英文

ko-sbert-multitask

这是一个 sentence-transformers 模型:它将句子和段落映射到一个768维的稠密向量空间,并可以用于聚类或语义搜索等任务。

使用方法(Sentence-Transformers)

安装了 sentence-transformers 后,使用该模型变得很简单:

pip install -U sentence-transformers

然后可以像这样使用模型:

from sentence_transformers import SentenceTransformer
sentences = ["안녕하세요?", "한국어 문장 임베딩을 위한 버트 모델입니다."]

model = SentenceTransformer('jhgan/ko-sbert-multitask')
embeddings = model.encode(sentences)
print(embeddings)

使用方法(HuggingFace Transformers)

如果没有 sentence-transformers ,可以这样使用该模型:首先,将输入通过transformer模型,然后必须在上下文化的单词嵌入之上应用正确的汇聚操作。

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 = ['This is an example sentence', 'Each sentence is converted']

# Load model from HuggingFace Hub
tokenizer = AutoTokenizer.from_pretrained('jhgan/ko-sbert-multitask')
model = AutoModel.from_pretrained('jhgan/ko-sbert-multitask')

# 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)

评估结果

使用KorSTS、KorNLI训练数据集进行多任务训练,然后使用KorSTS评估数据集进行评估的结果如下:

  • 余弦皮尔逊相关系数:84.13
  • 余弦斯皮尔曼相关系数:84.71
  • 欧氏距离皮尔逊相关系数:82.42
  • 欧氏距离斯皮尔曼相关系数:82.66
  • 曼哈顿距离皮尔逊相关系数:81.41
  • 曼哈顿距离斯皮尔曼相关系数:81.69
  • 点积皮尔逊相关系数:80.05
  • 点积斯皮尔曼相关系数:79.69

训练

使用以下参数训练模型:

数据加载器 :

sentence_transformers.datasets.NoDuplicatesDataLoader.NoDuplicatesDataLoader,长度为8885,具有以下参数:

{'batch_size': 64}

损失函数 :

sentence_transformers.losses.MultipleNegativesRankingLoss.MultipleNegativesRankingLoss,具有以下参数:

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

数据加载器 :

torch.utils.data.dataloader.DataLoader,长度为719,具有以下参数:

{'batch_size': 8, 'sampler': 'torch.utils.data.sampler.RandomSampler', 'batch_sampler': 'torch.utils.data.sampler.BatchSampler'}

损失函数 :

sentence_transformers.losses.CosineSimilarityLoss.CosineSimilarityLoss

fit()方法的参数:

{
    "epochs": 5,
    "evaluation_steps": 1000,
    "evaluator": "sentence_transformers.evaluation.EmbeddingSimilarityEvaluator.EmbeddingSimilarityEvaluator",
    "max_grad_norm": 1,
    "optimizer_class": "<class 'transformers.optimization.AdamW'>",
    "optimizer_params": {
        "lr": 2e-05
    },
    "scheduler": "WarmupLinear",
    "steps_per_epoch": null,
    "warmup_steps": 360,
    "weight_decay": 0.01
}

完整的模型架构

SentenceTransformer(
  (0): Transformer({'max_seq_length': 128, 'do_lower_case': False}) with Transformer model: BertModel 
  (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})
)

引用和作者

  • Ham, J., Choe, Y. J., Park, K., Choi, I., & Soh, H. (2020). Kornli和korsts:用于韩文自然语言理解的新基准数据集。arXiv预印本arXiv:2004.03289
  • Reimers, Nils和Iryna Gurevych。“Sentence-BERT: 使用Siamese BERT-Networks生成句子嵌入。”ArXiv abs/1908.10084(2019)
  • Reimers, Nils和Iryna Gurevych。“使用知识蒸馏使单语句子嵌入多语言化。”EMNLP(2020)。