_data:
    samples.append(InputExample(
        texts=[sample['sentence1'], sample['sentence2']],
        label=sample['label']
    ))
    
#初始化评估器
evaluator = EmbeddingSimilarityEvaluator.from_input_examples(
    samples, write_csv=False
)

#加载模型
model = SentenceTransformer('mrm8488/distiluse-base-multilingual-cased-v2-finetuned-stsb_multi_mt-es')

#执行评估
evaluator(model)

#输出结果：0.7604056195656299

评估结果

斯皮尔曼等级相关系数：0.7604056195656299

自动化评估请参见句子嵌入基准测试：https://seb.sbert.net

训练参数

数据加载器：

使用sentence_transformers.datasets.NoDuplicatesDataLoader.NoDuplicatesDataLoader，参数为：

{'batch_size': 16}

损失函数：

采用sentence_transformers.losses.MultipleNegativesRankingLoss.MultipleNegativesRankingLoss，参数为：

{'scale': 20.0, 'similarity_fct': 'cos_sim'}

训练方法参数：

{
    "epochs": 3,
    "evaluation_steps": 0,
    "evaluator": "NoneType",
    "max_grad_norm": 1,
    "optimizer_class": "<class 'transformers.optimization.AdamW'>",
    "optimizer_params": {
        "lr": 2e-05
    },
    "scheduler": "WarmupLinear",
    "steps_per_epoch": null,
    "warmup_steps": 271,
    "weight_decay": 0.01
}

完整模型架构

SentenceTransformer(
  (0): Transformer({'max_seq_length': 512, 'do_lower_case': False}) 内含Transformer模型: DistilBertModel 
  (1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': True, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False})
)

引用与作者