语言:

• 德语
• 法语
• 英语
• 罗马尼亚语
• 中文 缩略图: 标签:
• 句子对齐 许可证: bsd-3-clause

AWESOME：多语言编码器的词嵌入空间对齐

该模型源自以下GitHub仓库：https://github.com/neulab/awesome-align

对应论文：https://arxiv.org/abs/2101.08231

若使用此模型，请引用原论文：

@inproceedings{dou2021word,
  title={Word Alignment by Fine-tuning Embeddings on Parallel Corpora},
  author={Dou, Zi-Yi and Neubig, Graham},
  booktitle={Conference of the European Chapter of the Association for Computational Linguistics (EACL)},
  year={2021}
}

awesome-align 是一款工具，可从多语言BERT（mBERT）中提取词对齐 演示，并允许您在平行语料库上微调mBERT以获得更好的对齐质量（详见论文）。

使用方法（复制自演示）

from transformers import AutoModel, AutoTokenizer
import itertools
import torch

# 加载模型
model = AutoModel.from_pretrained("aneuraz/awesome-align-with-co")
tokenizer = AutoTokenizer.from_pretrained("aneuraz/awesome-align-with-co")

# 模型参数
align_layer = 8
threshold = 1e-3

# 定义输入
src = 'awesome-align is awesome !'
tgt = '牛对齐 是 牛 ！'

# 预处理
sent_src, sent_tgt = src.strip().split(), tgt.strip().split()
token_src, token_tgt = [tokenizer.tokenize(word) for word in sent_src], [tokenizer.tokenize(word) for word in sent_tgt]
wid_src, wid_tgt = [tokenizer.convert_tokens_to_ids(x) for x in token_src], [tokenizer.convert_tokens_to_ids(x) for x in token_tgt]
ids_src, ids_tgt = tokenizer.prepare_for_model(list(itertools.chain(*wid_src)), return_tensors='pt', model_max_length=tokenizer.model_max_length, truncation=True)['input_ids'], tokenizer.prepare_for_model(list(itertools.chain(*wid_tgt)), return_tensors='pt', truncation=True, model_max_length=tokenizer.model_max_length)['input_ids']
sub2word_map_src = []
for i, word_list in enumerate(token_src):
  sub2word_map_src += [i for x in word_list]
sub2word_map_tgt = []
for i, word_list in enumerate(token_tgt):
  sub2word_map_tgt += [i for x in word_list]
  
# 对齐
align_layer = 8
threshold = 1e-3
model.eval()
with torch.no_grad():
  out_src = model(ids_src.unsqueeze(0), output_hidden_states=True)[2][align_layer][0, 1:-1]
  out_tgt = model(ids_tgt.unsqueeze(0), output_hidden_states=True)[2][align_layer][0, 1:-1]

  dot_prod = torch.matmul(out_src, out_tgt.transpose(-1, -2))

  softmax_srctgt = torch.nn.Softmax(dim=-1)(dot_prod)
  softmax_tgtsrc = torch.nn.Softmax(dim=-2)(dot_prod)

  softmax_inter = (softmax_srctgt > threshold)*(softmax_tgtsrc > threshold)

align_subwords = torch.nonzero(softmax_inter, as_tuple=False)
align_words = set()
for i, j in align_subwords:
  align_words.add( (sub2word_map_src[i], sub2word_map_tgt[j]) )
  
print(align_words)