imvladikon/sentence-transformers-alephbert[WIP]

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

当前版本是在私有语料库上进行的 LaBSE 模型的蒸馏。

使用方式（Sentence-Transformers）

当您安装了 sentence-transformers 后，使用此模型变得很容易：

pip install -U sentence-transformers

然后您可以像这样使用该模型：

from sentence_transformers import SentenceTransformer
from sentence_transformers.util import cos_sim

sentences = [
"הם היו שמחים לראות את האירוע שהתקיים.",
"לראות את האירוע שהתקיים היה מאוד משמח להם."
]

model = SentenceTransformer('imvladikon/sentence-transformers-alephbert')
embeddings = model.encode(sentences)


print(cos_sim(*tuple(embeddings)).item())
# 0.883316159248352

使用方式（HuggingFace Transformers）

如果没有 sentence-transformers ，您可以像这样使用该模型：首先，通过变换器模型进行输入，然后必须在上下文化的词嵌入之上应用正确的汇集操作。

import torch
from torch import nn
from transformers import AutoTokenizer, AutoModel


#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 = [
"הם היו שמחים לראות את האירוע שהתקיים.",
"לראות את האירוע שהתקיים היה מאוד משמח להם."
]

# Load model from HuggingFace Hub
tokenizer = AutoTokenizer.from_pretrained('imvladikon/sentence-transformers-alephbert')
model = AutoModel.from_pretrained('imvladikon/sentence-transformers-alephbert')

# 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'])

cos_sim = nn.CosineSimilarity(dim=0, eps=1e-6)
print(cos_sim(sentence_embeddings[0], sentence_embeddings[1]).item())

评估结果

有关此模型的自动评估，请参见Sentence Embeddings Benchmark： https://seb.sbert.net

训练

该模型训练时使用了以下参数：

DataLoader：

torch.utils.data.dataloader.DataLoader长度为44999，参数为：

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

Loss：

sentence_transformers.losses.MultipleNegativesRankingLoss.MultipleNegativesRankingLoss参数为：

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

fit()方法的参数：

{
    "epochs": 10,
    "evaluation_steps": 0,
    "evaluator": "NoneType",
    "max_grad_norm": 1,
    "optimizer_class": "<class 'torch.optim.adamw.AdamW'>",
    "optimizer_params": {
        "lr": 2e-05
    },
    "scheduler": "WarmupLinear",
    "steps_per_epoch": null,
    "warmup_steps": 44999,
    "weight_decay": 0.01
}

完整的模型架构

SentenceTransformer(
  (0): Transformer({'max_seq_length': 512, '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

@misc{seker2021alephberta,
      title={AlephBERT:A Hebrew Large Pre-Trained Language Model to Start-off your Hebrew NLP Application With}, 
      author={Amit Seker and Elron Bandel and Dan Bareket and Idan Brusilovsky and Refael Shaked Greenfeld and Reut Tsarfaty},
      year={2021},
      eprint={2104.04052},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

@misc{reimers2019sentencebert,
      title={Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks}, 
      author={Nils Reimers and Iryna Gurevych},
      year={2019},
      eprint={1908.10084},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}