- \n
- \u4f20\u7edf\u201c\u5148\u8bc6\u522bgloss\u518d\u7ffb\u8bd1\u201d\u7684\u65b9\u6cd5\u867d\u7136\u6709\u6548，\u4f46Gloss \u6807\u6ce8\u975e\u5e38\u6602\u8d35\u3001\u4e14Gloss \u4f1a\u5f62\u6210\u4fe1\u606f\u74f6\u9888，\u5373\u624b\u8bed\u4e2d\u7684\u8868\u60c5\u3001\u52a8\u4f5c\u5e45\u5ea6\u3001\u7a7a\u95f4\u4f4d\u7f6e\u4ee5\u53ca\u8fde\u7eed\u52a8\u4f5c\u5173\u7cfb，\u53ef\u80fd\u65e0\u6cd5\u88ab\u7b80\u5355\u7684 gloss \u5e8f\u5217\u5b8c\u6574\u8868\u8fbe\u3002\u4e00\u65e6\u4fe1\u606f\u5728 gloss \u9636\u6bb5\u4e22\u5931，\u540e\u7eed\u7ffb\u8bd1\u6a21\u578b\u65e0\u6cd5\u6062\u590d\u3002<\/li>\n
- \u73b0\u6709\u7684gloss-free\u65b9\u6cd5\u53ef\u80fd\u5b66\u5230\u4e86\u52a8\u4f5c\u6216\u5916\u89c2\u4fe1\u606f，\u5374\u672a\u5fc5\u77e5\u9053\u54ea\u4e9b\u89c6\u89c9\u53d8\u5316\u5bf9\u5e94\u53e5\u5b50\u7684\u6838\u5fc3\u8bed\u4e49，\u5373\u4e0d\u591f\u201c\u8bed\u8a00\u5316\u201d，\u7f3a\u5c11\u4e86\u660e\u786e\u7684\u4e2d\u95f4\u76d1\u7763\u4fe1\u53f7\u3002<\/li>\n<\/ul>\n
\u8bba\u6587\u7684\u6838\u5fc3\u60f3\u6cd5<\/h3>\n
\u8bad\u7ec3\u89c6\u89c9\u7f16\u7801\u5668\u65f6，\u4e0d\u53ea\u662f\u8981\u6c42\u5b83\u8bc6\u522b\u89c6\u9891\u5185\u5bb9，\u800c\u662f\u8981\u6c42\u5b83\u63d0\u53d6\u51fa\u7684\u6574\u4f53\u89c6\u9891\u8868\u793a，\u5728\u8bed\u4e49\u7a7a\u95f4\u4e2d\u63a5\u8fd1\u5bf9\u5e94\u7684\u81ea\u7136\u8bed\u8a00\u53e5\u5b50，\u8be5\u8fc7\u7a0b\u5206\u4e3a\u4e24\u9636\u6bb5：<\/p>\n
- \u89c6\u89c9\u2014\u8bed\u8a00\u9884\u8bad\u7ec3 VLP（\u5c06\u201c\u89c6\u89c9\u7f16\u7801\u5668\u7406\u89e3\u624b\u8bed\u89c6\u9891\u201d\u548c\u201c\u6587\u672c\u89e3\u7801\u5668\u751f\u6210\u81ea\u7136\u8bed\u8a00\u53e5\u5b50\u201d\u4e24\u79cd\u9884\u8bad\u7ec3\u4efb\u52a1\u7ed3\u5408\u8d77\u6765），\u4efb\u52a1\u4e00：\u89c6\u9891\u2014\u6587\u672c\u5bf9\u6bd4\u5b66\u4e60（\u8bed\u8a00\u53e5\u5b50\u76f8\u5f53\u4e8e\u4e3a\u89c6\u9891\u7f16\u7801\u5668\u63d0\u4f9b\u4e86\u9ad8\u5c42\u8bed\u4e49\u76d1\u7763）\u3001\u4efb\u52a1\u4e8c：\u63a9\u7801\u53e5\u5b50\u6062\u590d（\u7c7b\u4f3c\u53bb\u566a\u8bed\u8a00\u6a21\u578b，\u5176\u4f5c\u7528\u662f\u8ba9 Text Decoder \u63d0\u524d\u5b66\u4e60\u8bed\u6cd5\u548c\u8bed\u4e49\u6d41\u7545），\u5728\u8fd9\u4e2a\u9636\u6bb5\u7684\u4e24\u4e2a\u4efb\u52a1\u4e2d，\u90fd\u7528\u5230\u4e86（\u540c\u4e00\u4e2a）mBART\u00a0Encoder，\u5e76\u4e14\u5b83\u4e0d\u88ab\u51bb\u7ed3\u65f6\u6548\u679c\u6709\u660e\u663e\u63d0\u5347\u3002<\/li>\n
- \u65e0 Gloss \u7aef\u5230\u7aef\u7ffb\u8bd1，Visual Encoder（ResNet18-1DCNN-Transformer Encoder）\u6700\u7ec8\u5f97\u5230\u89c6\u9891\u9690\u85cf\u8868\u793a，Text Decoder （\u5e26\u56e0\u679c\u63a9\u7801\u7684 Transformer Decoder）\u6839\u636e\u5df2\u7ecf\u751f\u6210\u7684\u8bcd\u548c\u89c6\u9891\u8868\u793a\u9884\u6d4b\u4e0b\u4e00\u4e2a\u8bcd\u3002\u5728\u8fd9\u4e00\u9636\u6bb5text encoder\u4e0d\u518d\u4fdd\u7559\u3002<\/li>\n
\u5e76\u4e14\u4f5c\u8005\u901a\u8fc7\u5b9e\u9a8c\u53d1\u73b0\u201c\u65e0 gloss \u624b\u8bed\u7ffb\u8bd1\u7684\u4e3b\u8981\u74f6\u9888\u662f\u89c6\u89c9\u8868\u793a，\u800c\u4e0d\u662f\u5355\u7eaf\u7684\u8bed\u8a00\u751f\u6210\u80fd\u529b\u201d\u4ee5\u53ca\u201c\u7b2c\u4e8c\u9636\u6bb5\u7684\u8bad\u7ec3epoch\u6570\u63d0\u5347\u548c\u6570\u636e\u89c4\u6a21\u7684\u63d0\u5347\u4e5f\u80fd\u5e26\u6765\u7a33\u5b9a\u6536\u76ca\u201d\u3002<\/p>\n
VLP \u7684\u4f5c\u7528\u4e0d\u4ec5\u662f\u8ba9\u53e5\u5b50\u66f4\u6d41\u7545，\u66f4\u91cd\u8981\u7684\u662f：\u63d0\u9ad8\u89c6\u9891\u8bed\u4e49\u4e0e\u5177\u4f53\u8bed\u8a00\u8bcd\u6c47\u4e4b\u95f4\u7684\u5bf9\u5e94\u51c6\u786e\u6027\u3002<\/p>\n
\u9884\u8bad\u7ec3\u548c\u5bf9\u6bd4\u5b66\u4e60\u76f8\u5173\u6982\u5ff5\u8865\u5145<\/h3>\n
\u9884\u8bad\u7ec3\u7684\u4f5c\u7528：<\/p>\n
- \n
- \u89c6\u89c9\u7f16\u7801\u5668：\u901a\u8fc7\u89c6\u9891\u2014\u6587\u672c\u5bf9\u6bd4\u5b66\u4e60，\u5b66\u4f1a\u628a\u624b\u8bed\u89c6\u9891\u7f16\u7801\u6210\u5e26\u6709\u8bed\u8a00\u8bed\u4e49\u7684\u89c6\u89c9\u8868\u793a\u3002<\/li>\n
- \u6587\u672c\u7f16\u7801\u5668：\u4f5c\u4e3a\u9884\u8bad\u7ec3\u9636\u6bb5\u7684\u8f85\u52a9\u6865\u6881，\u7528\u6765\u7f16\u7801\u6587\u672c\u3001\u6307\u5bfc\u89c6\u89c9\u5bf9\u9f50\u4ee5\u53ca\u652f\u6301\u63a9\u7801\u6062\u590d；\u5b83\u4e0d\u4f1a\u8fdb\u5165\u7b2c\u4e8c\u9636\u6bb5（\u540e\u7eed\u5fae\u8c03），\u5b83\u7684\u4f5c\u7528\u5c31\u662f\u7f16\u7801\u6587\u672c，\u8ba9\u89c6\u89c9\u7f16\u7801\u5668\u77e5\u9053\u54ea\u4e9b\u89c6\u9891\u4fe1\u606f\u4e0e\u53e5\u5b50\u7684\u8bed\u8a00\u8bed\u4e49\u76f8\u5173（\u540c\u65f6\u5bf9\u4e8e\u672c\u6587，\u5b83\u8fd8\u4e0e\u8bad\u7ec3\u4e86\u6587\u672c\u89e3\u7801\u5668\u7684\u53e5\u6cd5\u6062\u590d\u80fd\u529b）\u3002<\/li>\n
- \u6587\u672c\u89e3\u7801\u5668：\u901a\u8fc7\u63a9\u7801\u53e5\u5b50\u6062\u590d，\u63d0\u524d\u5b66\u4e60\u53e5\u6cd5\u3001\u8bed\u4e49\u548c\u6587\u672c\u751f\u6210\u80fd\u529b\u3002<\/li>\n<\/ul>\n
\u5fae\u8c03\u9636\u6bb5：\u628a\u9884\u8bad\u7ec3\u7684 Visual Encoder \u548c Text Decoder \u63a5\u8d77\u6765，\u4e24\u8005\u4e4b\u95f4\u7684\u8de8\u6ce8\u610f\u529b\u548c\u8f93\u51fa\u5c42\u4e5f\u4e00\u8d77\u9002\u914d\u3002\u56e0\u4e3a\u7b2c\u4e00\u9636\u6bb5\u7684\u6587\u672c\u7f16\u7801\u6765\u81ea\u6587\u672c\u7f16\u7801\u5668，\u4f46\u7b2c\u4e8c\u9636\u6bb5\u4e0d\u7528\u6587\u672c\u7f16\u7801\u5668，\u6240\u4ee5\u9700\u8981\u6709\u7b2c\u4e8c\u9636\u6bb5\u7684\u8bad\u7ec3，\u5373\u5fae\u8c03\u3002<\/p>\n
\u5927\u89c4\u6a21\u5916\u90e8\u76f8\u5173\u6570\u636e\u9884\u8bad\u7ec3\u2192\u76ee\u6807\u8bad\u7ec3\u96c6\u4e0a\u7684\u7ee7\u7eed\u9884\u8bad\u7ec3\u2192\u76ee\u6807\u4efb\u52a1\u5fae\u8c03，\u5373\u5982\u672c\u6587\u7684，ImageNet \u4e0a\u9884\u8bad\u7ec3\u7684 ResNet18 \u548c CC25 \u6587\u672c\u8bed\u6599\u4e0a\u9884\u8bad\u7ec3\u7684 mBART \u2192 VLP \u9636\u6bb5\u5728\u5404\u81ea\u7684\u8bad\u7ec3\u96c6\u8fdb\u884c\u4efb\u52a1\u76f8\u5173\u9884\u8bad\u7ec3\u2192\u7b2c\u4e8c\u9636\u6bb5\u4ecd\u5728\u540c\u4e00\u8bad\u7ec3\u96c6\u4e0a\u8fdb\u884c\u7aef\u5230\u7aef SLT \u5fae\u8c03\u3002\u56e0\u4e3a\u4e24\u4e2a\u9636\u6bb5\u5b66\u4e60\u7684\u4efb\u52a1\u4e0d\u540c，\u6240\u4ee5\u4f7f\u7528\u76f8\u540c\u7684（\u89c6\u9891，\u6587\u672c）\u5bf9\u6ca1\u6709\u5173\u7cfb，\u53ea\u8981\u4e0d\u5728\u9884\u8bad\u7ec3\u9636\u6bb5\u4f7f\u7528\u6d4b\u8bd5\u96c6\u7684\u89c6\u9891\u6216\u5bf9\u5e94\u6587\u672c\u5373\u53ef\u3002<\/p>\n
\u7b80\u5355\u5b9e\u73b0\u53cc\u5411\u5bf9\u6bd4\u5b66\u4e60：<\/p>\n
import torch
\nimport torch.nn as nn
\nimport torch.nn.functional as F<\/p>\nclass SimpleVLP(nn.Module):
\n \u00a0 \u00a0def __init__(
\n \u00a0 \u00a0 \u00a0 \u00a0self,
\n \u00a0 \u00a0 \u00a0 \u00a0video_input_dim: int,
\n \u00a0 \u00a0 \u00a0 \u00a0text_input_dim: int,
\n \u00a0 \u00a0 \u00a0 \u00a0projection_dim: int = 256,
\n \u00a0 \u00a0):
\n \u00a0 \u00a0 \u00a0 \u00a0super().__init__()<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0# \u7b80\u5316\u7684\u89c6\u89c9\u7f16\u7801\u5668 + \u6295\u5f71\u5934
\n \u00a0 \u00a0 \u00a0 \u00a0self.video_encoder = nn.Sequential(
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0nn.Linear(video_input_dim, 512),
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0nn.ReLU(),
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0nn.Linear(512, projection_dim),
\n \u00a0 \u00a0 \u00a0 \u00a0)<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0# \u7b80\u5316\u7684\u6587\u672c\u7f16\u7801\u5668 + \u6295\u5f71\u5934
\n \u00a0 \u00a0 \u00a0 \u00a0# \u8bba\u6587\u4e2d\u8fd9\u91cc\u5bf9\u5e94 mBART Text Encoder
\n \u00a0 \u00a0 \u00a0 \u00a0self.text_encoder = nn.Sequential(
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0nn.Linear(text_input_dim, 512),
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0nn.ReLU(),
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0nn.Linear(512, projection_dim),
\n \u00a0 \u00a0 \u00a0 \u00a0)<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0# \u6e29\u5ea6\u7cfb\u6570，\u63a7\u5236\u76f8\u4f3c\u5ea6\u5206\u5e03\u7684\u5c16\u9510\u7a0b\u5ea6
\n \u00a0 \u00a0 \u00a0 \u00a0self.logit_scale = nn.Parameter(
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0torch.tensor(1.0 \/ 0.07).log()
\n \u00a0 \u00a0 \u00a0 \u00a0)<\/p>\n\u00a0 \u00a0def forward(
\n \u00a0 \u00a0 \u00a0 \u00a0self,
\n \u00a0 \u00a0 \u00a0 \u00a0video_features: torch.Tensor,
\n \u00a0 \u00a0 \u00a0 \u00a0text_features: torch.Tensor,
\n \u00a0 \u00a0):
\n \u00a0 \u00a0 \u00a0 \u00a0# video_embedding: [batch_size, projection_dim]
\n \u00a0 \u00a0 \u00a0 \u00a0video_embedding = self.video_encoder(video_features)<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0# text_embedding: [batch_size, projection_dim]
\n \u00a0 \u00a0 \u00a0 \u00a0text_embedding = self.text_encoder(text_features)<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0# L2 \u5f52\u4e00\u5316，\u4f7f\u70b9\u79ef\u7b49\u4ef7\u4e8e\u4f59\u5f26\u76f8\u4f3c\u5ea6
\n \u00a0 \u00a0 \u00a0 \u00a0video_embedding = F.normalize(video_embedding, dim=-1)
\n \u00a0 \u00a0 \u00a0 \u00a0text_embedding = F.normalize(text_embedding, dim=-1)<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0# \u8ba1\u7b97\u4e00\u4e2a\u6279\u6b21\u5185\u6240\u6709\u89c6\u9891\u548c\u6587\u672c\u4e4b\u95f4\u7684\u76f8\u4f3c\u5ea6
\n \u00a0 \u00a0 \u00a0 \u00a0# logits[i, j] \u8868\u793a\u7b2c i \u4e2a\u89c6\u9891\u4e0e\u7b2c j \u4e2a\u53e5\u5b50\u7684\u76f8\u4f3c\u5ea6
\n \u00a0 \u00a0 \u00a0 \u00a0logits = self.logit_scale.exp() * (
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0video_embedding @ text_embedding.T
\n \u00a0 \u00a0 \u00a0 \u00a0) # \u6784\u9020\u5bf9\u89d2\u7ebflabel\u6b63\u6837\u672c\u6700\u6838\u5fc3\u7684\u4e00\u6b65<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0batch_size = logits.size(0)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0# \u6b63\u786e\u914d\u5bf9\u4f4d\u4e8e\u76f8\u4f3c\u5ea6\u77e9\u9635\u7684\u5bf9\u89d2\u7ebf\u4e0a：
\n \u00a0 \u00a0 \u00a0 \u00a0# \u7b2c i \u4e2a\u89c6\u9891\u5e94\u8be5\u5bf9\u5e94\u7b2c i \u4e2a\u6587\u672c
\n \u00a0 \u00a0 \u00a0 \u00a0labels = torch.arange(batch_size, device=logits.device)<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0# \u89c6\u9891\u627e\u6587\u672c（\u5355\u5411InfoNCE，\u5c31\u662f\u4e00\u6b21\u57fa\u4e8e\u6b63\u8d1f\u6837\u672c\u76f8\u4f3c\u5ea6\u7684\u4ea4\u53c9\u71b5）
\n \u00a0 \u00a0 \u00a0 \u00a0loss_video_to_text = F.cross_entropy(logits, labels)<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0# \u6587\u672c\u627e\u89c6\u9891（\u5355\u5411InfoNCE）
\n \u00a0 \u00a0 \u00a0 \u00a0loss_text_to_video = F.cross_entropy(logits.T, labels)<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0# \u5bf9\u79f0\u5bf9\u6bd4\u635f\u5931
\n \u00a0 \u00a0 \u00a0 \u00a0loss = (loss_video_to_text + loss_text_to_video) \/ 2<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0return loss, logits<\/p>\n
# —————- \u4f7f\u7528\u793a\u4f8b —————-<\/p>\n
batch_size = 8<\/p>\n
# \u5047\u8bbe\u5df2\u7ecf\u83b7\u5f97\u89c6\u9891\u7279\u5f81\u548c\u6587\u672c\u7279\u5f81
\nvideo_features = torch.randn(batch_size, 1024)
\ntext_features = torch.randn(batch_size, 768)<\/p>\nmodel = SimpleVLP(
\n \u00a0 \u00a0video_input_dim=1024,
\n \u00a0 \u00a0text_input_dim=768,
\n \u00a0 \u00a0projection_dim=256,
\n)<\/p>\noptimizer = torch.optim.Adam(model.parameters(), lr=1e-4)<\/p>\n
loss, similarity_matrix = model(video_features, text_features)<\/p>\n
optimizer.zero_grad()
\nloss.backward()
\noptimizer.step()<\/p>\nprint("\u5bf9\u6bd4\u5b66\u4e60\u635f\u5931：", loss.item())
\nprint("\u76f8\u4f3c\u5ea6\u77e9\u9635\u5f62\u72b6：", similarity_matrix.shape) <\/p>\n\u7b80\u5355\u5b9e\u73b0\u9884\u8bad\u7ec3\u63a9\u7801\u6062\u590d：<\/p>\n
"""
\n\u63a9\u7801\u53e5\u5b50\u6062\u590d\u7684\u6574\u4f53\u6d41\u7a0b
\n========================================<\/p>\n\u539f\u59cb\u53e5\u5b50：
\n \u00a0 \u00a0<bos> \u4eca\u5929 \u5929\u6c14 \u5f88 \u597d <eos><\/p>\n\u6b65\u9aa4 1：\u968f\u673a\u63a9\u7801\u90e8\u5206\u5185\u5bb9
\n \u00a0 \u00a0<bos> \u4eca\u5929 <mask> \u5f88 \u597d <eos><\/p>\n\u6b65\u9aa4 2：Text Encoder \u7f16\u7801\u63a9\u7801\u540e\u7684\u53e5\u5b50
\n \u00a0 \u00a0masked_sentence_ids
\n \u00a0 \u00a0 \u00a0 \u00a0-> Token Embedding + Position Embedding
\n \u00a0 \u00a0 \u00a0 \u00a0-> Transformer Encoder
\n \u00a0 \u00a0 \u00a0 \u00a0-> \u5e26\u6709\u4e0a\u4e0b\u6587\u4fe1\u606f\u7684\u7f16\u7801\u8868\u793a memory<\/p>\n\u6b65\u9aa4 3：Text Decoder \u6839\u636e\u4e24\u4e2a\u4fe1\u606f\u6062\u590d\u539f\u53e5
\n \u00a0 \u00a0\u2460 Text Encoder \u8f93\u51fa\u7684 memory
\n \u00a0 \u00a0\u2461 \u5df2\u7ecf\u751f\u6210\u6216\u5df2\u77e5\u7684\u524d\u6587<\/p>\n\u8bad\u7ec3\u65f6\u4f7f\u7528\u6559\u5e08\u5f3a\u5236：
\n \u00a0 \u00a0Decoder \u8f93\u5165：
\n \u00a0 \u00a0 \u00a0 \u00a0<bos> \u4eca\u5929 \u5929\u6c14 \u5f88 \u597d<\/p>\n\u00a0 \u00a0\u76d1\u7763\u6807\u7b7e：
\n \u00a0 \u00a0 \u00a0 \u00a0\u4eca\u5929 \u5929\u6c14 \u5f88 \u597d <eos><\/p>\n\u6b65\u9aa4 4：\u4f7f\u7528\u4ea4\u53c9\u71b5\u635f\u5931\u8bad\u7ec3
\n \u00a0 \u00a0Decoder \u5728\u6bcf\u4e2a\u4f4d\u7f6e\u9884\u6d4b\u539f\u59cb\u53e5\u5b50\u7684\u4e0b\u4e00\u4e2a\u8bcd<\/p>\n\u6ce8\u610f：
\n \u00a0 \u00a0\u8fd9\u91cc\u6f14\u793a\u7684\u662f\u5b8c\u6574\u53e5\u5b50\u91cd\u5efa，\u800c\u4e0d\u4ec5\u4ec5\u9884\u6d4b\u88ab\u63a9\u7801\u7684\u4f4d\u7f6e\u3002
\n \u00a0 \u00a0\u8fd9\u66f4\u63a5\u8fd1\u7f16\u7801\u5668\u2014\u89e3\u7801\u5668\u5f0f\u7684\u53bb\u566a\u9884\u8bad\u7ec3\u3002
\n"""<\/p>\nimport torch
\nimport torch.nn as nn
\nimport torch.nn.functional as F<\/p>\nclass MaskedSentenceRecovery(nn.Module):
\n \u00a0 \u00a0"""
\n \u00a0 \u00a0\u7b80\u5316\u7248\u63a9\u7801\u53e5\u5b50\u6062\u590d\u6a21\u578b\u3002<\/p>\n\u00a0 \u00a0\u8f93\u5165：
\n \u00a0 \u00a0 \u00a0 \u00a01. \u88ab\u63a9\u7801\u7684\u53e5\u5b50
\n \u00a0 \u00a0 \u00a0 \u00a02. \u53f3\u79fb\u540e\u7684\u539f\u59cb\u53e5\u5b50<\/p>\n\u00a0 \u00a0\u8f93\u51fa：
\n \u00a0 \u00a0 \u00a0 \u00a0\u539f\u59cb\u53e5\u5b50\u4e2d\u6bcf\u4e2a\u4f4d\u7f6e\u7684\u8bcd\u8868\u6982\u7387
\n \u00a0 \u00a0"""<\/p>\n\u00a0 \u00a0def __init__(
\n \u00a0 \u00a0 \u00a0 \u00a0self,
\n \u00a0 \u00a0 \u00a0 \u00a0vocab_size: int,
\n \u00a0 \u00a0 \u00a0 \u00a0hidden_dim: int = 128,
\n \u00a0 \u00a0 \u00a0 \u00a0num_heads: int = 4,
\n \u00a0 \u00a0 \u00a0 \u00a0num_layers: int = 2,
\n \u00a0 \u00a0 \u00a0 \u00a0max_length: int = 32,
\n \u00a0 \u00a0):
\n \u00a0 \u00a0 \u00a0 \u00a0super().__init__()<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0# \u5c06\u8bcd\u7f16\u53f7\u8f6c\u6362\u4e3a\u8fde\u7eed\u5411\u91cf
\n \u00a0 \u00a0 \u00a0 \u00a0self.token_embedding = nn.Embedding(
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0num_embeddings=vocab_size,
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0embedding_dim=hidden_dim,
\n \u00a0 \u00a0 \u00a0 \u00a0)<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0# \u4e3a\u6bcf\u4e2a\u8bcd\u52a0\u5165\u4f4d\u7f6e\u4fe1\u606f
\n \u00a0 \u00a0 \u00a0 \u00a0self.position_embedding = nn.Embedding(
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0num_embeddings=max_length,
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0embedding_dim=hidden_dim,
\n \u00a0 \u00a0 \u00a0 \u00a0)<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0# \u5305\u542b Text Encoder \u548c Text Decoder \u7684\u5b8c\u6574 Transformer
\n \u00a0 \u00a0 \u00a0 \u00a0self.transformer = nn.Transformer(
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0d_model=hidden_dim,
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0nhead=num_heads,
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0num_encoder_layers=num_layers,
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0num_decoder_layers=num_layers,
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0dim_feedforward=hidden_dim * 4,
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0dropout=0.1,
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0batch_first=True,
\n \u00a0 \u00a0 \u00a0 \u00a0)<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0# \u5c06 Decoder \u7684\u9690\u85cf\u5411\u91cf\u6620\u5c04\u5230\u8bcd\u8868\u5927\u5c0f
\n \u00a0 \u00a0 \u00a0 \u00a0self.output_layer = nn.Linear(
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0hidden_dim,
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0vocab_size,
\n \u00a0 \u00a0 \u00a0 \u00a0)<\/p>\n\u00a0 \u00a0def embed_tokens(
\n \u00a0 \u00a0 \u00a0 \u00a0self,
\n \u00a0 \u00a0 \u00a0 \u00a0token_ids: torch.Tensor,
\n \u00a0 \u00a0) -> torch.Tensor:
\n \u00a0 \u00a0 \u00a0 \u00a0"""
\n \u00a0 \u00a0 \u00a0 \u00a0\u7ed9\u8bcd\u7f16\u53f7\u52a0\u5165 Token Embedding \u548c Position Embedding\u3002<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0\u53c2\u6570：
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0token_ids:
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0[batch_size, sequence_length]<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0\u8fd4\u56de：
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0embeddings:
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0[batch_size, sequence_length, hidden_dim]
\n \u00a0 \u00a0 \u00a0 \u00a0"""<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0batch_size, sequence_length = token_ids.shape<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0# \u4e3a\u6bcf\u4e2a\u4f4d\u7f6e\u751f\u6210\u7f16\u53f7：0, 1, 2, …
\n \u00a0 \u00a0 \u00a0 \u00a0position_ids = torch.arange(
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0sequence_length,
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0device=token_ids.device,
\n \u00a0 \u00a0 \u00a0 \u00a0)<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0# \u6269\u5c55\u5230\u6574\u4e2a batch
\n \u00a0 \u00a0 \u00a0 \u00a0position_ids = position_ids.unsqueeze(0).expand(
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0batch_size,
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0sequence_length,
\n \u00a0 \u00a0 \u00a0 \u00a0)<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0token_embeddings = self.token_embedding(token_ids)
\n \u00a0 \u00a0 \u00a0 \u00a0position_embeddings = self.position_embedding(position_ids)<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0return token_embeddings + position_embeddings<\/p>\n
\u00a0 \u00a0def forward(
\n \u00a0 \u00a0 \u00a0 \u00a0self,
\n \u00a0 \u00a0 \u00a0 \u00a0masked_sentence_ids: torch.Tensor,
\n \u00a0 \u00a0 \u00a0 \u00a0decoder_input_ids: torch.Tensor,
\n \u00a0 \u00a0) -> torch.Tensor:
\n \u00a0 \u00a0 \u00a0 \u00a0"""
\n \u00a0 \u00a0 \u00a0 \u00a0masked_sentence_ids：
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u88ab\u63a9\u7801\u540e\u7684\u53e5\u5b50\u3002
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u8fd9\u662f Text Encoder \u7684\u8f93\u5165\u3002<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0decoder_input_ids：
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u53f3\u79fb\u540e\u7684\u539f\u59cb\u53e5\u5b50\u3002
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u8fd9\u662f Text Decoder \u7684\u8f93\u5165\u3002
\n \u00a0 \u00a0 \u00a0 \u00a0"""<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0# ====================================================
\n \u00a0 \u00a0 \u00a0 \u00a0# \u7b2c\u4e00\u6b65：\u51c6\u5907 Text Encoder \u7684\u8f93\u5165
\n \u00a0 \u00a0 \u00a0 \u00a0# ====================================================<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0# \u6ce8\u610f：\u8fd9\u91cc\u7f16\u7801\u7684\u662f\u201c\u63a9\u7801\u540e\u7684\u53e5\u5b50\u201d
\n \u00a0 \u00a0 \u00a0 \u00a0encoder_embeddings = self.embed_tokens(
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0masked_sentence_ids
\n \u00a0 \u00a0 \u00a0 \u00a0)<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0# ====================================================
\n \u00a0 \u00a0 \u00a0 \u00a0# \u7b2c\u4e8c\u6b65：\u51c6\u5907 Text Decoder \u7684\u8f93\u5165
\n \u00a0 \u00a0 \u00a0 \u00a0# ====================================================<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0decoder_embeddings = self.embed_tokens(
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0decoder_input_ids
\n \u00a0 \u00a0 \u00a0 \u00a0)<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0target_length = decoder_input_ids.size(1)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0# Decoder \u4f7f\u7528\u56e0\u679c\u63a9\u7801：
\n \u00a0 \u00a0 \u00a0 \u00a0# \u5f53\u524d\u4f4d\u7f6e\u53ea\u80fd\u770b\u5230\u81ea\u5df1\u4ee5\u53ca\u4e4b\u524d\u7684\u8bcd，
\n \u00a0 \u00a0 \u00a0 \u00a0# \u4e0d\u80fd\u5077\u770b\u672a\u6765\u7684\u76ee\u6807\u8bcd\u3002
\n \u00a0 \u00a0 \u00a0 \u00a0causal_mask = nn.Transformer.generate_square_subsequent_mask(
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0target_length,
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0device=decoder_input_ids.device,
\n \u00a0 \u00a0 \u00a0 \u00a0)<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0# ====================================================
\n \u00a0 \u00a0 \u00a0 \u00a0# \u7b2c\u4e09\u6b65：\u6267\u884c\u7f16\u7801\u548c\u89e3\u7801
\n \u00a0 \u00a0 \u00a0 \u00a0# ====================================================<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0# \u5185\u90e8\u8ba1\u7b97\u8fc7\u7a0b\u53ef\u4ee5\u7406\u89e3\u4e3a：
\n \u00a0 \u00a0 \u00a0 \u00a0#
\n \u00a0 \u00a0 \u00a0 \u00a0# memory = TextEncoder(encoder_embeddings)
\n \u00a0 \u00a0 \u00a0 \u00a0# hidden = TextDecoder(
\n \u00a0 \u00a0 \u00a0 \u00a0# \u00a0 \u00a0 decoder_embeddings,
\n \u00a0 \u00a0 \u00a0 \u00a0# \u00a0 \u00a0 memory
\n \u00a0 \u00a0 \u00a0 \u00a0# )
\n \u00a0 \u00a0 \u00a0 \u00a0#
\n \u00a0 \u00a0 \u00a0 \u00a0# \u8fd9\u91cc nn.Transformer \u5c06\u4e24\u6b65\u5c01\u88c5\u5728\u4e00\u8d77\u3002
\n \u00a0 \u00a0 \u00a0 \u00a0decoder_hidden_states = self.transformer(
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0src=encoder_embeddings,
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0tgt=decoder_embeddings,
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0tgt_mask=causal_mask,
\n \u00a0 \u00a0 \u00a0 \u00a0)<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0# ====================================================
\n \u00a0 \u00a0 \u00a0 \u00a0# \u7b2c\u56db\u6b65：\u9884\u6d4b\u6bcf\u4e2a\u4f4d\u7f6e\u7684\u8bcd
\n \u00a0 \u00a0 \u00a0 \u00a0# ====================================================<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0# \u8f93\u51fa\u5f62\u72b6：
\n \u00a0 \u00a0 \u00a0 \u00a0# [batch_size, target_length, vocab_size]
\n \u00a0 \u00a0 \u00a0 \u00a0logits = self.output_layer(
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0decoder_hidden_states
\n \u00a0 \u00a0 \u00a0 \u00a0)<\/p>\n\u00a0 \u00a0 \u00a0 \u00a0return logits<\/p>\n
# ============================================================
\n# \u4e00\u3001\u6784\u5efa\u4e00\u4e2a\u7528\u4e8e\u6f14\u793a\u7684\u5c0f\u8bcd\u8868
\n# ============================================================<\/p>\nvocab = [
\n \u00a0 \u00a0"<pad>",
\n \u00a0 \u00a0"<bos>",
\n \u00a0 \u00a0"<eos>",
\n \u00a0 \u00a0"<mask>",
\n \u00a0 \u00a0"\u4eca\u5929",
\n \u00a0 \u00a0"\u5929\u6c14",
\n \u00a0 \u00a0"\u5f88",
\n \u00a0 \u00a0"\u597d",
\n \u00a0 \u00a0"\u660e\u5929",
\n \u00a0 \u00a0"\u53ef\u80fd",
\n \u00a0 \u00a0"\u4f1a",
\n \u00a0 \u00a0"\u4e0b\u96e8",
\n]<\/p>\nword_to_id = {
\n \u00a0 \u00a0word: index
\n \u00a0 \u00a0for index, word in enumerate(vocab)
\n}<\/p>\nid_to_word = {
\n \u00a0 \u00a0index: word
\n \u00a0 \u00a0for word, index in word_to_id.items()
\n}<\/p>\nPAD_ID = word_to_id["<pad>"]
\nBOS_ID = word_to_id["<bos>"]
\nEOS_ID = word_to_id["<eos>"]
\nMASK_ID = word_to_id["<mask>"]<\/p>\n# ============================================================
\n# \u4e8c\u3001\u6784\u9020 Text Encoder \u7684\u8f93\u5165
\n# ============================================================<\/p>\n# \u539f\u59cb\u53e5\u5b50 1：
\n# <bos> \u4eca\u5929 \u5929\u6c14 \u5f88 \u597d <eos>
\n#
\n# \u63a9\u7801\u53e5\u5b50 1：
\n# <bos> \u4eca\u5929 <mask> \u5f88 \u597d <eos>
\n#
\n# \u539f\u59cb\u53e5\u5b50 2：
\n# <bos> \u660e\u5929 \u53ef\u80fd \u4f1a \u4e0b\u96e8 <eos>
\n#
\n# \u63a9\u7801\u53e5\u5b50 2：
\n# <bos> \u660e\u5929 \u53ef\u80fd <mask> \u4e0b\u96e8 <eos><\/p>\n# \u8fd9\u4e2a\u5f20\u91cf\u5c31\u662f Text Encoder \u7684\u8f93\u5165\u3002
\n# \u56e0\u6b64，Text Encoder \u7f16\u7801\u7684\u662f\u63a9\u7801\u540e\u7684\u53e5\u5b50\u3002
\nmasked_sentence_ids = torch.tensor([
\n \u00a0 \u00a0[
\n \u00a0 \u00a0 \u00a0 \u00a0BOS_ID,
\n \u00a0 \u00a0 \u00a0 \u00a0word_to_id["\u4eca\u5929"],
\n \u00a0 \u00a0 \u00a0 \u00a0MASK_ID,
\n \u00a0 \u00a0 \u00a0 \u00a0word_to_id["\u5f88"],
\n \u00a0 \u00a0 \u00a0 \u00a0word_to_id["\u597d"],
\n \u00a0 \u00a0 \u00a0 \u00a0EOS_ID,
\n \u00a0 \u00a0],
\n \u00a0 \u00a0[
\n \u00a0 \u00a0 \u00a0 \u00a0BOS_ID,
\n \u00a0 \u00a0 \u00a0 \u00a0word_to_id["\u660e\u5929"],
\n \u00a0 \u00a0 \u00a0 \u00a0word_to_id["\u53ef\u80fd"],
\n \u00a0 \u00a0 \u00a0 \u00a0MASK_ID,
\n \u00a0 \u00a0 \u00a0 \u00a0word_to_id["\u4e0b\u96e8"],
\n \u00a0 \u00a0 \u00a0 \u00a0EOS_ID,
\n \u00a0 \u00a0],
\n])<\/p>\n# ============================================================
\n# \u4e09\u3001\u6784\u9020 Text Decoder \u7684\u8f93\u5165
\n# ============================================================<\/p>\n# \u8bad\u7ec3\u9636\u6bb5\u91c7\u7528\u6559\u5e08\u5f3a\u5236\u3002
\n#
\n# \u539f\u76ee\u6807：
\n# <bos> \u4eca\u5929 \u5929\u6c14 \u5f88 \u597d <eos>
\n#
\n# Decoder \u8f93\u5165\u53bb\u6389\u6700\u540e\u4e00\u4e2a\u8bcd：
\n# <bos> \u4eca\u5929 \u5929\u6c14 \u5f88 \u597d
\n#
\n# \u76d1\u7763\u6807\u7b7e\u53bb\u6389\u7b2c\u4e00\u4e2a\u8bcd：
\n# \u4eca\u5929 \u5929\u6c14 \u5f88 \u597d <eos>
\n#
\n# \u56e0\u800c Decoder \u5b66\u4e60：
\n# \u770b\u5230 <bos> \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 -> \u9884\u6d4b \u4eca\u5929
\n# \u770b\u5230 <bos> \u4eca\u5929 \u00a0 \u00a0 \u00a0 \u00a0-> \u9884\u6d4b \u5929\u6c14
\n# \u770b\u5230 <bos> \u4eca\u5929 \u5929\u6c14 \u00a0 -> \u9884\u6d4b \u5f88
\n# …<\/p>\ndecoder_input_ids = torch.tensor([
\n \u00a0 \u00a0[
\n \u00a0 \u00a0 \u00a0 \u00a0BOS_ID,
\n \u00a0 \u00a0 \u00a0 \u00a0word_to_id["\u4eca\u5929"],
\n \u00a0 \u00a0 \u00a0 \u00a0word_to_id["\u5929\u6c14"],
\n \u00a0 \u00a0 \u00a0 \u00a0word_to_id["\u5f88"],
\n \u00a0 \u00a0 \u00a0 \u00a0word_to_id["\u597d"],
\n \u00a0 \u00a0],
\n \u00a0 \u00a0[
\n \u00a0 \u00a0 \u00a0 \u00a0BOS_ID,
\n \u00a0 \u00a0 \u00a0 \u00a0word_to_id["\u660e\u5929"],
\n \u00a0 \u00a0 \u00a0 \u00a0word_to_id["\u53ef\u80fd"],
\n \u00a0 \u00a0 \u00a0 \u00a0word_to_id["\u4f1a"],
\n \u00a0 \u00a0 \u00a0 \u00a0word_to_id["\u4e0b\u96e8"],
\n \u00a0 \u00a0],
\n])<\/p>\n# ============================================================
\n# \u56db\u3001\u6784\u9020\u76d1\u7763\u6807\u7b7e
\n# ============================================================<\/p>\ntarget_ids = torch.tensor([
\n \u00a0 \u00a0[
\n \u00a0 \u00a0 \u00a0 \u00a0word_to_id["\u4eca\u5929"],
\n \u00a0 \u00a0 \u00a0 \u00a0word_to_id["\u5929\u6c14"],
\n \u00a0 \u00a0 \u00a0 \u00a0word_to_id["\u5f88"],
\n \u00a0 \u00a0 \u00a0 \u00a0word_to_id["\u597d"],
\n \u00a0 \u00a0 \u00a0 \u00a0EOS_ID,
\n \u00a0 \u00a0],
\n \u00a0 \u00a0[
\n \u00a0 \u00a0 \u00a0 \u00a0word_to_id["\u660e\u5929"],
\n \u00a0 \u00a0 \u00a0 \u00a0word_to_id["\u53ef\u80fd"],
\n \u00a0 \u00a0 \u00a0 \u00a0word_to_id["\u4f1a"],
\n \u00a0 \u00a0 \u00a0 \u00a0word_to_id["\u4e0b\u96e8"],
\n \u00a0 \u00a0 \u00a0 \u00a0EOS_ID,
\n \u00a0 \u00a0],
\n])<\/p>\n# ============================================================
\n# \u4e94\u3001\u521b\u5efa\u6a21\u578b\u548c\u4f18\u5316\u5668
\n# ============================================================<\/p>\nmodel = MaskedSentenceRecovery(
\n \u00a0 \u00a0vocab_size=len(vocab),
\n \u00a0 \u00a0hidden_dim=128,
\n \u00a0 \u00a0num_heads=4,
\n \u00a0 \u00a0num_layers=2,
\n)<\/p>\noptimizer = torch.optim.Adam(
\n \u00a0 \u00a0model.parameters(),
\n \u00a0 \u00a0lr=1e-3,
\n)<\/p>\n# ============================================================
\n# \u516d\u3001\u8bad\u7ec3\u6a21\u578b
\n# ============================================================<\/p>\nmodel.train()<\/p>\n
for step in range(200):
\n \u00a0 \u00a0# \u524d\u5411\u4f20\u64ad：
\n \u00a0 \u00a0# \u63a9\u7801\u53e5\u5b50\u8fdb\u5165 Encoder，
\n \u00a0 \u00a0# \u53f3\u79fb\u539f\u53e5\u8fdb\u5165 Decoder\u3002
\n \u00a0 \u00a0logits = model(
\n \u00a0 \u00a0 \u00a0 \u00a0masked_sentence_ids=masked_sentence_ids,
\n \u00a0 \u00a0 \u00a0 \u00a0decoder_input_ids=decoder_input_ids,
\n \u00a0 \u00a0)<\/p>\n\u00a0 \u00a0# logits \u539f\u59cb\u5f62\u72b6：
\n \u00a0 \u00a0# [batch_size, target_length, vocab_size]
\n \u00a0 \u00a0#
\n \u00a0 \u00a0# cross_entropy \u9700\u8981：
\n \u00a0 \u00a0# \u8f93\u5165：[\u6837\u672c\u603b\u6570, vocab_size]
\n \u00a0 \u00a0# \u6807\u7b7e：[\u6837\u672c\u603b\u6570]
\n \u00a0 \u00a0loss = F.cross_entropy(
\n \u00a0 \u00a0 \u00a0 \u00a0logits.reshape(-1, len(vocab)),
\n \u00a0 \u00a0 \u00a0 \u00a0target_ids.reshape(-1),
\n \u00a0 \u00a0 \u00a0 \u00a0ignore_index=PAD_ID,
\n \u00a0 \u00a0)<\/p>\n\u00a0 \u00a0optimizer.zero_grad()
\n \u00a0 \u00a0loss.backward()
\n \u00a0 \u00a0optimizer.step()<\/p>\n\u00a0 \u00a0if (step + 1) % 50 == 0:
\n \u00a0 \u00a0 \u00a0 \u00a0print(
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0f"Step {step + 1:3d}, "
\n \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0f"loss = {loss.item():.4f}"
\n \u00a0 \u00a0 \u00a0 \u00a0)<\/p>\n# ============================================================
\n# \u4e03\u3001\u67e5\u770b\u9884\u6d4b\u7ed3\u679c
\n# ============================================================<\/p>\nmodel.eval()<\/p>\n
with torch.no_grad():
\n \u00a0 \u00a0logits = model(
\n \u00a0 \u00a0 \u00a0 \u00a0masked_sentence_ids=masked_sentence_ids,
\n \u00a0 \u00a0 \u00a0 \u00a0decoder_input_ids=decoder_input_ids,
\n \u00a0 \u00a0)<\/p>\n\u00a0 \u00a0# \u5728\u8bcd\u8868\u7ef4\u5ea6\u4e0a\u9009\u62e9\u6982\u7387\u6700\u5927\u7684\u8bcd
\n \u00a0 \u00a0predicted_ids = logits.argmax(dim=-1)<\/p>\nfor sentence_index, sentence_ids in enumerate(predicted_ids):
\n \u00a0 \u00a0predicted_words = [
\n \u00a0 \u00a0 \u00a0 \u00a0id_to_word[token_id.item()]
\n \u00a0 \u00a0 \u00a0 \u00a0for token_id in sentence_ids
\n \u00a0 \u00a0]<\/p>\n\u00a0 \u00a0print(
\n \u00a0 \u00a0 \u00a0 \u00a0f"\u7b2c {sentence_index + 1} \u4e2a\u9884\u6d4b\u53e5\u5b50：",
\n \u00a0 \u00a0 \u00a0 \u00a0" ".join(predicted_words),
\n \u00a0 \u00a0) <\/p>\n0.\u6458\u8981<\/h2>\n
\u624b\u8bed\u7ffb\u8bd1（Sign Language Translation，SLT）\u662f\u4e00\u9879\u5177\u6709\u6311\u6218\u6027\u7684\u4efb\u52a1，\u56e0\u4e3a\u5b83\u5177\u6709\u8de8\u6a21\u6001\u3001\u8de8\u9886\u57df\u7684\u6027\u8d28，\u9700\u8981\u5c06\u89c6\u89c9\u2014\u624b\u52bf\u8bed\u8a00\u7ffb\u8bd1\u4e3a\u6587\u672c\u3002\u8bb8\u591a\u4ee5\u5f80\u7684\u65b9\u6cd5\u91c7\u7528\u4e00\u79cd\u4e2d\u95f4\u8868\u793a，\u5373\u624b\u8bed\u8bcd\u6c47\u6807\u6ce8\u5e8f\u5217（gloss sequence），\u6765\u8f85\u52a9\u5b8c\u6210\u624b\u8bed\u7ffb\u8bd1，\u4ece\u800c\u5c06\u8be5\u4efb\u52a1\u8f6c\u5316\u4e3a\u4e00\u4e2a\u4e24\u9636\u6bb5\u8fc7\u7a0b：\u9996\u5148\u8fdb\u884c\u624b\u8bed\u8bc6\u522b（Sign Language Recognition，SLR），\u7136\u540e\u8fdb\u884c\u624b\u8bed\u7ffb\u8bd1（SLT）\u3002\u7136\u800c，\u5e26\u6709 gloss \u6807\u6ce8\u7684\u624b\u8bed\u6570\u636e\u5341\u5206\u7a00\u7f3a，\u540c\u65f6，\u4e2d\u95f4\u5c42\u7684 gloss \u8868\u793a\u8fd8\u4f1a\u9020\u6210\u4fe1\u606f\u74f6\u9888，\u8fd9\u4e9b\u95ee\u9898\u963b\u788d\u4e86\u624b\u8bed\u7ffb\u8bd1\u4efb\u52a1\u7684\u8fdb\u4e00\u6b65\u53d1\u5c55\u3002<\/p>\n
\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u6311\u6218，\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u57fa\u4e8e\u89c6\u89c9\u2014\u8bed\u8a00\u9884\u8bad\u7ec3\u7684\u65b0\u578b\u65e0 gloss \u624b\u8bed\u7ffb\u8bd1\u65b9\u6cd5，\u79f0\u4e3a GFSLT-VLP（Gloss-Free Sign Language Translation based on Visual-Language Pretraining）\u3002\u8be5\u65b9\u6cd5\u5b8c\u5168\u4e0d\u501f\u52a9\u4efb\u4f55 gloss \u6807\u6ce8，\u800c\u662f\u901a\u8fc7\u7ee7\u627f\u9884\u8bad\u7ec3\u6a21\u578b\u4e2d\u9762\u5411\u8bed\u8a00\u7684\u5148\u9a8c\u77e5\u8bc6\u6765\u63d0\u5347\u624b\u8bed\u7ffb\u8bd1\u6027\u80fd\u3002\u672c\u6587\u7684\u65b9\u6cd5\u5305\u542b\u4e24\u4e2a\u9636\u6bb5：<\/p>\n
\u7b2c\u4e00\u9636\u6bb5，\u5c06\u5bf9\u6bd4\u5f0f\u8bed\u8a00\u2014\u56fe\u50cf\u9884\u8bad\u7ec3（Contrastive Language-Image Pretraining，CLIP）\u4e0e\u63a9\u7801\u81ea\u76d1\u7763\u5b66\u4e60\u76f8\u7ed3\u5408，\u6784\u5efa\u76f8\u5e94\u7684\u9884\u8bad\u7ec3\u4efb\u52a1\u3002\u8fd9\u4e9b\u4efb\u52a1\u4e00\u65b9\u9762\u7528\u4e8e\u7f29\u5c0f\u89c6\u89c9\u8868\u793a\u4e0e\u6587\u672c\u8868\u793a\u4e4b\u95f4\u7684\u8bed\u4e49\u5dee\u8ddd，\u53e6\u4e00\u65b9\u9762\u7528\u4e8e\u6062\u590d\u88ab\u638d\u7801\u7684\u53e5\u5b50\u5185\u5bb9\u3002<\/p>\n
\u7b2c\u4e8c\u9636\u6bb5，\u6784\u5efa\u4e00\u4e2a\u7aef\u5230\u7aef\u7684\u7f16\u7801\u5668\u2014\u89e3\u7801\u5668\u5f0f\u7f51\u7edc\u67b6\u6784，\u5e76\u7ee7\u627f\u7b2c\u4e00\u9636\u6bb5\u4e2d\u9884\u8bad\u7ec3\u5f97\u5230\u7684\u89c6\u89c9\u7f16\u7801\u5668\u548c\u6587\u672c\u89e3\u7801\u5668\u53c2\u6570\u3002<\/p>\n
\u8fd9\u4e9b\u65b0\u8bbe\u8ba1\u7684\u6709\u673a\u7ed3\u5408，\u4f7f\u6a21\u578b\u80fd\u591f\u5b66\u4e60\u5230\u66f4\u52a0\u9c81\u68d2\u7684\u624b\u8bed\u8868\u793a，\u5e76\u663e\u8457\u63d0\u5347\u65e0 gloss \u624b\u8bed\u7ffb\u8bd1\u7684\u6027\u80fd\u3002<\/p>\n
\u5177\u4f53\u800c\u8a00，\u4e0e\u5f53\u524d\u6700\u5148\u8fdb\u7684\u65e0 gloss \u624b\u8bed\u7ffb\u8bd1\u65b9\u6cd5\u76f8\u6bd4，\u672c\u6587\u65b9\u6cd5\u5728 PHOENIX14T \u6570\u636e\u96c6\u4e0a\u7684 BLEU-4 \u5206\u6570\u53d6\u5f97\u4e86\u81f3\u5c11 5\u5206\u4ee5\u4e0a\u7684\u63d0\u5347，\u5728 CSL-Daily \u6570\u636e\u96c6\u4e0a\u53d6\u5f97\u4e86\u81f3\u5c11 3\u5206\u4ee5\u4e0a\u7684\u63d0\u5347，\u8fbe\u5230\u4e86\u6b64\u524d\u672a\u6709\u8fc7\u7684\u6539\u8fdb\u5e45\u5ea6\u3002\u6b64\u5916，\u5728 PHOENIX14T \u6570\u636e\u96c6\u4e0a，\u4e0e\u5927\u591a\u6570\u4f9d\u8d56 gloss \u6807\u6ce8\u7684\u65b9\u6cd5\u76f8\u6bd4，\u672c\u6587\u65b9\u6cd5\u4e5f\u53d6\u5f97\u4e86\u5177\u6709\u7ade\u4e89\u529b\u7684\u7ed3\u679c\u3002<\/p>\n
1.Introduction<\/h2>\n
\u624b\u8bed\u662f\u804b\u4eba\u7fa4\u4f53\u4e4b\u95f4\u4e3b\u8981\u7684\u4ea4\u6d41\u5a92\u4ecb\u3002\u4e3a\u4e86\u4fc3\u8fdb\u804b\u4eba\u4e0e\u542c\u969c\u7a0b\u5ea6\u8f83\u8f7b\u8005\u4e4b\u95f4\u7684\u6709\u6548\u6c9f\u901a，\u53d1\u5c55\u624b\u8bed\u7ffb\u8bd1（Sign Language Translation，SLT）\u6280\u672f\u662f\u4e00\u4e2a\u5f88\u6709\u524d\u666f\u7684\u65b9\u5411\u3002\u624b\u8bed\u7ffb\u8bd1\u662f\u6307\u5c06\u624b\u8bed\u7ffb\u8bd1\u6210\u6d41\u7545\u7684\u53e3\u8bed\u8bed\u8a00\u53e5\u5b50\u3002\u7531\u4e8e\u624b\u8bed\u7ffb\u8bd1\u5177\u6709\u8de8\u6a21\u6001\u7ffb\u8bd1\u7684\u6027\u8d28，\u5e76\u4e14\u53ef\u7528\u7684\u6807\u6ce8\u6570\u636e\u5341\u5206\u7a00\u7f3a，\u56e0\u6b64\u5b83\u6bd4\u4f20\u7edf\u7684\u795e\u7ecf\u673a\u5668\u7ffb\u8bd1（Neural Machine Translation，NMT）\u66f4\u5177\u6311\u6218\u6027\u3002<\/p>\n
\u8fd1\u5e74\u6765，\u8d8a\u6765\u8d8a\u591a\u7684\u7814\u7a76\u5de5\u4f5c\u901a\u8fc7\u76f4\u63a5\u6216\u95f4\u63a5\u5730\u4f7f\u7528\u4e00\u79cd\u4e2d\u95f4\u8868\u793a\u2014\u2014\u624b\u8bed gloss，\u6765\u63d0\u5347\u624b\u8bed\u7ffb\u8bd1\u6027\u80fd\u3002\u5982 Figure 1(a) \u6240\u793a，gloss \u662f\u5bf9\u8fde\u7eed\u89c6\u9891\u4e2d\u5404\u4e2a\u624b\u8bed\u5355\u5143\u7684\u4e00\u79cd\u7b80\u5316\u8868\u793a\u3002\u5c3d\u7ba1\u4e0e\u7aef\u5230\u7aef\u7684\u65e0 gloss \u65b9\u6cd5\u76f8\u6bd4（\u5982 Figure 1(b) \u6240\u793a），\u57fa\u4e8e gloss \u7684\u65b9\u6cd5\u5df2\u7ecf\u663e\u8457\u63d0\u5347\u4e86\u624b\u8bed\u7ffb\u8bd1\u6027\u80fd，\u4f46\u524d\u8005\u4ecd\u7136\u5b58\u5728\u4ee5\u4e0b\u95ee\u9898：\u7b2c\u4e00，gloss \u6807\u6ce8\u662f\u4e00\u9879\u52b3\u52a8\u5bc6\u96c6\u578b\u5de5\u4f5c，\u9700\u8981\u8fdb\u884c\u7ec6\u7c92\u5ea6\u5bf9\u9f50，\u5e76\u7531\u4e13\u4e1a\u4eba\u5458\u5b8c\u6210\u6807\u6ce8，\u8fd9\u6781\u5927\u5730\u9650\u5236\u4e86\u57fa\u4e8e gloss \u7684\u624b\u8bed\u7ffb\u8bd1\u65b9\u6cd5\u7684\u53ef\u6269\u5c55\u6027；\u7b2c\u4e8c，\u57fa\u4e8e gloss \u7684\u65b9\u6cd5\u5728\u4e2d\u95f4\u5c42 gloss \u8868\u793a\u5904\u5f15\u5165\u4e86\u4fe1\u606f\u74f6\u9888，\u8fd9\u9650\u5236\u4e86\u7f51\u7edc\u7406\u89e3\u624b\u8bed\u7684\u80fd\u529b，\u56e0\u4e3a\u7ffb\u8bd1\u6a21\u578b\u7684\u6027\u80fd\u4e0a\u9650\u53d7\u5236\u4e8e\u8bad\u7ec3\u65f6\u4f7f\u7528\u7684\u624b\u8bed gloss \u6807\u6ce8\u8d28\u91cf\u3002<\/p>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/07\/20260729010841-6a6952996b783.png\")
<\/p>\n\u53d7\u5230 CLIP \u7684\u542f\u53d1，CLIP \u5229\u7528\u81ea\u7136\u8bed\u8a00\u76d1\u7763\u6765\u5b66\u4e60\u56fe\u50cf\u8868\u793a；\u4f5c\u8005\u53d1\u73b0，\u4ece\u624b\u8bed\u89c6\u9891\u4e2d\u5b66\u4e60\u8bed\u8a00\u6307\u793a\u7684\u89c6\u89c9\u8868\u793a，\u53ef\u4ee5\u6210\u4e3a\u4e00\u79cd\u6709\u6548\u7684\u624b\u8bed\u7ffb\u8bd1\u9884\u8bad\u7ec3\u4efb\u52a1，\u56e0\u4e3a\u5b83\u80fd\u591f\u5728\u89c6\u89c9\u624b\u8bed\u52a8\u4f5c\u4e0e\u8bed\u8a00\u4e0a\u4e0b\u6587\u4e4b\u95f4\u5efa\u7acb\u6f5c\u5728\u8054\u7cfb\u3002\u7136\u800c，\u5c06 CLIP \u76f4\u63a5\u5e94\u7528\u4e8e\u624b\u8bed\u7ffb\u8bd1\u4ecd\u7136\u9762\u4e34\u4e24\u4e2a\u6311\u6218：\u7b2c\u4e00，\u5b83\u4e0d\u5177\u5907\u4e3a\u624b\u8bed\u7ffb\u8bd1\u8054\u5408\u9884\u8bad\u7ec3 Visual Encoder \u548c Text Decoder \u7684\u80fd\u529b；\u7b2c\u4e8c，\u8fd9\u79cd\u9884\u8bad\u7ec3\u4efb\u52a1\u9700\u8981\u8db3\u591f\u591a\u7684\u624b\u8bed\u7ffb\u8bd1\u6570\u636e\u3002\u4e3a\u89e3\u51b3\u8fd9\u4e9b\u969c\u788d，\u4f5c\u8005\u9700\u8981\u56de\u7b54\u4e24\u4e2a\u5173\u952e\u95ee\u9898：\u7b2c\u4e00，\u5982\u4f55\u5229\u7528\u6709\u9650\u7684\u624b\u8bed\u7ffb\u8bd1\u6570\u636e\u96c6，\u9ad8\u6548\u5730\u5b8c\u6210\u8054\u5408\u9884\u8bad\u7ec3？\u7b2c\u4e8c，\u5982\u4f55\u4fdd\u8bc1\u9884\u8bad\u7ec3\u6a21\u578b\u80fd\u591f\u4e3a\u4e0b\u6e38\u624b\u8bed\u7ffb\u8bd1\u4efb\u52a1\u63d0\u4f9b\u6700\u6709\u6548\u7684\u5e2e\u52a9？<\/p>\n
\u4e3a\u4e86\u514b\u670d\u7b2c\u4e00\u4e2a\u6311\u6218，\u4f5c\u8005\u5206\u522b\u4ece\u7b97\u6cd5\u5c42\u9762\u548c\u6570\u636e\u5c42\u9762\u63d0\u51fa\u4e86\u89e3\u51b3\u65b9\u6848\u3002\u5728\u7b97\u6cd5\u5c42\u9762，\u4f5c\u8005\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u9884\u8bad\u7ec3\u65b9\u6cd5，\u79f0\u4e3a\u89c6\u89c9\u2014\u8bed\u8a00\u9884\u8bad\u7ec3（Visual-Language Pretraining，VLP），\u5982 Figure 2(a) \u6240\u793a\u3002\u8be5\u65b9\u6cd5\u5c06\u63a9\u7801\u81ea\u76d1\u7763\u5b66\u4e60\u4e0e CLIP \u7ed3\u5408\u8d77\u6765\u3002\u5177\u4f53\u6765\u8bf4，\u4f5c\u8005\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u9884\u6587\u672c\u4efb\u52a1，\u5c06\u89c6\u89c9\u8868\u793a\u548c\u6587\u672c\u8868\u793a\u5bf9\u9f50\u5230\u4e00\u4e2a\u5171\u4eab\u7684\u591a\u6a21\u6001\u8bed\u4e49\u7a7a\u95f4\u4e2d，\u4ece\u800c\u5f15\u5bfc Visual Encoder \u5b66\u4e60\u8bed\u8a00\u6307\u793a\u7684\u89c6\u89c9\u8868\u793a\u3002\u4e0e\u6b64\u540c\u65f6，\u4f5c\u8005\u5728\u9884\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5f15\u5165\u63a9\u7801\u81ea\u76d1\u7763\u5b66\u4e60，\u5e2e\u52a9 Text Decoder \u6355\u6349\u624b\u8bed\u76ee\u6807\u53e5\u5b50\u7684\u53e5\u6cd5\u548c\u8bed\u4e49\u5c5e\u6027\u3002\u5728\u6570\u636e\u5c42\u9762，\u4f5c\u8005\u7814\u7a76\u4e86\u4e00\u7ec4\u7528\u4e8e\u624b\u8bed\u89c6\u9891\u7684\u5f3a\u6570\u636e\u589e\u5f3a\u6280\u672f，\u4ee5\u589e\u52a0\u89c6\u89c9\u6570\u636e\u7684\u591a\u6837\u6027；\u8fd9\u4e00\u65b9\u9762\u5728\u4ee5\u5f80\u7684\u624b\u8bed\u7ffb\u8bd1\u65b9\u6cd5\u4e2d\u7ecf\u5e38\u88ab\u5ffd\u89c6\u3002<\/p>\n
\u4e3a\u4e86\u5e94\u5bf9\u7b2c\u4e8c\u4e2a\u95ee\u9898，\u5982 Figure 2(b) \u6240\u793a，\u4f5c\u8005\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u7aef\u5230\u7aef\u7684\u65e0 gloss \u624b\u8bed\u7ffb\u8bd1\u67b6\u6784，\u79f0\u4e3a GFSLT\u3002\u8be5\u67b6\u6784\u91c7\u7528\u7f16\u7801\u5668\u2014\u89e3\u7801\u5668\u5f62\u5f0f，\u5e76\u7ee7\u627f\u7b2c\u4e00\u9636\u6bb5\u9884\u8bad\u7ec3\u5f97\u5230\u7684 Visual Encoder \u548c Text Decoder \u53c2\u6570\u3002GFSLT \u80fd\u591f\u5728\u4e0d\u9700\u8981\u4efb\u4f55\u4e2d\u95f4\u6b65\u9aa4\u6216\u4e2d\u95f4\u6307\u5bfc\u7684\u60c5\u51b5\u4e0b，\u76f4\u63a5\u5c06\u89c6\u89c9\u8868\u793a\u8f6c\u6362\u6210\u81ea\u7136\u8bed\u8a00\u53e5\u5b50\u3002\u6b64\u5916，\u4e0e\u5176\u4ed6\u53ea\u5fae\u8c03 Visual Encoder \u4e2d\u7a7a\u95f4\u7279\u5f81\u63d0\u53d6\u5668，\u5373\u89c6\u89c9\u5d4c\u5165\u6a21\u5757\u7684\u65b9\u6cd5\u4e0d\u540c，\u4f5c\u8005\u5c06\u7a7a\u95f4\u7279\u5f81\u63d0\u53d6\u5668\u4e0e\u65f6\u95f4\u5173\u7cfb\u5efa\u6a21\u7f51\u7edc，\u5373 Transformer Encoder，\u4f5c\u4e3a\u4e00\u4e2a\u7edf\u4e00\u6574\u4f53\u8fdb\u884c\u5fae\u8c03\u3002<\/p>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/07\/20260729010842-6a69529a29a94.png\")
<\/p>\n- \n
- （a）\u5305\u542b\u4e24\u6761\u540c\u65f6\u8fdb\u884c\u7684\u8bad\u7ec3\u5206\u652f\n
- \n
- \u89c6\u9891\u2014\u6587\u672c\u5bf9\u6bd4\u5b66\u4e60\u5206\u652f，\u4e24\u79cd\u7279\u5f81\u7ecf\u8fc7\u5404\u81ea\u7684\u6295\u5f71\u5934\u540e\u8fdb\u5165\u5171\u4eab\u591a\u6a21\u6001\u7a7a\u95f4，\u8ba1\u7b97\u89c6\u9891\u548c\u6587\u672c\u7684\u76f8\u4f3c\u5ea6，\u4f7f\u6b63\u786e\u7684\u89c6\u9891\u2014\u53e5\u5b50\u5bf9\u66f4\u52a0\u63a5\u8fd1（\u56fe\u4e2d\u5b9e\u7ebf）\u3002<\/li>\n
- \u63a9\u7801\u53e5\u5b50\u6062\u590d\u5206\u652f，\u8d1f\u8d23\u9884\u8bad\u7ec3 Text Decoder \u7684\u8bed\u8a00\u7406\u89e3\u4e0e\u751f\u6210\u80fd\u529b（\u56fe\u4e2d\u865a\u7ebf）\u3002<\/li>\n<\/ul>\n<\/li>\n
- （b）Visual Encoder \u548c Text Decoder \u7ee7\u627f\u7b2c\u4e00\u9636\u6bb5\u53c2\u6570，\u7136\u540e\u4f7f\u7528\u89c6\u9891\u2014\u53e5\u5b50\u7ffb\u8bd1\u76ee\u6807\u8fdb\u884c\u7aef\u5230\u7aef\u5fae\u8c03\u3002<\/li>\n<\/ul>\n
\u603b\u800c\u8a00\u4e4b，\u672c\u6587\u7684\u4e3b\u8981\u8d21\u732e\u5982\u4e0b：<\/p>\n
- \n
- \u5728\u4e0d\u4f7f\u7528 gloss \u6807\u6ce8\u7684\u60c5\u51b5\u4e0b，\u672c\u6587\u5728\u624b\u8bed\u7ffb\u8bd1\u7684 BLEU-4 \u6307\u6807\u4e0a\u53d6\u5f97\u4e86\u524d\u6240\u672a\u6709\u7684\u63d0\u5347\u3002\u5177\u4f53\u6765\u8bf4，\u4e0e\u5f53\u65f6\u6700\u5148\u8fdb\u7684\u65e0 gloss \u624b\u8bed\u7ffb\u8bd1\u65b9\u6cd5\u76f8\u6bd4，\u672c\u6587\u65b9\u6cd5\u5728 PHOENIX14T \u6570\u636e\u96c6\u548c CSL-Daily \u6570\u636e\u96c6\u4e0a\u5206\u522b\u53d6\u5f97\u4e86\u81f3\u5c11 5 \u5206\u548c\u81f3\u5c11 3 \u5206\u7684\u63d0\u5347\u3002\u4f5c\u8005\u8ba4\u4e3a，\u8fd9\u4e9b\u63d0\u5347\u4ee3\u8868\u4e86\u65e0 gloss \u624b\u8bed\u7ffb\u8bd1\u4efb\u52a1\u4e0a\u7684\u4e00\u6b21\u91cd\u8981\u7a81\u7834\u3002<\/li>\n
- \u636e\u4f5c\u8005\u6240\u77e5，\u8fd9\u662f\u9996\u6b21\u5728\u65e0 gloss \u624b\u8bed\u7ffb\u8bd1\u4efb\u52a1\u4e2d\u5f15\u5165 VLP \u7b56\u7565，\u5c06\u89c6\u89c9\u8868\u793a\u4e0e\u6587\u672c\u8868\u793a\u5bf9\u9f50\u5230\u4e00\u4e2a\u5171\u4eab\u7684\u8bed\u4e49\u7a7a\u95f4\u4e2d\u3002<\/li>\n
- \u4f5c\u8005\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u9884\u8bad\u7ec3\u8303\u5f0f，\u5c06\u63a9\u7801\u81ea\u76d1\u7763\u5b66\u4e60\u4e0e\u5bf9\u6bd4\u5f0f\u8bed\u8a00\u2014\u56fe\u50cf\u9884\u8bad\u7ec3\u7ed3\u5408\u8d77\u6765，\u4ee5\u4fc3\u8fdb\u65e0 gloss \u624b\u8bed\u7ffb\u8bd1\u3002\u4e0e\u4ee5\u5f80\u65b9\u6cd5\u76f8\u6bd4，\u8be5\u65b9\u6cd5\u53d6\u5f97\u4e86\u663e\u8457\u6539\u8fdb，\u5e76\u6709\u6f5c\u529b\u5927\u5e45\u63d0\u5347\u624b\u8bed\u7ffb\u8bd1\u7cfb\u7edf\u7684\u51c6\u786e\u6027\u548c\u6548\u7387\u3002<\/li>\n<\/ul>\n
2.\u76f8\u5173\u5de5\u4f5c<\/h2>\n
\u603b\u4f53\u800c\u8a00，\u624b\u8bed\u7ffb\u8bd1（Sign Language Translation，SLT）\u65b9\u6cd5\u53ef\u4ee5\u5206\u4e3a\u4e24\u7c7b：\u57fa\u4e8e gloss \u7684\u65b9\u6cd5\u548c\u65e0 gloss \u7684\u65b9\u6cd5\u3002\u5728\u7b80\u8981\u56de\u987e\u8fd9\u4e24\u4e2a\u65b9\u5411\u7684\u7814\u7a76\u5de5\u4f5c\u4e4b\u524d，\u6211\u4eec\u9996\u5148\u4ecb\u7ecd\u624b\u8bed\u8bc6\u522b（Sign Language Recognition，SLR）\u4efb\u52a1，\u56e0\u4e3a\u5b83\u662f\u57fa\u4e8e gloss \u7684\u624b\u8bed\u7ffb\u8bd1\u65b9\u6cd5\u4e2d\u7684\u4e00\u4e2a\u5173\u952e\u6b65\u9aa4\u3002<\/p>\n
2.1.\u624b\u8bed\u8bc6\u522b<\/h3>\n
\u624b\u8bed\u8bc6\u522b（SLR）\u5305\u542b\u4e24\u79cd\u4e0d\u540c\u7684\u4efb\u52a1：\u5b64\u7acb\u624b\u8bed\u8bc6\u522b（Isolated Sign Language Recognition，ISLR）\u548c\u8fde\u7eed\u624b\u8bed\u8bc6\u522b（Continuous Sign Language Recognition，CSLR）\u3002ISLR \u7684\u76ee\u6807\u662f\u5c06\u4e00\u4e2a\u5b64\u7acb\u7684\u624b\u8bed\u52a8\u4f5c\u8f6c\u6362\u4e3a\u5bf9\u5e94\u7684\u5355\u4e2a\u624b\u8bed\u8bcd\u6c47，\u8fd9\u4e00\u4efb\u52a1\u4e0e\u5b64\u7acb\u624b\u52bf\u8bc6\u522b\u4efb\u52a1\u6709\u4e9b\u76f8\u4f3c\u3002CSLR \u5219\u662f\u4e00\u9879\u66f4\u5177\u6311\u6218\u6027\u7684\u4efb\u52a1，\u5b83\u81f4\u529b\u4e8e\u5c06\u4e00\u6bb5\u8fde\u7eed\u7684\u624b\u8bed\u89c6\u9891\u8bc6\u522b\u4e3a\u6309\u987a\u5e8f\u6392\u5217\u7684\u624b\u8bed\u8bcd\u6c47，\u8fd9\u4e9b\u8bcd\u6c47\u5e8f\u5217\u88ab\u79f0\u4e3a gloss \u5e8f\u5217\u3002\u4ee5\u5f80\u7684\u624b\u8bed\u7ffb\u8bd1\u7814\u7a76\u901a\u5e38\u5c06 CSLR \u4f5c\u4e3a\u4e00\u4e2a\u9884\u5907\u4efb\u52a1，\u7528\u4e8e\u9884\u6d4b gloss，\u6216\u8005\u7528\u4e8e\u83b7\u5f97\u66f4\u597d\u7684\u89c6\u89c9\u8868\u793a\u3002\u8fd9\u7c7b\u65b9\u6cd5\u901a\u5e38\u5bf9 CSLR \u7684\u51c6\u786e\u7387\u6709\u5f88\u9ad8\u7684\u8981\u6c42\u3002\u7136\u800c，\u5728\u672c\u6587\u4e2d，\u4f5c\u8005\u5b8c\u5168\u820d\u5f03\u4e86 gloss \u5e8f\u5217，\u5e76\u63a2\u7d22\u4e00\u79cd\u65b0\u7684\u65e0 gloss \u624b\u8bed\u7ffb\u8bd1\u65b9\u6cd5\u3002<\/p>\n
2.2.\u57fa\u4e8e Gloss \u7684\u624b\u8bed\u7ffb\u8bd1<\/h3>\n
\u4e3a\u4e86\u63d0\u5347\u624b\u8bed\u7ffb\u8bd1\u6027\u80fd，\u4e00\u4e9b\u7814\u7a76\u91c7\u7528\u624b\u8bed gloss \u4f5c\u4e3a\u4e2d\u95f4\u5c42\u8868\u793a\u3002SLRT \u9996\u6b21\u5f15\u5165\u4e86\u4e00\u79cd\u57fa\u4e8e Transformer \u7684\u7f16\u7801\u5668\u2014\u89e3\u7801\u5668\u6846\u67b6，\u4ee5\u6267\u884c\u7aef\u5230\u7aef\u624b\u8bed\u7ffb\u8bd1\u3002\u8be5\u65b9\u6cd5\u5229\u7528\u8fde\u63a5\u65f6\u5e8f\u5206\u7c7b（Connectionist Temporal Classification，CTC）\u635f\u5931，\u5728\u624b\u8bed\u89c6\u89c9\u8868\u793a\u4e0e gloss \u5e8f\u5217\u4e4b\u95f4\u8fdb\u884c\u8f6f\u5339\u914d，\u4ece\u800c\u63d0\u5347\u6a21\u578b\u6027\u80fd\u3002STMC-T \u91c7\u7528\u591a\u7ebf\u7d22\u5b66\u4e60\u7684\u65b9\u5f0f\u7406\u89e3\u624b\u8bed，\u5e76\u901a\u8fc7\u5f15\u5165\u7ebf\u7d22\u5185\u90e8\u548c\u7ebf\u7d22\u4e4b\u95f4\u7684 CTC \u635f\u5931\u6765\u5efa\u6a21\u5e8f\u5217\u4fe1\u606f\u3002SignBack \u5c1d\u8bd5\u5c06\u56de\u8bd1（back-translation） \u7b49\u5148\u8fdb\u7684\u673a\u5668\u7ffb\u8bd1\u6280\u672f\u5f15\u5165\u624b\u8bed\u7ffb\u8bd1\u3002\u6b64\u5916，\u5f97\u76ca\u4e8e\u8fc1\u79fb\u5b66\u4e60\u5728\u795e\u7ecf\u673a\u5668\u7ffb\u8bd1\u4e2d\u7684\u6210\u529f\u5e94\u7528，Chen \u7b49\u4eba\u9996\u6b21\u5c1d\u8bd5\u5c06\u5927\u578b\u8bed\u8a00\u6a21\u578b\u5f15\u5165\u624b\u8bed\u7ffb\u8bd1\u3002\u4e0a\u8ff0\u6240\u6709\u65b9\u6cd5\u5728\u624b\u8bed\u7ffb\u8bd1\u6a21\u578b\u7684\u8bad\u7ec3\u8fc7\u7a0b\u4e2d，\u90fd\u76f4\u63a5\u6216\u95f4\u63a5\u5730\u4f7f\u7528\u4e86 gloss \u6807\u6ce8\u3002\u7136\u800c，\u5728\u672c\u6587\u7684\u65b9\u6cd5\u4e2d，\u4f5c\u8005\u5b8c\u5168\u820d\u5f03\u4e86 gloss \u6807\u6ce8，\u56e0\u4e3a gloss \u7684\u5b58\u5728\u9650\u5236\u4e86\u624b\u8bed\u6570\u636e\u96c6\u89c4\u6a21\u7684\u6269\u5927\u3002\u4f5c\u4e3a\u66ff\u4ee3，\u672c\u6587\u5f15\u5165\u4e86\u4e00\u79cd\u66f4\u52a0\u901a\u7528\u7684\u624b\u8bed\u9884\u8bad\u7ec3\u8bbe\u8ba1\u3002<\/p>\n
- \n
- Transformer \u7f16\u7801\u5668\u2014\u89e3\u7801\u5668\u5982\u4f55\u6267\u884c\u7aef\u5230\u7aef\u624b\u8bed\u7ffb\u8bd1？\u57fa\u672c\u6d41\u7a0b：\u624b\u8bed\u89c6\u9891\u2192\u89c6\u89c9\u7279\u5f81（2DCNN）\u2192Transformer\u00a0Encoder（\u6bcf\u4e00\u5e27\u7684\u7279\u5f81\u5148\u52a0\u5165\u4f4d\u7f6e\u7f16\u7801，\u7136\u540e\u7406\u89e3\u6574\u6bb5\u52a8\u4f5c\u5305\u62ec\u5176\u524d\u540e\u6587）\u2192Transformer\u00a0Decoder（\u9010\u8bcd\u751f\u6210）\u2192\u81ea\u7136\u8bed\u8a00\u53e5\u5b50；\u5728 Encoder \u540e\u8fd8\u6709\u53e6\u4e00\u4e2a\u4f7f\u7528 CTC \u4f5c\u4e3a\u635f\u5931\u51fd\u6570\u7684\u8bc6\u522b\u5206\u652f，\u5b83\u4eec\u5171\u540c\u4f18\u5316 Encoder\u3002<\/li>\n
- \u591a\u7ebf\u7d22\u5b66\u4e60？\u4e0d\u8ba9\u4e00\u4e2a\u7f51\u7edc\u53ea\u5173\u6ce8\u6700\u660e\u663e\u7684\u624b\u90e8\u52a8\u4f5c，\u800c\u662f\u5206\u522b\u63d0\u53d6\u591a\u79cd\u89c6\u89c9\u4fe1\u606f，\u518d\u5b66\u4e60\u5b83\u4eec\u600e\u6837\u5171\u540c\u8868\u8fbe\u624b\u8bed\u542b\u4e49\u3002\u201c\u7ebf\u7d22\u5185\u90e8\u201d：\u5bf9\u4e8e\u6bcf\u4e00\u79cd\u7ebf\u7d22，\u5982\u624b\u90e8\u3001\u9762\u90e8\u7b49，\u5355\u72ec\u5b66\u4e60\u5b83\u4eec\u81ea\u8eab\u72ec\u7279\u7684\u65f6\u5e8f\u89c4\u5f8b，\u907f\u514d\u6709\u53ef\u80fd\u5f31\u5c0f\u4f46\u91cd\u8981\u7684\u9762\u90e8\u4fe1\u606f\u88ab\u624b\u90e8\u7279\u5f81\u6df9\u6ca1，\u6bcf\u6761\u7ebf\u7d22\u4f7f\u7528 CTC \u662f\u4e3a\u4e86\u907f\u514d\u67d0\u6761\u5206\u652f\u4ec0\u4e48\u90fd\u5b66\u4e0d\u5230，\u8fd9\u662f\u8f85\u52a9\u76ee\u6807；\u201c\u7ebf\u7d22\u4e4b\u95f4\u201d：\u878d\u5408\u4e0d\u540c\u7ebf\u7d22\u5e76\u5b66\u4e60\u5b83\u4eec\u4e4b\u95f4\u7684\u914d\u5408\u5173\u7cfb，\u4e0d\u540c\u7ebf\u7d22\u4e0d\u4ec5\u53ef\u80fd\u5728\u540c\u4e00\u65f6\u95f4\u53d1\u751f，\u4e5f\u53ef\u80fd\u5b58\u5728\u65f6\u95f4\u4e0a\u7684\u5148\u540e\u914d\u5408\u3002STMC \u7684\u7ebf\u7d22\u5185\u90e8\u8def\u5f84\u5206\u522b\u4fdd\u7559\u5404\u7ebf\u7d22\u7279\u5f81，\u7ebf\u7d22\u4e4b\u95f4\u8def\u5f84\u5219\u5728\u4e0d\u540c\u65f6\u95f4\u5c3a\u5ea6\u4e0a\u878d\u5408\u8fd9\u4e9b\u7279\u5f81，\u8fd9\u662f\u4e3b\u8981\u76ee\u6807，\u4f9d\u8d56\u5b83\u5b8c\u6210\u6700\u7ec8\u8bc6\u522b\u3002\u5b83\u662f\u4e00\u4e2a\u7aef\u5230\u7aef\u7f51\u7edc，\u5185\u90e8\u5305\u542b\u5171\u4eab\u6a21\u5757（\u9aa8\u5e72\u89c6\u89c9\u7f51\u7edc）\u3001\u591a\u4e2a\u5355\u7ebf\u7d22\u5206\u652f\u548c\u4e00\u4e2a\u878d\u5408\u5206\u652f\u3002<\/li>\n<\/ul>\n
2.3.\u65e0 Gloss \u624b\u8bed\u7ffb\u8bd1<\/h3>\n
\u201c\u65e0 gloss \u624b\u8bed\u7ffb\u8bd1\u201d\u662f\u6307\u5728\u6574\u4e2a\u8bad\u7ec3\u548c\u6d4b\u8bd5\u8fc7\u7a0b\u4e2d\u90fd\u4e0d\u5b58\u5728 gloss \u76d1\u7763，\u5176\u4e2d\u5305\u62ec\u9884\u8bad\u7ec3\u9636\u6bb5\u548c\u5fae\u8c03\u9636\u6bb5\u3002NSLT \u4f7f\u7528 CNN \u4e0e RNN \u7684\u7ec4\u5408\u6267\u884c\u7aef\u5230\u7aef\u624b\u8bed\u7ffb\u8bd1，\u5176\u4e2d CNN \u7528\u4e8e\u5b66\u4e60\u624b\u8bed\u7684\u89c6\u89c9\u7279\u5f81，\u800c\u5e26\u6709\u6ce8\u610f\u529b\u673a\u5236\u7684 RNN \u5219\u8d1f\u8d23\u5e8f\u5217\u5efa\u6a21\u548c\u6587\u672c\u5efa\u6a21\u3002TSPNet \u4f7f\u7528\u5c3a\u5ea6\u95f4\u6ce8\u610f\u529b\u548c\u5c3a\u5ea6\u5185\u6ce8\u610f\u529b，\u901a\u8fc7\u6355\u6349\u624b\u8bed\u89c6\u9891\u4e2d\u7684\u5c40\u90e8\u4e0a\u4e0b\u6587\u4e0e\u5168\u5c40\u4e0a\u4e0b\u6587\u6765\u589e\u5f3a\u89c6\u89c9\u7279\u5f81\u5b66\u4e60\u3002GASLT \u63ed\u793a\u4e86 gloss \u6807\u6ce8\u5bf9\u624b\u8bed\u7ffb\u8bd1\u7684\u91cd\u8981\u4f5c\u7528，\u5e76\u5f15\u5165\u4e86 gloss-attention，\u4ee5\u5229\u7528 gloss \u6240\u5e26\u6765\u7684\u76f8\u5173\u4f18\u52bf\u3002\u4e0e\u4e4b\u4e0d\u540c，CSGCR \u63d0\u51fa\u4e86\u4e09\u4e2a\u6a21\u5757\u2014\u2014\u8bcd\u8bed\u5b58\u5728\u6027\u9a8c\u8bc1\u3001\u6761\u4ef6\u53e5\u5b50\u751f\u6210\u4ee5\u53ca\u8de8\u6a21\u6001\u91cd\u6392\u5e8f\u2014\u2014\u4ee5\u5b66\u4e60\u66f4\u597d\u7684\u8bed\u6cd5\u7279\u5f81，\u4ece\u800c\u63d0\u5347\u624b\u8bed\u7ffb\u8bd1\u7684\u51c6\u786e\u6027\u548c\u6d41\u7545\u6027\u3002\u7136\u800c，\u7531\u4e8e\u624b\u8bed\u89c6\u9891\u4e2d\u7684\u52a8\u4f5c\u987a\u5e8f\u4e0e\u53e3\u8bed\u53e5\u5b50\u7684\u8bcd\u8bed\u987a\u5e8f\u5b58\u5728\u663e\u8457\u5dee\u5f02，\u5728\u6ca1\u6709 gloss \u6307\u5bfc\u7684\u60c5\u51b5\u4e0b\u8fdb\u884c\u6a21\u6001\u5bf9\u9f50\u5341\u5206\u56f0\u96be\u3002\u56e0\u6b64，\u65e0 gloss \u624b\u8bed\u7ffb\u8bd1\u65b9\u6cd5\u7684\u6027\u80fd\u901a\u5e38\u4f4e\u4e8e\u57fa\u4e8e gloss \u7684\u65b9\u6cd5\u3002\u672c\u6587\u91c7\u7528\u4e00\u79cd\u57fa\u4e8e VLP \u7684\u65b9\u6cd5\u6765\u83b7\u5f97\u66f4\u597d\u7684\u8de8\u6a21\u6001\u8868\u793a，\u4ece\u800c\u6709\u6548\u7f29\u5c0f\u65e0 gloss \u65b9\u6cd5\u4e0e\u57fa\u4e8e gloss \u65b9\u6cd5\u4e4b\u95f4\u7684\u6027\u80fd\u5dee\u8ddd\u3002<\/p>\n
3.\u65b9\u6cd5<\/h2>\n
\u672c\u6587\u8ba4\u4e3a，\u8bed\u8a00\u6307\u793a\u7684\u89c6\u89c9\u8868\u5f81\u540c\u65f6\u5177\u5907\u8bed\u8a00\u4fe1\u606f\u7684\u4f4e\u5197\u4f59\u6027\u4e0e\u9ad8\u5ea6\u62bd\u8c61\u6027，\u56e0\u6b64\u80fd\u591f\u6539\u5584\u624b\u8bed\u7ffb\u8bd1\u3002\u4e3a\u6b64，\u672c\u6587\u4e3a\u624b\u8bed\u7ffb\u8bd1\u5f15\u5165\u4e86\u4e00\u79cd\u65b0\u7684\u9884\u8bad\u7ec3\u8303\u5f0f，\u5c06\u63a9\u7801\u81ea\u76d1\u7763\u5b66\u4e60\u4e0e CLIP \u76f8\u7ed3\u5408，\u4ece\u800c\u4e3a\u4e0b\u6e38 GFSLT \u6a21\u578b\u8054\u5408\u9884\u8bad\u7ec3 Visual Encoder $\\\\psi_{VE}(\\\\cdot)$ \u548c Text Decoder $\\\\psi_{TD}(\\\\cdot)$，\u5177\u4f53\u89c1\u7b2c 3.1 \u8282\u3002\u968f\u540e，\u672c\u6587\u5c06\u9884\u8bad\u7ec3\u540e\u7684 Visual Encoder $\\\\psi_{VE}^{*}(\\\\cdot)$ \u548c Text Decoder $\\\\psi_{TD}^{*}(\\\\cdot)$ \u7684\u53c2\u6570\u8fc1\u79fb\u5230 GFSLT \u6a21\u578b $\\\\psi_{GFSLT}(\\\\cdot)$ \u4e2d，\u4ee5\u589e\u5f3a\u5176\u7ffb\u8bd1\u80fd\u529b，\u5177\u4f53\u89c1\u7b2c 3.2 \u8282\u3002Algorithm 1 \u8be6\u7ec6\u63cf\u8ff0\u4e86\u6574\u4e2a\u7b97\u6cd5\u6d41\u7a0b\u3002<\/p>\n
- \n
- \n
\u8fd9\u91cc\u7684\u201c\u8bed\u8a00\u6307\u793a\u201d\u53ef\u4ee5\u7406\u89e3\u4e3a：Visual Encoder \u5b66\u5230\u7684\u4e0d\u662f\u666e\u901a\u52a8\u4f5c\u8bc6\u522b\u7279\u5f81，\u800c\u662f\u5728\u6587\u672c\u8bed\u4e49\u76d1\u7763\u4e0b\u5f62\u6210\u7684\u89c6\u89c9\u7279\u5f81\u3002<\/p>\n<\/li>\n<\/ul>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/07\/20260729010842-6a69529af1db8.png\")
<\/p>\n- \n
- \n
Stage1 \u4e2d\u91cd\u590d\u6267\u884c\u7684\u5185\u5bb9\u662f：\u5bf9\u6570\u636e\u96c6 $\\\\mathcal D$ \u4e2d\u7684\u6bcf\u4e00\u5bf9 $V^{(i)},S^{(i)}$：\u4f7f\u7528\u89c6\u9891\u4e0e\u6587\u672c\u4e4b\u95f4\u7684\u5bf9\u6bd4\u635f\u5931\u6267\u884c\u68af\u5ea6\u4e0b\u964d\u4ece\u800c\u66f4\u65b0\u5bf9\u6bd4\u5206\u652f\u4e2d\u7684 Visual Encoder \u548c Text Encoder\u3002\u83b7\u5f97\u63a9\u7801\u540e\u7684\u53e5\u5b50\u5e76\u7528\u5b83\u6062\u590d\u635f\u5931\u6267\u884c\u68af\u5ea6\u4e0b\u964d\u4ece\u800c\u66f4\u65b0 Text Decoder\u3002<\/p>\n<\/li>\n
- \n
Stage2 \u4e2d\u91cd\u590d\u6267\u884c\u7684\u5185\u5bb9\u662f：\u5bf9\u6570\u636e\u96c6 $\\\\mathcal D$ \u4e2d\u7684\u6bcf\u4e00\u5bf9 $V^{(i)},S^{(i)}$：\u4f7f\u7528\u7ffb\u8bd1\u635f\u5931\u6267\u884c\u68af\u5ea6\u4e0b\u964d\u4ece\u800c\u66f4\u65b0\u6574\u4e2a $\\\\psi_{GFSLT}(\\\\cdot)$\u3002<\/p>\n<\/li>\n<\/ul>\n
3.1.\u89c6\u89c9\u2014\u8bed\u8a00\u9884\u8bad\u7ec3<\/h3>\n
\u4e3a\u4e86\u4ece\u624b\u8bed\u89c6\u9891\u4e2d\u5b66\u4e60\u8bed\u8a00\u6307\u793a\u7684\u89c6\u89c9\u8868\u5f81，\u9700\u8981\u8003\u8651\u4e24\u4e2a\u5173\u952e\u95ee\u9898：\u7b2c\u4e00，\u5982\u4f55\u8bbe\u8ba1\u4e00\u79cd\u80fd\u591f\u6709\u6548\u7f29\u5c0f\u89c6\u89c9\u8868\u5f81\u548c\u6587\u672c\u8868\u5f81\u4e4b\u95f4\u8bed\u4e49\u5dee\u8ddd\u7684\u9884\u6587\u672c\u4efb\u52a1？\u7b2c\u4e8c，\u5982\u4f55\u5728\u6709\u9650\u7684\u624b\u8bed\u7ffb\u8bd1\u6570\u636e\u96c6\u4e0a\u5b9e\u73b0\u8054\u5408\u9884\u8bad\u7ec3？<\/p>\n
\u4e3a\u4e86\u89e3\u51b3\u7b2c\u4e00\u4e2a\u95ee\u9898，\u6211\u4eec\u4ece\u96f6\u6837\u672c\u8fc1\u79fb\u5b66\u4e60\u9886\u57df\u7684 CLIP \u4e2d\u83b7\u5f97\u542f\u53d1\u3002CLIP \u5c06\u201c\u56fe\u50cf\u5230\u6587\u672c\u201d\u53d1\u5c55\u4e3a\u4e00\u79cd\u6807\u51c6\u5316\u7684\u8f93\u5165\u2014\u8f93\u51fa\u63a5\u53e3，\u4f7f\u89c6\u89c9\u6a21\u578b\u80fd\u591f\u901a\u8fc7\u81ea\u7136\u8bed\u8a00\u76d1\u7763\u83b7\u5f97\u8fc1\u79fb\u80fd\u529b\u3002CLIP \u5c55\u793a\u4e86\u4ece\u81ea\u7136\u8bed\u8a00\u4e2d\u5b66\u4e60\u76f8\u5bf9\u4e8e\u5176\u4ed6\u4efb\u52a1\u65e0\u5173\u9884\u8bad\u7ec3\u65b9\u5f0f\u7684\u4f18\u52bf，\u56e0\u6b64\u7279\u522b\u9002\u5408\u624b\u8bed\u7ffb\u8bd1\u4efb\u52a1\u3002\u6362\u53e5\u8bdd\u8bf4，\u7531\u4e8e\u624b\u8bed\u7ffb\u8bd1\u6570\u636e\u5929\u7136\u5177\u6709\u56fe\u50cf\u2014\u6587\u672c\u914d\u5bf9\u7ed3\u6784，\u5229\u7528\u8bed\u8a00\u76d1\u7763\u5b66\u4e60\u89c6\u89c9\u8868\u5f81，\u662f\u4e00\u79cd\u76f4\u63a5\u800c\u6709\u6548\u7684\u624b\u8bed\u7ffb\u8bd1\u9884\u6587\u672c\u4efb\u52a1\u3002\u57fa\u4e8e\u8fd9\u4e00\u8ba4\u8bc6，\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u89c6\u89c9\u2014\u8bed\u8a00\u9884\u8bad\u7ec3\u65b9\u6848，\u79f0\u4e3a VLP，\u5982 Figure 2(a) \u6240\u793a\u3002\u8be5\u65b9\u6848\u8054\u5408\u8bad\u7ec3 Visual Encoder $\\\\psi_{VE}(\\\\cdot)$ \u548c Text Encoder $\\\\psi_{TE}(\\\\cdot)$，\u4f7f\u6a21\u578b\u80fd\u591f\u4ece\u4e00\u4e2a\u6279\u6b21\u7684\u201c\u624b\u8bed\u89c6\u9891\u2014\u81ea\u7136\u8bed\u8a00\u53e5\u5b50\u201d\u8bad\u7ec3\u6837\u672c\u4e2d\u9884\u6d4b\u6b63\u786e\u7684\u914d\u5bf9\u5173\u7cfb\u3002\u5f62\u5f0f\u5316\u5730，\u9996\u5148\u5c06\u89c6\u9891\u2014\u6587\u672c\u5bf9\u5206\u522b\u8f93\u5165 $\\\\psi_{VE}(\\\\cdot)$ \u548c $\\\\psi_{TE}(\\\\cdot)$，\u5f97\u5230\u76f8\u5e94\u7684\u9ad8\u7ef4\u8bed\u4e49\u7279\u5f81：<\/p>\n
$$I_f=\\\\psi_{VE}(V),
\n\\\\qquad
\nV=(v_1,\\\\ldots,v_T);$$
\n$$I_y=\\\\psi_{TE}(S),
\n\\\\qquad
\nS=(s_1,\\\\ldots,s_U).$$ <\/p>\n\u5176\u4e2d，$V$ \u662f\u4e00\u6bb5\u5305\u542b $T$ \u5e27\u7684\u624b\u8bed\u89c6\u9891，$v_t$ \u662f\u89c6\u9891\u4e2d\u7684\u7b2c $t$ \u5e27，$S$ \u662f\u4e00\u53e5\u5305\u542b $U$ \u4e2a\u8bcd\u7684\u53e3\u8bed\u8bed\u8a00\u53e5\u5b50，$s_u$\u662f\u53e5\u5b50\u4e2d\u7684\u7b2c $u$ \u4e2a\u8bcd\u6216\u5b50\u8bcd Token，$I_f$\u662fVisual Encoder \u8f93\u51fa\u7684\u89c6\u89c9\u7279\u5f81\u5e8f\u5217\u3002<\/p>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/07\/20260729010843-6a69529b400b2.png\")
<\/p>\n\u89c6\u89c9\u7f16\u7801\u5668<\/h4>\n
Visual Encoder \u7531\u4e00\u4e2a Vision Embedding \u5c42\u4ee5\u53ca\u968f\u540e\u5806\u53e0\u7684\u591a\u5c42 Transformer Encoder \u6784\u6210，\u5176 Vision Embedding \u7ed3\u6784\u5982 Figure 3(b) \u6240\u793a\u3002\u89c6\u9891\u4e2d\u7684\u6bcf\u4e00\u5e27\u9996\u5148\u7ecf\u8fc7\u5171\u4eab\u6743\u91cd\u7684\u4e8c\u7ef4 CNN \u7f16\u7801；\u5f97\u5230\u7684\u89c6\u89c9\u7f16\u7801\u968f\u540e\u901a\u8fc7\u4e24\u4e2a\u65f6\u95f4\u6a21\u5757，\u8fd9\u4e24\u4e2a\u6a21\u5757\u91c7\u7528 Conv1D-BN-ReLU-MaxPooling \u7684\u7ec4\u5408\u6765\u6355\u6349\u77ed\u671f\u4f9d\u8d56\u5173\u7cfb\u3002\u6700\u540e，\u7279\u5f81\u88ab\u9001\u5165 Transformer Encoder，\u4ee5\u6355\u6349\u89c6\u9891\u4e2d\u7684\u957f\u671f\u4f9d\u8d56\u5173\u7cfb\u3002<\/p>\n
\u6587\u672c\u7f16\u7801\u5668<\/h4>\n
\u4e3a\u4e86\u6709\u6548\u5730\u7f16\u7801\u6587\u672c\u6570\u636e，\u4e00\u4e2a\u5f3a\u5927\u7684 Text Encoder \u81f3\u5173\u91cd\u8981\u3002\u56e0\u6b64，\u4f5c\u8005\u9009\u62e9\u4f7f\u7528 mBART \u4e2d\u7684 12 \u5c42\u7f16\u7801\u5668\u8fdb\u884c\u53c2\u6570\u521d\u59cb\u5316\u3002mBART \u662f\u4e00\u4e2a\u5df2\u7ecf\u5728 CC25 \u4e0a\u5b8c\u6210\u9884\u8bad\u7ec3\u7684\u795e\u7ecf\u673a\u5668\u7ffb\u8bd1\u6a21\u578b；CC25 \u662f\u4e00\u4e2a\u8986\u76d6 25 \u79cd\u8bed\u8a00\u7684\u591a\u8bed\u8a00\u8bed\u6599\u5e93\u3002<\/p>\n
\u968f\u540e，\u5c06\u63d0\u53d6\u5230\u7684\u89c6\u89c9\u7279\u5f81 $I_f$ \u548c\u6587\u672c\u7279\u5f81 $I_y$ \u5206\u522b\u8f93\u5165\u5bf9\u5e94\u7684\u6620\u5c04\u5934，\u901a\u8fc7\u7ebf\u6027\u6295\u5f71\u8fdb\u5165\u5171\u4eab\u7684\u591a\u6a21\u6001\u8bed\u4e49\u7a7a\u95f4，\u4ee5\u8ba1\u7b97\u76f8\u4f3c\u5ea6\u3002\u8fd9\u91cc，\u4e24\u4e2a\u6620\u5c04\u5934\u90fd\u53ea\u7531\u4e00\u4e2a\u7b80\u5355\u7684 Linear \u5c42\u7ec4\u6210\u3002\u5f62\u5f0f\u5316\u5730，\u8be5\u8fc7\u7a0b\u8868\u793a\u4e3a：<\/p>\n
$$\\\\widetilde I_{f,c}
\n=
\n\\\\operatorname{Linear}(I_{f,c}),
\n\\\\qquad
\n\\\\widetilde I_{y,k}
\n=
\n\\\\operatorname{Linear}(I_{y,k}).$$ <\/p>\n\u5176\u4e2d，$I_{f,c}$ \u8868\u793a Visual Encoder \u6700\u540e\u4e00\u5c42\u5728 [CLS] Token \u4f4d\u7f6e\u7684\u6fc0\u6d3b\u503c；$I_{y,k}$ \u8868\u793a Text Encoder \u6700\u540e\u4e00\u5c42\u5728 <EOS> Token \u4f4d\u7f6e\u7684\u6fc0\u6d3b\u503c\u3002\u968f\u540e，\u4e0e CLIP \u7c7b\u4f3c，\u5bf9 $\\\\widetilde I_{f,c}$ \u548c $\\\\widetilde I_{y,k}$ \u8fdb\u884c\u5c42\u5f52\u4e00\u5316\u548c\u6210\u5bf9\u7f29\u653e，\u5e76\u5229\u7528\u5bf9\u79f0\u4ea4\u53c9\u71b5\u635f\u5931\u8ba1\u7b97\u635f\u5931\u503c：<\/p>\n
$$\\\\mathcal L_s
\n=
\n-\\\\frac{1}{2}
\n\\\\left(
\n\\\\sum V\\\\log\\\\left(\\\\widetilde I_{f,c}\\\\right)
\n+
\n\\\\sum S\\\\log\\\\left(\\\\widetilde I_{y,k}\\\\right)
\n\\\\right).$$ <\/p>\n- \n
- \n
[CLS] \u662f\u4e3a\u4e86\u5f62\u6210\u6574\u4e2a\u89c6\u9891\u5e8f\u5217\u7684\u5168\u5c40\u8868\u793a\u800c\u6dfb\u52a0\u7684\u7279\u6b8a Token\u3002\u6587\u672c\u5e8f\u5217\u5219\u7531 <BOS> \u548c <EOS> \u5305\u56f4，\u4f5c\u8005\u53d6 <EOS> \u4f4d\u7f6e\u7684\u6fc0\u6d3b\u4f5c\u4e3a\u6574\u53e5\u8bdd\u7684\u5168\u5c40\u8868\u793a\u3002<\/p>\n<\/li>\n<\/ul>\n
\u4e3a\u4e86\u89e3\u51b3\u7b2c\u4e8c\u4e2a\u95ee\u9898，\u4f5c\u8005\u5728\u7b97\u6cd5\u5c42\u9762\u548c\u6570\u636e\u5c42\u9762\u540c\u65f6\u91c7\u53d6\u63aa\u65bd\u3002\u5728\u7b97\u6cd5\u5c42\u9762，\u5982 Figure 2(a) \u6240\u793a，\u6211\u4eec\u989d\u5916\u5f15\u5165\u4e86\u4e00\u6761\u8f85\u52a9\u76d1\u7763\u6d41\u2014\u2014\u63a9\u7801\u81ea\u76d1\u7763\u5b66\u4e60，\u5373\u56fe\u4e2d\u7684\u865a\u7ebf\u5206\u652f，\u4ece\u800c\u5b9e\u73b0\u8054\u5408\u9884\u8bad\u7ec3\u3002\u5728\u8be5\u5206\u652f\u4e2d，Text Decoder $\\\\psi_{TD}(\\\\cdot)$ \u63a5\u6536\u7531\u5171\u4eab\u6743\u91cd\u7684 Text Encoder\u4ece\u63a9\u7801\u53e5\u5b50\u4e2d\u63d0\u53d6\u7684\u8bed\u8a00\u7279\u5f81，\u5e76\u9884\u6d4b\u63a9\u7801\u533a\u57df\u4e2d\u7684\u8bcd\u3002<\/p>\n
- \n
- \u8fd9\u91cc\u7684\u201c\u5171\u4eab\u6743\u91cd\u201d\u610f\u5473\u7740：\n
- \n
- \u5bf9\u6bd4\u5b66\u4e60\u5206\u652f\u4e2d\u7684\u6b63\u5e38\u53e5\u5b50\u8fdb\u5165 $\\\\psi_{TE}$；<\/li>\n
- \u63a9\u7801\u6062\u590d\u5206\u652f\u4e2d\u7684\u63a9\u7801\u53e5\u5b50\u4e5f\u8fdb\u5165\u540c\u4e00\u4e2a $\\\\psi_{TE}$；<\/li>\n
- \u4e24\u6761\u5206\u652f\u4e0d\u662f\u5404\u81ea\u4f7f\u7528\u4e00\u5957\u72ec\u7acb\u7684 Text Encoder\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n
\u6587\u672c\u89e3\u7801\u5668<\/h4>\n
Text Decoder \u5b9e\u9645\u4e0a\u662f\u4e00\u4e2a\u5e26\u6709\u56e0\u679c\u63a9\u7801\u7684\u6807\u51c6 Transformer Decoder\u3002\u5176\u4f18\u5316\u76ee\u6807\u53ef\u4ee5\u8868\u793a\u4e3a：<\/p>\n
$$\\\\min_{\\\\Theta}
\n\\\\frac{1}{N}
\n\\\\sum_{i=1}^{N}
\n\\\\mathcal L_c
\n\\\\left(
\n\\\\psi_{TD}
\n\\\\left(
\n\\\\psi_{TE}^{*}
\n\\\\left(
\n\\\\widetilde S^{(i)}
\n\\\\right)
\n\\\\right),
\nS^{(i)}
\n\\\\right).$$ <\/p>\n\u5176\u4e2d，$\\\\widetilde S$ \u8868\u793a\u63a9\u7801\u540e\u7684\u53e5\u5b50，$N$ \u8868\u793a\u8bad\u7ec3\u6837\u672c\u6570，$\\\\mathcal L_c$ \u8868\u793a\u635f\u5931\u51fd\u6570\u3002\u53ef\u4ee5\u62c6\u89e3\u4e3a：<\/p>\n
$$\\\\widetilde S^{(i)}
\n\\\\rightarrow
\n\\\\psi_{TE}^{*}
\n\\\\left(
\n\\\\widetilde S^{(i)}
\n\\\\right)
\n\\\\rightarrow
\n\\\\psi_{TD}
\n\\\\left(
\n\\\\psi_{TE}^{*}
\n\\\\left(
\n\\\\widetilde S^{(i)}
\n\\\\right)
\n\\\\right)
\n\\\\rightarrow
\nS^{(i)}.$$ <\/p>\n\u5404\u7b26\u53f7\u542b\u4e49\u5982\u4e0b：<\/p>\n
- \n
- $S^{(i)}$：\u7b2c $i$ \u4e2a\u5b8c\u6574\u539f\u59cb\u53e5\u5b50\u3002<\/li>\n
- $\\\\widetilde S^{(i)}$：\u5bf9 $S^{(i)}$ \u8fdb\u884c\u63a9\u7801\u540e\u5f97\u5230\u7684\u53e5\u5b50\u3002<\/li>\n
- $\\\\psi_{TE}^{*}$：\u4e3a Decoder \u63d0\u4f9b\u4e0a\u4e0b\u6587\u7279\u5f81\u7684 Text Encoder\u3002<\/li>\n
- $\\\\psi_{TD}$：\u6839\u636e\u7f16\u7801\u7279\u5f81\u548c\u76ee\u6807\u524d\u6587\u6062\u590d\u53e5\u5b50\u7684 Text Decoder\u3002<\/li>\n
- $\\\\Theta$：\u8be5\u76ee\u6807\u4e0b\u88ab\u4f18\u5316\u7684\u53c2\u6570\u96c6\u5408，\u4e3b\u8981\u5bf9\u5e94 Text Decoder \u53c2\u6570\u3002<\/li>\n
- $\\\\mathcal L_c$：\u63a9\u7801\u53e5\u5b50\u6062\u590d\u635f\u5931\u3002<\/li>\n<\/ul>\n
\u56e0\u679c\u63a9\u7801\u4fdd\u8bc1 Decoder \u5728\u9884\u6d4b\u7b2c $u$ \u4e2a\u8bcd\u65f6，\u53ea\u80fd\u770b\u5230：<\/p>\n
$$s_1,\\\\ldots,s_{u-1}$$ <\/p>\n
\u800c\u4e0d\u80fd\u770b\u5230\u7b2c $u$ \u4e2a\u8bcd\u4e4b\u540e\u7684\u5185\u5bb9\u3002<\/p>\n
\u91c7\u7528\u8054\u5408\u9884\u8bad\u7ec3\u7684\u539f\u56e0\u5728\u4e8e，\u4e0d\u540c\u7684\u9884\u8bad\u7ec3\u8303\u5f0f\u80fd\u591f\u6355\u6349\u6570\u636e\u7684\u4e0d\u540c\u65b9\u9762，\u800c\u5c06\u5b83\u4eec\u7ed3\u5408\u8d77\u6765\u53ef\u4ee5\u83b7\u5f97\u66f4\u5168\u9762\u7684\u6570\u636e\u8868\u5f81\u3002<\/p>\n
\u5728\u6570\u636e\u5c42\u9762，\u4f5c\u8005\u4e3a\u624b\u8bed\u7ffb\u8bd1\u7684\u8f93\u5165\u89c6\u9891\u5f15\u5165\u4e86\u5f3a\u6570\u636e\u589e\u5f3a，\u8be5\u589e\u5f3a\u7531 VIDAUG \u5e93\u5b9e\u73b0，\u5305\u62ec\u51e0\u4f55\u53d8\u6362\u3001\u989c\u8272\u7a7a\u95f4\u53d8\u6362\u548c\u65f6\u95f4\u53d8\u6362\u3002\u5728\u8bad\u7ec3\u671f\u95f4，\u6211\u4eec\u968f\u673a\u7ec4\u5408\u8fd9\u4e09\u7c7b\u589e\u5f3a\u65b9\u6cd5，\u4ee5\u6269\u5927\u6570\u636e\u7a7a\u95f4\u3002<\/p>\n
\u4e8b\u5b9e\u4e0a，\u8fd9\u79cd\u9884\u8bad\u7ec3\u8303\u5f0f\u4f7f Visual Encoder \u80fd\u591f\u83b7\u5f97\u4e0e Text Encoder \u76f8\u4f3c\u7684\u3001\u5f3a\u5927\u7684\u8bed\u8a00\u8868\u5f81\u80fd\u529b，\u4ece\u800c\u751f\u6210\u66f4\u52a0\u9c81\u68d2\u3001\u66f4\u5177\u4ee3\u8868\u6027\u7684\u89c6\u89c9\u7279\u5f81\u3002\u8fd9\u5c31\u662f\u4e3a\u4ec0\u4e48\u4e3a\u624b\u8bed\u7ffb\u8bd1\u5b66\u4e60\u8bed\u8a00\u6307\u793a\u7684\u89c6\u89c9\u7279\u5f81\u662f\u53ef\u884c\u7684\u3002\u56e0\u6b64，\u5f53 Visual Encoder \u548c Text Decoder \u5efa\u7acb\u8d77\u8fd9\u6837\u7684\u5efa\u6a21\u80fd\u529b\u4e4b\u540e，\u6211\u4eec\u4fbf\u5728\u7b2c\u4e8c\u9636\u6bb5\u4f7f\u7528\u5b83\u4eec\u6267\u884c\u624b\u8bed\u7ffb\u8bd1\u4efb\u52a1\u3002<\/p>\n
3.2.\u65e0 Gloss \u624b\u8bed\u7ffb\u8bd1<\/h3>\n
\u5728\u672c\u8282\u4e2d，\u4f5c\u8005\u63d0\u51fa\u65e0 Gloss \u624b\u8bed\u7ffb\u8bd1\u7f51\u7edc GFSLT\u3002\u8be5\u7f51\u7edc\u80fd\u591f\u5728\u4e0d\u501f\u52a9\u4efb\u4f55 Gloss \u6807\u6ce8\u7684\u60c5\u51b5\u4e0b，\u6839\u636e\u7ed9\u5b9a\u7684\u624b\u8bed\u89c6\u9891 $V$ \u751f\u6210\u5bf9\u5e94\u53e5\u5b50 $S$\u3002\u4e3a\u4e86\u5b9e\u73b0\u8fd9\u4e00\u76ee\u6807，\u5982 Figure 3(a) \u6240\u793a，\u6211\u4eec\u91c7\u7528 Transformer \u4f5c\u4e3a\u6a21\u578b\u7684\u4e3b\u8981\u6846\u67b6，\u56e0\u4e3a\u5b83\u5df2\u7ecf\u5728\u795e\u7ecf\u673a\u5668\u7ffb\u8bd1\u4efb\u52a1\u4e2d\u8868\u73b0\u51fa\u4f18\u5f02\u6027\u80fd\u3002\u9996\u5148，\u5c06\u624b\u8bed\u89c6\u9891\u8f93\u5165\u5728 VLP \u9636\u6bb5\u9884\u8bad\u7ec3\u5f97\u5230\u7684 Visual Encoder $\\\\psi_{VE}^{*}$，\u4ece\u800c\u5f97\u5230\u9690\u85cf\u8bed\u4e49\u5411\u91cf：<\/p>\n
$$h_{1:M}
\n=
\n\\\\psi_{VE}^{*}
\n\\\\left(
\nv_{1:T}
\n\\\\right).$$ <\/p>\n\u5176\u4e2d：<\/p>\n
$$M=\\\\frac{T}{4}.$$ <\/p>\n
\u4e0e\u6b64\u540c\u65f6，\u5728 VLP \u9636\u6bb5\u9884\u8bad\u7ec3\u5f97\u5230\u7684 Text Decoder $\\\\psi_{TD}^{*}$，\u5c06\u5bf9\u5e94\u53e5\u5b50：<\/p>\n
$$S=(s_1,\\\\ldots,s_U)$$ <\/p>\n
\u548c Encoder \u7684\u6700\u540e\u4e00\u5c42\u9690\u85cf\u72b6\u6001\u4f5c\u4e3a\u8f93\u5165，\u6bcf\u6b21\u751f\u6210\u4e00\u4e2a\u8bcd：<\/p>\n
$$z_u
\n=
\n\\\\psi_{TD}^{*}
\n\\\\left(
\ns_{1:u-1},
\nh_{1:M}
\n\\\\right).$$ <\/p>\n\u53e5\u5b50\u7684\u7b2c\u4e00\u4e2a\u8bcd\u88ab\u4eba\u4e3a\u8bbe\u7f6e\u4e3a\u7279\u6b8a\u6807\u8bb0 <BOS>，Transformer Decoder \u4f1a\u4e00\u76f4\u751f\u6210，\u76f4\u5230\u8f93\u51fa\u7279\u6b8a\u6807\u8bb0 <EOS>\u3002\u6700\u540e，\u7ecf\u8fc7 Linear \u5c42\u548c Softmax \u5c42\u540e，\u8ba1\u7b97\u6761\u4ef6\u6982\u7387 $p(S\\\\mid V)$，\u5e76\u901a\u8fc7\u6700\u5c0f\u5316\u89c6\u9891\u5230\u53e5\u5b50\u7684\u4ea4\u53c9\u71b5\u635f\u5931，\u4f18\u5316\u6574\u4e2a\u7f51\u7edc：<\/p>\n
$$p(S\\\\mid V)
\n=
\n\\\\prod_{u=1}^{U}
\np(s_u\\\\mid o_u),
\n\\\\qquad
\no_u
\n=
\n\\\\operatorname{softmax}
\n\\\\left(
\nWz_u+b
\n\\\\right),$$
\n$$\\\\mathcal L_g
\n=
\n-\\\\log p(S\\\\mid V).$$ <\/p>\n4.\u5b9e\u9a8c<\/h2>\n
4.1.\u6570\u636e\u96c6\u548c\u8bc4\u4f30\u65b9\u6cd5<\/h3>\n
\u6570\u636e\u96c6\u3002\u6211\u4eec\u5728\u4e24\u4e2a\u88ab\u5e7f\u6cdb\u4f7f\u7528\u7684\u624b\u8bed\u7ffb\u8bd1\u6570\u636e\u96c6\u4e0a\u8bc4\u4f30\u6240\u63d0\u51fa\u7684\u65b9\u6cd5：RWTH-PHOENIX-Weather 2014T，\u7b80\u79f0 PHOENIX-2014T，\u4ee5\u53ca CSL-Daily\u3002PHOENIX-2014T \u5305\u542b\u6765\u81ea\u5929\u6c14\u9884\u62a5\u8282\u76ee\u7684 8257 \u4e2a\u5fb7\u56fd\u624b\u8bed（DGS）\u89c6\u9891\u53ca\u5176\u5bf9\u5e94\u7684\u5fb7\u8bed\u7ffb\u8bd1\u3002\u8be5\u6570\u636e\u96c6\u88ab\u5212\u5206\u4e3a\u8bad\u7ec3\u96c6\u3001\u9a8c\u8bc1\u96c6\u548c\u6d4b\u8bd5\u96c6，\u6837\u672c\u6570\u91cf\u5206\u522b\u4e3a 7096\u3001519 \u548c 642\u3002\u5176\u5fb7\u8bed\u7ffb\u8bd1\u6587\u672c\u7684\u8bcd\u8868\u5927\u5c0f\u4e3a 2887\u3002CSL-Daily \u4e3b\u8981\u5173\u6ce8\u4e2d\u56fd\u624b\u8bed\u4e2d\u7684\u65e5\u5e38\u751f\u6d3b\u4e3b\u9898，\u5171\u5305\u542b 20654 \u4e2a\u4e2d\u56fd\u624b\u8bed\u89c6\u9891\u53ca\u5176\u5bf9\u5e94\u7684\u4e2d\u6587\u7ffb\u8bd1\u3002\u8be5\u6570\u636e\u96c6\u88ab\u5212\u5206\u4e3a\u8bad\u7ec3\u96c6\u3001\u9a8c\u8bc1\u96c6\u548c\u6d4b\u8bd5\u96c6，\u6837\u672c\u6570\u91cf\u5206\u522b\u4e3a 18401\u30011077 \u548c 1176，\u4e2d\u6587\u7ffb\u8bd1\u6587\u672c\u7684\u8bcd\u8868\u5927\u5c0f\u4e3a 2343\u3002<\/p>\n
\u8bc4\u4f30\u6307\u6807\u3002\u9075\u5faa\u4ee5\u5f80\u5de5\u4f5c，\u6211\u4eec\u91c7\u7528 BLEU \u548c ROUGE \u8bc4\u4f30\u624b\u8bed\u7ffb\u8bd1\u6027\u80fd\u3002BLEU \u548c ROUGE-L \u7684\u6570\u503c\u8d8a\u9ad8，\u8868\u793a\u7ffb\u8bd1\u6027\u80fd\u8d8a\u597d\u3002<\/p>\n
- \n
- BLEU-$n$ \u4e3b\u8981\u8003\u5bdf\u6a21\u578b\u8bd1\u6587\u4e0e\u53c2\u8003\u8bd1\u6587\u4e4b\u95f4 $n$-gram \u7684\u5339\u914d\u60c5\u51b5\u3002\u4f8b\u5982：\n
- \n
- BLEU-1 \u66f4\u4fa7\u91cd\u5355\u8bcd\u662f\u5426\u6b63\u786e；<\/li>\n
- BLEU-2 \u8003\u5bdf\u8fde\u7eed\u4e24\u4e2a\u8bcd\u662f\u5426\u5339\u914d；<\/li>\n
- BLEU-4 \u8003\u5bdf\u8fde\u7eed\u56db\u4e2a\u8bcd\u7684\u5339\u914d，\u66f4\u80fd\u53cd\u6620\u77ed\u8bed\u7ed3\u6784\u548c\u7ffb\u8bd1\u6d41\u7545\u6027\u3002\u56e0\u6b64，BLEU-4 \u901a\u5e38\u6bd4 BLEU-1 \u66f4\u4e25\u683c，\u4e5f\u662f\u672c\u6587\u540e\u7eed\u6d88\u878d\u5b9e\u9a8c\u4e3b\u8981\u5173\u6ce8\u7684\u6307\u6807\u3002<\/li>\n<\/ul>\n<\/li>\n
- ROUGE-L \u57fa\u4e8e\u6a21\u578b\u8bd1\u6587\u4e0e\u53c2\u8003\u8bd1\u6587\u4e4b\u95f4\u7684\u6700\u957f\u516c\u5171\u5b50\u5e8f\u5217，\u66f4\u91cd\u89c6\u8bcd\u8bed\u987a\u5e8f\u4ee5\u53ca\u53e5\u5b50\u4e2d\u6709\u591a\u5c11\u53c2\u8003\u5185\u5bb9\u88ab\u8986\u76d6\u3002<\/li>\n<\/ul>\n
4.2.\u5b9e\u65bd\u7ec6\u8282<\/h3>\n
GFSLT \u6a21\u578b\u3002\u4f5c\u8005\u4f7f\u7528\u5728 ImageNet \u4e0a\u9884\u8bad\u7ec3\u7684 ResNet18 \u4f5c\u4e3a\u4e8c\u7ef4 CNN\u3002\u5bf9\u4e8e\u65f6\u95f4\u6a21\u5757，\u6211\u4eec\u6cbf\u7528\u73b0\u6709\u5de5\u4f5c\u7684\u914d\u7f6e：Conv1D \u5c42\u7684\u6b65\u957f\u548c\u5377\u79ef\u6838\u5927\u5c0f\u5206\u522b\u8bbe\u7f6e\u4e3a 1 \u548c 5，MaxPooling \u5c42\u7684\u6b65\u957f\u548c\u6c60\u5316\u6838\u5927\u5c0f\u5206\u522b\u8bbe\u7f6e\u4e3a 2 \u548c 2\u3002Transformer Encoder \u548c Transformer Decoder \u5747\u5305\u542b 3 \u5c42，\u9690\u85cf\u5c42\u7ef4\u5ea6\u4e3a 1024，\u524d\u9988\u7f51\u7edc\u7ef4\u5ea6\u4e3a 4096\u3002\u6bcf\u4e00\u5c42\u90fd\u5305\u542b 8 \u4e2a\u6ce8\u610f\u529b\u5934，\u5e76\u5c06 Dropout \u8bbe\u7f6e\u4e3a 0.1，\u4ee5\u907f\u514d\u8fc7\u62df\u5408\u3002<\/p>\n
\u89c6\u89c9\u2014\u8bed\u8a00\u9884\u8bad\u7ec3\u3002\u4f5c\u8005\u5206\u522b\u5728\u4e24\u4e2a\u624b\u8bed\u6570\u636e\u96c6\u5404\u81ea\u7684\u8bad\u7ec3\u96c6\u4e0a\u6267\u884c\u9884\u8bad\u7ec3\u4efb\u52a1\u3002Mini-batch \u5927\u5c0f\u8bbe\u7f6e\u4e3a 16；\u4f7f\u7528 AMP \u6280\u672f\u6269\u5927\u6279\u6b21\u5927\u5c0f\u3002\u8f93\u5165\u5e8f\u5217\u9996\u5148\u88ab\u7f29\u653e\u81f3 $256\\\\times256$，\u968f\u540e\u5728\u8bad\u7ec3\u548c\u63a8\u7406\u9636\u6bb5\u5206\u522b\u8fdb\u884c\u968f\u673a\u88c1\u526a\u548c\u4e2d\u5fc3\u88c1\u526a，\u5f97\u5230 $224\\\\times224$ \u7684\u8f93\u5165\u3002\u91c7\u7528\u52a8\u91cf\u4e3a 0.9 \u7684 SGD \u4f5c\u4e3a\u4f18\u5316\u5668，\u5e76\u4f7f\u7528\u4f59\u5f26\u5b66\u4e60\u7387\u8c03\u5ea6\u7b56\u7565，\u4f7f\u5b66\u4e60\u7387\u4ece\u6700\u5927\u503c 0.01 \u8870\u51cf\u81f3\u6700\u5c0f\u503c $1\\\\times10^{-5}$\u3002\u9884\u8bad\u7ec3\u603b\u5171\u8fdb\u884c 80 \u4e2a Epoch\u3002<\/p>\n
- \n
- \n
AMP \u662f Automatic Mixed Precision，\u81ea\u52a8\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\u3002\u5b83\u901a\u5e38\u7ed3\u5408\u534a\u7cbe\u5ea6\u548c\u5355\u7cbe\u5ea6\u8ba1\u7b97，\u4ee5\u51cf\u5c11\u663e\u5b58\u5360\u7528\u3001\u63d0\u9ad8\u8bad\u7ec3\u901f\u5ea6，\u56e0\u6b64\u4f5c\u8005\u80fd\u591f\u91c7\u7528\u66f4\u5927\u7684\u6709\u6548\u6279\u6b21\u3002<\/p>\n<\/li>\n<\/ul>\n
\u624b\u8bed\u7ffb\u8bd1\u8bad\u7ec3\u4e0e\u63a8\u7406\u3002GFSLT \u7f51\u7edc\u91c7\u7528\u5e26\u6709 0.2 \u6807\u7b7e\u5e73\u6ed1\u7684\u4ea4\u53c9\u71b5\u635f\u5931\u8fdb\u884c\u7aef\u5230\u7aef\u8bad\u7ec3，Mini-batch \u5927\u5c0f\u8bbe\u7f6e\u4e3a 8\u3002\u4f7f\u7528\u52a8\u91cf\u4e3a 0.9 \u7684 SGD \u4f18\u5316\u5668，\u5c06\u521d\u59cb\u5b66\u4e60\u7387\u8bbe\u7f6e\u4e3a 0.01，\u5e76\u91c7\u7528\u4f59\u5f26\u9000\u706b\u5b66\u4e60\u7387\u8c03\u5ea6\u5668\u3002\u6574\u4e2a\u7f51\u7edc\u8bad\u7ec3 200 \u4e2a Epoch\u3002\u5728\u63a8\u7406\u9636\u6bb5，\u6a21\u578b\u4f7f\u7528\u675f\u641c\u7d22\u7b56\u7565\u8fdb\u884c\u89e3\u7801，\u957f\u5ea6\u60e9\u7f5a\u8bbe\u7f6e\u4e3a 1，Beam Size \u8bbe\u7f6e\u4e3a 5\u3002<\/p>\n
- \n
- \u6807\u7b7e\u5e73\u6ed1：\u666e\u901a\u4ea4\u53c9\u71b5\u901a\u5e38\u628a\u6b63\u786e\u8bcd\u7684\u76ee\u6807\u6982\u7387\u8bbe\u7f6e\u4e3a 1，\u5176\u4ed6\u8bcd\u8bbe\u7f6e\u4e3a 0\u3002\u6807\u7b7e\u5e73\u6ed1\u4e3a 0.2 \u65f6，\u4f1a\u628a\u4e00\u90e8\u5206\u6982\u7387\u5206\u914d\u7ed9\u8bcd\u8868\u4e2d\u7684\u5176\u4ed6\u8bcd，\u907f\u514d\u6a21\u578b\u5bf9\u8bad\u7ec3\u6807\u7b7e\u8fc7\u5ea6\u81ea\u4fe1\u3002<\/li>\n
- \u675f\u641c\u7d22：\u4fdd\u7559\u6982\u7387\u6700\u9ad8\u7684\u5019\u9009\u5e8f\u5217\u3002<\/li>\n
- \u957f\u5ea6\u60e9\u7f5a\u8bbe\u7f6e\u4e3a 1，\u7528\u4e8e\u5728\u53e5\u5b50\u6982\u7387\u548c\u53e5\u5b50\u957f\u5ea6\u4e4b\u95f4\u8fdb\u884c\u5e73\u8861，\u907f\u514d\u6a21\u578b\u4ec5\u4ec5\u56e0\u4e3a\u77ed\u53e5\u9700\u8981\u76f8\u4e58\u7684\u8bcd\u6982\u7387\u8f83\u5c11，\u5c31\u8fc7\u5ea6\u504f\u5411\u751f\u6210\u8f83\u77ed\u53e5\u5b50\u3002<\/li>\n<\/ul>\n
\u89c6\u89c9\u7f16\u7801\u5668<\/h4>\n
\u624b\u8bed\u89c6\u9891\u5e27
\n \u00a0 \u00a0\u2193
\nResNet18
\n \u00a0 \u00a0\u2193
\nConv1D(kernel=5, stride=1)
\n \u00a0 \u00a0\u2193
\nBN \u2192 ReLU
\n \u00a0 \u00a0\u2193
\nMaxPool1D(kernel=2, stride=2)
\n \u00a0 \u00a0\u2193
\nConv1D(kernel=5, stride=1)
\n \u00a0 \u00a0\u2193
\nBN \u2192 ReLU
\n \u00a0 \u00a0\u2193
\nMaxPool1D(kernel=2, stride=2)
\n \u00a0 \u00a0\u2193
\nLinear
\n \u00a0 \u00a0\u2193
\n\u4f4d\u7f6e\u7f16\u7801 PE
\n \u00a0 \u00a0\u2193
\nTransformer Encoder <\/p>\n\u6587\u672c\u89e3\u7801\u5668<\/h4>\n
\u4e4b\u524d\u7684\u76ee\u6807\u8bcd
\n \u00a0 \u00a0\u2193
\nWord Embedding + \u4f4d\u7f6e\u7f16\u7801
\n \u00a0 \u00a0\u2193
\nTransformer Decoder \u2b05 \u89c6\u89c9\u7f16\u7801\u5668\u4e2d Transformer Encoder \u7684\u8f93\u51fa
\n \u00a0 \u00a0\u2193
\nLinear
\n \u00a0 \u00a0\u2193
\nSoftmax
\n \u00a0 \u00a0\u2193
\n\u4e0b\u4e00\u4e2a\u8bcd <\/p>\n\u9884\u8bad\u7ec3\u9636\u6bb5<\/h4>\n
\u9884\u8bad\u7ec3\u65f6\u8fd8\u4f7f\u7528\u4e86 mBART \u7684 12 \u5c42 Text Encoder\u3002<\/p>\n
\u5b8c\u6574\u53e5\u5b50\u6216\u63a9\u7801\u53e5\u5b50
\n \u00a0 \u00a0\u2193
\nmBART 12 \u5c42 Text Encoder <\/p>\n\u5176\u4e2d：<\/p>\n
- \n
- mBART \u7684 12 \u5c42 Text Encoder \u4f7f\u7528\u5916\u90e8\u9884\u8bad\u7ec3\u53c2\u6570；<\/li>\n
- \u89c6\u89c9\u7f16\u7801\u5668\u4e2d\u7684 Transformer Encoder \u662f\u6807\u51c6 Transformer Encoder；<\/li>\n
- \u6587\u672c\u89e3\u7801\u5668\u4e2d\u7684 Transformer Decoder \u662f\u6807\u51c6 Transformer Decoder；<\/li>\n
- mBART Text Encoder \u4ec5\u5728\u9884\u8bad\u7ec3\u9636\u6bb5\u4f7f\u7528，\u4e0d\u8fdb\u5165\u7b2c\u4e8c\u9636\u6bb5\u7684\u6700\u7ec8\u7ffb\u8bd1\u6a21\u578b\u3002<\/li>\n<\/ul>\n
4.3.\u548cSOTA\u7684\u5bf9\u6bd4<\/h3>\n
Table 1 （\u51e4\u51f0\u6570\u636e\u96c6）\u5c55\u793a\u4e86\u672c\u6587\u65b9\u6cd5\u4e0e\u5f53\u524d\u6700\u5148\u8fdb\u7684\u57fa\u4e8e Gloss \u65b9\u6cd5\u548c\u65e0 Gloss \u65b9\u6cd5\u5728\u624b\u8bed\u7ffb\u8bd1\u4efb\u52a1\u4e0a\u7684\u6bd4\u8f83\u7ed3\u679c\u3002\u4e0e CSGCR \u7b49\u5176\u4ed6\u65e0 Gloss \u65b9\u6cd5\u76f8\u6bd4，\u672c\u6587\u65b9\u6cd5\u53d6\u5f97\u4e86\u663e\u8457\u7684\u6027\u80fd\u63d0\u5347\u3002\u5177\u4f53\u800c\u8a00，\u672c\u6587\u65b9\u6cd5\u5728\u9a8c\u8bc1\u96c6\u548c\u6d4b\u8bd5\u96c6\u4e0a\u7684 BLEU-4 \u5206\u6570\u5206\u522b\u63d0\u5347\u4e86\u7ea6 7.0 \u5206\u548c 5.7 \u5206，ROUGE \u5206\u6570\u5206\u522b\u63d0\u5347\u4e86\u7ea6 4.8 \u5206\u548c 2.6 \u5206\u3002\u6b64\u5916，\u4e0e\u5927\u591a\u6570\u57fa\u4e8e Gloss \u7684\u65b9\u6cd5\u76f8\u6bd4，\u672c\u6587\u65b9\u6cd5\u4e5f\u5177\u6709\u5f88\u5f3a\u7684\u7ade\u4e89\u529b\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f，\u672c\u6587\u65b9\u6cd5\u53d6\u5f97\u4e86\u4e0e SLRT \u548c STMC-T \u63a5\u8fd1\u7684\u6027\u80fd：\u5728\u9a8c\u8bc1\u96c6 BLEU-4 \u4e0a，SLRT\u3001STMC-T \u548c\u672c\u6587\u65b9\u6cd5\u7684\u7ed3\u679c\u5206\u522b\u4e3a 22.38\u300124.09 \u548c 22.12\u3002\u8fd9\u4e00\u7ed3\u679c\u663e\u793a\u4e86\u65e0 Gloss \u65b9\u6cd5\u7684\u6f5c\u529b\u3002<\/p>\n
- \n
- \n
\u5355\u72ec\u4f7f\u7528\u672c\u6587\u8bbe\u8ba1\u7684 GFSLT \u67b6\u6784，\u5373\u4e0d\u8fdb\u884c VLP，\u5df2\u7ecf\u5f97\u5230\u6d4b\u8bd5\u96c6 19.66 BLEU-4，\u660e\u663e\u8d85\u8fc7\u6b64\u524d\u65e0 Gloss \u65b9\u6cd5 GASLT \u7684 15.74\u3002<\/p>\n<\/li>\n<\/ul>\n
Table 2 （CSL-Daily\u6570\u636e\u96c6）\u5c06\u672c\u6587\u65b9\u6cd5\u4e0e CSL-Daily \u6570\u636e\u96c6\u4e0a\u7684\u6700\u5148\u8fdb\u65b9\u6cd5\u8fdb\u884c\u4e86\u6bd4\u8f83\u3002CSL-Daily \u662f\u4e8e 2021 \u5e74\u53d1\u5e03\u7684\u5927\u89c4\u6a21\u4e2d\u56fd\u624b\u8bed\u6570\u636e\u96c6，\u56e0\u6b64，\u5f53\u65f6\u5728\u8be5\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u6d4b\u8bd5\u7684\u65b9\u6cd5\u6570\u91cf\u8f83\u5c11，\u7279\u522b\u662f\u65e0 Gloss \u65b9\u6cd5\u66f4\u5c11\u3002\u4ece\u7ed3\u679c\u4e2d\u53ef\u4ee5\u770b\u51fa，\u672c\u6587\u65b9\u6cd5\u5728\u6240\u6709\u6307\u6807\u4e0a\u90fd\u8d85\u8fc7\u4e86\u65e0 Gloss \u65b9\u6cd5 NSLT + Luong\u3002\u6b64\u5916，\u5728\u4e0e\u57fa\u4e8e Gloss \u7684\u65b9\u6cd5\u6bd4\u8f83\u65f6，\u672c\u6587\u65b9\u6cd5\u5df2\u7ecf\u63a5\u8fd1 SLRT \u548c BN-TIN-Transf.\u3002\u8fd9\u4e24\u79cd\u65b9\u6cd5\u6ca1\u6709\u91c7\u7528\u66f4\u590d\u6742\u7684\u8f85\u52a9\u8bad\u7ec3\u7b56\u7565\u3002<\/p>\n
4.4.\u6d88\u878d\u5b9e\u9a8c<\/h3>\n
\u6d88\u878d\u5b9e\u9a8c\u4e3b\u8981\u5728 PHOENIX14T \u6570\u636e\u96c6\u4e0a\u8fdb\u884c，\u5e76\u91cd\u70b9\u8003\u5bdf BLEU-4 \u7684\u63d0\u5347，\u56e0\u4e3a BLEU-4 \u662f\u8861\u91cf\u624b\u8bed\u7ffb\u8bd1\u51c6\u786e\u6027\u6700\u53ef\u9760\u7684\u6307\u6807\u3002\u9664\u975e\u53e6\u6709\u8bf4\u660e，\u7f51\u7edc\u5747\u91c7\u7528\u7b2c 4.2 \u8282\u6240\u8ff0\u7684\u914d\u7f6e\u4f5c\u4e3a\u57fa\u7ebf\u8bbe\u7f6e\u3002<\/p>\n
\u89c6\u89c9\u2014\u8bed\u8a00\u9884\u8bad\u7ec3<\/h4>\n
\u5728\u5bf9 VLP \u7684\u7814\u7a76\u4e2d，\u4f5c\u8005\u6df1\u5165\u5206\u6790\u4e86\u5f71\u54cd\u5176\u6709\u6548\u6027\u7684\u5173\u952e\u56e0\u7d20，\u5e76\u89c2\u5bdf\u5230\u4ee5\u4e0b\u73b0\u8c61\u3002<\/p>\n
\u9996\u5148，\u4ece Table 3 \u53ef\u4ee5\u770b\u51fa，\u5bf9\u624b\u8bed\u89c6\u9891\u8fdb\u884c\u6570\u636e\u589e\u5f3a，\u5bf9 VLP \u7684\u6210\u529f\u8d77\u7740\u91cd\u8981\u4f5c\u7528\u3002\u5177\u4f53\u800c\u8a00，\u5f53\u4ec5\u4f7f\u7528\u968f\u673a\u88c1\u526a\u7b49\u8f7b\u91cf\u7ea7\u6570\u636e\u589e\u5f3a\u65f6，VLP \u5bf9\u624b\u8bed\u7ffb\u8bd1\u7684\u63d0\u5347\u6bd4\u8f83\u6709\u9650：\u9a8c\u8bc1\u96c6 BLEU-4 \u4ec5\u63d0\u9ad8\u7ea6 0.3 \u5206，\u6d4b\u8bd5\u96c6\u4ec5\u63d0\u9ad8\u7ea6 0.1 \u5206\u3002\u7136\u800c，\u5f53 VLP \u4e0e\u5f3a\u6570\u636e\u589e\u5f3a\u7ed3\u5408\u65f6，\u5b83\u80fd\u591f\u663e\u8457\u6539\u5584\u624b\u8bed\u7ffb\u8bd1\u6027\u80fd，\u4f7f\u9a8c\u8bc1\u96c6 BLEU-4 \u4ece 19.84 \u63d0\u5347\u81f3 22.05，\u63d0\u9ad8\u4e86\u7ea6 2.2 \u5206\u3002\u8fd9\u4e00\u7ed3\u679c\u4f53\u73b0\u4e86 VLP \u5bf9\u6570\u636e\u91cf\u548c\u6570\u636e\u591a\u6837\u6027\u7684\u8f83\u9ad8\u9700\u6c42\u3002<\/p>\n
\u5176\u6b21，\u4f5c\u8005\u53d1\u73b0，\u5728\u6ca1\u6709 VLP \u7684\u60c5\u51b5\u4e0b，\u4ec5\u5728\u7b2c\u4e8c\u9636\u6bb5\u4f7f\u7528\u5f3a\u6570\u636e\u589e\u5f3a\u5e76\u4e0d\u80fd\u4e3a\u624b\u8bed\u7ffb\u8bd1\u5e26\u6765\u660e\u663e\u6536\u76ca，\u751a\u81f3\u53ef\u80fd\u7565\u5fae\u635f\u5bb3\u6a21\u578b\u6027\u80fd\u3002\u4f46\u5f53\u5f3a\u6570\u636e\u589e\u5f3a\u4e0e VLP \u7ed3\u5408\u65f6，\u624b\u8bed\u7ffb\u8bd1\u6027\u80fd\u53ef\u4ee5\u5f97\u5230\u8fdb\u4e00\u6b65\u63d0\u5347\u3002\u8fd9\u662f\u56e0\u4e3a，\u6fc0\u8fdb\u7684\u6570\u636e\u589e\u5f3a\u65b9\u6cd5\u53ef\u80fd\u4f1a\u5728\u8bad\u7ec3\u6570\u636e\u4e2d\u5f15\u5165\u8fc7\u591a\u53d8\u5316\u6216\u5931\u771f，\u4f7f\u624b\u8bed\u7ffb\u8bd1\u6a21\u578b\u96be\u4ee5\u9002\u5e94\u589e\u5f3a\u6570\u636e\u7684\u5206\u5e03\u3002\u76f8\u6bd4\u4e4b\u4e0b，VLP \u9636\u6bb5\u5229\u7528\u5927\u8bed\u8a00\u6a21\u578b\u63d0\u4f9b\u7684\u8bed\u8a00\u76d1\u7763，\u4fc3\u4f7f Visual Encoder \u9002\u5e94\u589e\u5f3a\u6570\u636e\u6240\u5e26\u6765\u7684\u5206\u5e03\u5dee\u5f02，\u4ece\u800c\u5e2e\u52a9\u4e0b\u6e38\u624b\u8bed\u7ffb\u8bd1\u6a21\u578b\u83b7\u5f97\u5bf9\u589e\u5f3a\u6570\u636e\u7684\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n
\u6b64\u5916，\u4ece Table 4 \u53ef\u4ee5\u53d1\u73b0，\u5c06 Visual Embedding \u6a21\u5757\u548c Transformer Encoder \u4f5c\u4e3a\u4e00\u4e2a\u7edf\u4e00\u6574\u4f53\u8fdb\u884c\u9884\u8bad\u7ec3\u53c2\u6570\u8fc1\u79fb\u4e0e\u5fae\u8c03，\u4e0e\u53ea\u8fc1\u79fb\u5176\u4e2d\u4e00\u4e2a\u6a21\u5757\u76f8\u6bd4，\u53ef\u4ee5\u5e26\u6765\u663e\u8457\u7684\u6027\u80fd\u63d0\u5347\u3002<\/p>\n
\u6700\u540e，\u4f5c\u8005\u89c2\u5bdf\u5230，\u4f7f\u7528\u9884\u8bad\u7ec3\u7684 Text Decoder \u4e5f\u80fd\u591f\u5e26\u6765\u4e00\u5b9a\u63d0\u5347，\u4f46\u589e\u76ca\u76f8\u5bf9\u6709\u9650，\u4e0d\u8d85\u8fc7 1 \u5206\u3002<\/p>\n
\u8fd9\u4e9b\u7ed3\u679c\u8bc1\u5b9e，\u9ad8\u8d28\u91cf\u7684\u89c6\u89c9\u7279\u5f81\u5bf9\u4e8e\u65e0 Gloss \u624b\u8bed\u7ffb\u8bd1\u81f3\u5173\u91cd\u8981\u3002VLP \u7b56\u7565\u4fc3\u4f7f Visual Encoder \u5b66\u4e60\u8bed\u8a00\u8868\u793a\u6240\u5177\u6709\u7684\u4f4e\u5197\u4f59\u3001\u9ad8\u62bd\u8c61\u7279\u5f81，\u56e0\u6b64 Visual Encoder \u662f\u6574\u4e2a\u7cfb\u7edf\u4e2d\u7684\u5173\u952e\u7ec4\u6210\u90e8\u5206\u3002<\/p>\n
\u8bad\u7ec3\u65f6\u95f4\u7814\u7a76<\/h4>\n
\u57fa\u4e8e Gloss \u7684\u624b\u8bed\u7ffb\u8bd1\u6a21\u578b，\u8bad\u7ec3\u65f6\u95f4\u901a\u5e38\u4e0d\u4f1a\u8d85\u8fc7 100 \u4e2a Epoch\u3002\u7136\u800c，\u5982 Table 5 \u6240\u793a，\u5728\u56fa\u5b9a\u9884\u8bad\u7ec3\u65f6\u95f4\u7684\u60c5\u51b5\u4e0b，\u65e0 Gloss \u624b\u8bed\u7ffb\u8bd1\u6a21\u578b\u9700\u8981\u66f4\u957f\u7684\u8bad\u7ec3\u5468\u671f，\u5373\u8d85\u8fc7 100 \u4e2a Epoch，\u624d\u80fd\u83b7\u5f97\u4ee4\u4eba\u6ee1\u610f\u7684\u6027\u80fd\u3002<\/p>\n
\u8fd9\u662f\u56e0\u4e3a，\u5728\u6ca1\u6709\u4e2d\u95f4\u8868\u793a\u8f85\u52a9\u7684\u60c5\u51b5\u4e0b，\u7f51\u7edc\u7684\u6536\u655b\u901f\u5ea6\u4f1a\u964d\u4f4e，\u56e0\u6b64\u9700\u8981\u66f4\u957f\u7684\u8bad\u7ec3\u65f6\u95f4，\u624d\u80fd\u4f7f\u6a21\u578b\u8fbe\u5230\u9884\u671f\u62df\u5408\u6548\u679c\u3002<\/p>\n
\u6b64\u5916，\u4f5c\u8005\u8fd8\u7814\u7a76\u4e86\u9884\u8bad\u7ec3\u65f6\u957f\u5bf9\u6a21\u578b\u6027\u80fd\u7684\u5f71\u54cd\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e，\u4f3c\u4e4e\u6ca1\u6709\u5fc5\u8981\u8fdb\u884c\u8fc7\u957f\u65f6\u95f4\u7684\u9884\u8bad\u7ec3\u3002\u7efc\u5408\u4e24\u4e2a\u624b\u8bed\u6570\u636e\u96c6\u7684\u60c5\u51b5，80 \u4e2a Epoch \u662f\u4e00\u4e2a\u8f83\u4e3a\u5408\u7406\u7684\u6298\u4e2d\u9009\u62e9\u3002\u5e76\u56fa\u5b9a\u7b2c\u4e8c\u9636\u6bb5\u4e3a 200 Epoch\u3002<\/p>\n
\u6a21\u578b\u53c2\u6570\u89c4\u6a21\u7684\u5f71\u54cd<\/h4>\n
\u4eba\u4eec\u666e\u904d\u8ba4\u4e3a，\u7f51\u7edc\u53c2\u6570\u89c4\u6a21\u4f1a\u663e\u8457\u5f71\u54cd\u6a21\u578b\u7684\u6700\u7ec8\u6027\u80fd；\u4e00\u79cd\u6bd4\u8f83\u76f4\u89c2\u7684\u89c2\u70b9\u662f，\u7f51\u7edc\u8d8a\u6df1，\u6027\u80fd\u8d8a\u597d\u3002<\/p>\n
\u7136\u800c，\u5bf9\u4e8e GFSLT \u7f51\u7edc，\u4f5c\u8005\u53d1\u73b0，\u589e\u52a0\u7f51\u7edc\u5c42\u6570\u4f1a\u5bfc\u81f4\u66f4\u52a0\u4e25\u91cd\u7684\u8fc7\u62df\u5408，\u5982 Figure 4 \u4e2d\u7684\u7eff\u8272\u66f2\u7ebf\u6240\u793a\u3002\u4f5c\u8005\u8ba4\u4e3a，\u8fd9\u662f\u7531\u4e8e\u624b\u8bed\u7ffb\u8bd1\u6570\u636e\u96c6\u89c4\u6a21\u6709\u9650\u3002\u62e5\u6709\u8db3\u591f\u5927\u7684\u624b\u8bed\u7ffb\u8bd1\u6570\u636e\u96c6，\u53ef\u80fd\u80fd\u591f\u7f13\u89e3\u8fd9\u4e00\u95ee\u9898\u3002<\/p>\n
\u51bb\u7ed3 Text Encoder \u7684\u5f71\u54cd<\/h4>\n
\u8003\u8651\u5230 Text Encoder \u6765\u6e90\u4e8e\u9884\u8bad\u7ec3\u7684 mBART，\u4f5c\u8005\u5728\u8fd9\u4e00\u5b9e\u9a8c\u4e2d\u5c1d\u8bd5\u51bb\u7ed3\u5176\u53c2\u6570，\u5c06\u5176\u4f5c\u4e3a\u6559\u5e08\u6a21\u578b，\u76d1\u7763 Visual Encoder \u7684\u5b66\u4e60\u3002<\/p>\n
\u4e0e\u9884\u671f\u76f8\u53cd，\u8fd9\u79cd\u9884\u8bad\u7ec3\u7b56\u7565\u5e76\u6ca1\u6709\u4ea7\u751f\u4ee4\u4eba\u6ee1\u610f\u7684\u7ed3\u679c，\u5982 Table 6 \u6240\u793a\u3002\u4f5c\u8005\u63a8\u6d4b，\u539f\u56e0\u53ef\u80fd\u5728\u4e8e\u6587\u672c\u7279\u5f81\u548c\u89c6\u89c9\u7279\u5f81\u5728\u5e95\u5c42\u8868\u793a\u4e0a\u5b58\u5728\u6839\u672c\u5dee\u5f02\u3002\u4e3a\u4e86\u8fdb\u884c\u6709\u610f\u4e49\u7684\u6bd4\u8f83\u548c\u5206\u6790，\u4e24\u79cd\u6a21\u6001\u9700\u8981\u88ab\u5171\u540c\u4f18\u5316\u5230\u4e00\u4e2a\u5171\u4eab\u8868\u793a\u7a7a\u95f4\u4e2d\u3002\u56e0\u6b64，\u76f4\u63a5\u51bb\u7ed3 Text Encoder \u7684\u53c2\u6570，\u53ef\u80fd\u65e0\u6cd5\u4e3a Visual Encoder \u5b66\u4e60\u9c81\u68d2\u8868\u793a\u63d0\u4f9b\u5145\u5206\u6307\u5bfc\u3002<\/p>\n
4.5.\u5b9a\u6027\u7ed3\u679c<\/h3>\n
\u4f5c\u8005\u5728 Table 7 \u4e2d，\u76f4\u89c2\u5c55\u793a\u4e86\u6a21\u578b\u5728 PHOENIX14T \u6d4b\u8bd5\u96c6\u82e5\u5e72\u624b\u8bed\u89c6\u9891\u4e0a\u7684\u8868\u73b0\u3002\u867d\u7136\u4e24\u4e2a\u6a21\u578b\u90fd\u80fd\u591f\u7406\u89e3\u624b\u8bed\u89c6\u9891\u7684\u6574\u4f53\u542b\u4e49，\u5e76\u751f\u6210\u5b8c\u6574\u7684\u53e5\u5b50，\u4f46\u57fa\u7ebf\u6a21\u578b\u5728\u67d0\u4e9b\u5173\u952e\u8bcd\u4e0a\u66f4\u5bb9\u6613\u51fa\u9519，\u4ece\u800c\u4ea7\u751f\u542b\u4e49\u5dee\u5f02\u5f88\u5927\u7684\u7ffb\u8bd1\u7ed3\u679c，\u5982\u8868\u4e2d\u7684\u7b2c\u4e00\u884c\u548c\u7b2c\u4e8c\u884c\u6240\u793a\u3002\u6b64\u5916，VLP \u6a21\u578b\u5728\u8bc6\u522b\u547d\u540d\u5b9e\u4f53\u65b9\u9762\u4f18\u4e8e\u57fa\u7ebf\u6a21\u578b，\u80fd\u591f\u66f4\u52a0\u51c6\u786e\u5730\u7ffb\u8bd1\u5730\u70b9\u540d\u79f0\u548c\u6708\u4efd，\u5982\u7b2c\u4e09\u884c\u548c\u7b2c\u56db\u884c\u6240\u793a\u3002<\/p>\n
5.\u7ed3\u8bba\u548c\u672a\u6765\u5de5\u4f5c<\/h2>\n
\u5728\u672c\u7814\u7a76\u4e2d，\u4f5c\u8005\u4ece\u7f29\u5c0f\u89c6\u89c9\u8868\u793a\u4e0e\u6587\u672c\u8868\u793a\u4e4b\u95f4\u8bed\u4e49\u5dee\u8ddd\u7684\u89d2\u5ea6，\u4e3a\u65e0 Gloss \u624b\u8bed\u7ffb\u8bd1\u4efb\u52a1\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u601d\u8def，\u4f7f\u6a21\u578b\u80fd\u591f\u4ece\u624b\u8bed\u89c6\u9891\u4e2d\u5b66\u4e60\u8bed\u8a00\u6307\u793a\u7684\u89c6\u89c9\u8868\u793a\u3002\u4e3a\u5b9e\u73b0\u8fd9\u4e00\u76ee\u6807，\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u9884\u8bad\u7ec3\u8303\u5f0f，\u5c06\u63a9\u7801\u81ea\u76d1\u7763\u5b66\u4e60\u4e0e\u89c6\u89c9\u2014\u8bed\u8a00\u76d1\u7763\u5b66\u4e60\u76f8\u7ed3\u5408\u3002<\/p>\n
\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e，\u6570\u636e\u89c4\u6a21\u548c\u6a21\u578b\u53c2\u6570\u89c4\u6a21\u90fd\u4f1a\u5bf9\u8be5\u65b9\u6cd5\u7684\u6027\u80fd\u4ea7\u751f\u663e\u8457\u5f71\u54cd\u3002\u5c3d\u7ba1\u672c\u6587\u63d0\u51fa\u7684\u9884\u8bad\u7ec3\u8303\u5f0f\u662f\u8fc8\u5411\u65e0 Gloss \u624b\u8bed\u7ffb\u8bd1\u7684\u91cd\u8981\u4e00\u6b65，\u4f46\u6211\u4eec\u4e5f\u8ba4\u8bc6\u5230\u4ecd\u9700\u5f00\u5c55\u8fdb\u4e00\u6b65\u7814\u7a76，\u5c24\u5176\u662f\u5728\u5927\u89c4\u6a21\u3001\u65e0 Gloss \u6807\u6ce8\u7684\u624b\u8bed\u7ffb\u8bd1\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u9884\u8bad\u7ec3\u3002\u6211\u4eec\u5e0c\u671b\u672c\u6587\u7684\u5de5\u4f5c\u80fd\u591f\u542f\u53d1\u8be5\u9886\u57df\u672a\u6765\u7684\u7814\u7a76\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
\u5199\u5728\u524d\u9762
\n\u672c\u535a\u5ba2\u662f\u6211\u5bf9\u8fd9\u7bc7\u8bba\u6587\u7684\u7ffb\u8bd1\u548c\u7b14\u8bb0\u7406\u89e3，\u8bba\u6587\u539f\u6587\u94fe\u63a5：ICCV 2023 Open Access Repository，\u4ee3\u7801\u5f00\u6e90\u3002GitHub – zhoubenjia\/GFSLT-VLP \u00b7 GitHub \u76ee\u5f55
\n\u5199\u5728\u524d\u9762
\n\u8bba\u6587\u201c\u6545\u4e8b\u201d\u6982\u8ff0
\n\u8bba\u6587\u8981\u89e3\u51b3\u7684\u95ee\u9898
\n\u8bba\u6587\u7684\u6838\u5fc3\u60f3\u6cd5
\n\u9884\u8bad\u7ec3\u548c\u5bf9\u6bd4\u5b66\u4e60\u76f8\u5173\u6982\u5ff5\u8865\u5145
\n0.\u6458\u8981
\n1.Introduction
\n2.\u76f8\u5173\u5de5\u4f5c
\n2.1.\u624b\u8bed\u8bc6\u522b
\n2.2.\u57fa\u4e8e Gloss \u7684\u624b<\/p>\n","protected":false},"author":2,"featured_media":86793,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[9801,9800,9799,50,3546],"topic":[],"class_list":["post-86797","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-9801","tag-9800","tag-9799","tag-50","tag-3546"],"yoast_head":"\n\u8bba\u6587\u7cbe\u8bfb-\u300aGloss-free Sign Language Translation: Improving from Visual-Language Pretraining\u300b - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/86797.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u8bba\u6587\u7cbe\u8bfb-\u300aGloss-free Sign Language Translation: Improving from Visual-Language Pretraining\u300b - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u5199\u5728\u524d\u9762 \u672c\u535a\u5ba2\u662f\u6211\u5bf9\u8fd9\u7bc7\u8bba\u6587\u7684\u7ffb\u8bd1\u548c\u7b14\u8bb0\u7406\u89e3，\u8bba\u6587\u539f\u6587\u94fe\u63a5：ICCV 2023 Open Access Repository，\u4ee3\u7801\u5f00\u6e90\u3002GitHub - zhoubenjia\/GFSLT-VLP \u00b7 GitHub \u76ee\u5f55 \u5199\u5728\u524d\u9762 \u8bba\u6587\u201c\u6545\u4e8b\u201d\u6982\u8ff0 \u8bba\u6587\u8981\u89e3\u51b3\u7684\u95ee\u9898 \u8bba\u6587\u7684\u6838\u5fc3\u60f3\u6cd5 \u9884\u8bad\u7ec3\u548c\u5bf9\u6bd4\u5b66\u4e60\u76f8\u5173\u6982\u5ff5\u8865\u5145 0.\u6458\u8981 1.Introduction 2.\u76f8\u5173\u5de5\u4f5c 2.1.\u624b\u8bed\u8bc6\u522b 2.2.\u57fa\u4e8e Gloss \u7684\u624b\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/86797.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-07-29T01:08:44+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/07\/20260729010841-6a6952996b783.png\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"15 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/86797.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/86797.html\",\"name\":\"\u8bba\u6587\u7cbe\u8bfb-\u300aGloss-free Sign Language Translation: Improving from Visual-Language Pretraining\u300b - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-07-29T01:08:44+00:00\",\"dateModified\":\"2026-07-29T01:08:44+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/86797.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/86797.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/86797.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u8bba\u6587\u7cbe\u8bfb--\u300aGloss-free Sign Language Translation: Improving from Visual-Language Pretraining\u300b\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u8bba\u6587\u7cbe\u8bfb-\u300aGloss-free Sign Language Translation: Improving from Visual-Language Pretraining\u300b - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/86797.html","og_locale":"zh_CN","og_type":"article","og_title":"\u8bba\u6587\u7cbe\u8bfb-\u300aGloss-free Sign Language Translation: Improving from Visual-Language Pretraining\u300b - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u5199\u5728\u524d\u9762 \u672c\u535a\u5ba2\u662f\u6211\u5bf9\u8fd9\u7bc7\u8bba\u6587\u7684\u7ffb\u8bd1\u548c\u7b14\u8bb0\u7406\u89e3，\u8bba\u6587\u539f\u6587\u94fe\u63a5：ICCV 2023 Open Access Repository，\u4ee3\u7801\u5f00\u6e90\u3002GitHub - zhoubenjia\/GFSLT-VLP \u00b7 GitHub \u76ee\u5f55 \u5199\u5728\u524d\u9762 \u8bba\u6587\u201c\u6545\u4e8b\u201d\u6982\u8ff0 \u8bba\u6587\u8981\u89e3\u51b3\u7684\u95ee\u9898 \u8bba\u6587\u7684\u6838\u5fc3\u60f3\u6cd5 \u9884\u8bad\u7ec3\u548c\u5bf9\u6bd4\u5b66\u4e60\u76f8\u5173\u6982\u5ff5\u8865\u5145 0.\u6458\u8981 1.Introduction 2.\u76f8\u5173\u5de5\u4f5c 2.1.\u624b\u8bed\u8bc6\u522b 2.2.\u57fa\u4e8e Gloss \u7684\u624b","og_url":"https:\/\/www.wsisp.com\/helps\/86797.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-07-29T01:08:44+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/07\/20260729010841-6a6952996b783.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"15 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/86797.html","url":"https:\/\/www.wsisp.com\/helps\/86797.html","name":"\u8bba\u6587\u7cbe\u8bfb-\u300aGloss-free Sign Language Translation: Improving from Visual-Language Pretraining\u300b - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-07-29T01:08:44+00:00","dateModified":"2026-07-29T01:08:44+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/86797.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/86797.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/86797.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u8bba\u6587\u7cbe\u8bfb--\u300aGloss-free Sign Language Translation: Improving from Visual-Language Pretraining\u300b"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/86797","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=86797"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/86797\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/86793"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=86797"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=86797"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=86797"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=86797"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
- \n
- \n
- BLEU-$n$ \u4e3b\u8981\u8003\u5bdf\u6a21\u578b\u8bd1\u6587\u4e0e\u53c2\u8003\u8bd1\u6587\u4e4b\u95f4 $n$-gram \u7684\u5339\u914d\u60c5\u51b5\u3002\u4f8b\u5982：\n
- \u8fd9\u91cc\u7684\u201c\u5171\u4eab\u6743\u91cd\u201d\u610f\u5473\u7740：\n
- \n
- \n
- \n
- （a）\u5305\u542b\u4e24\u6761\u540c\u65f6\u8fdb\u884c\u7684\u8bad\u7ec3\u5206\u652f\n