87 lines
2.5 KiB
Markdown
87 lines
2.5 KiB
Markdown
---
|
||
pipeline_tag: sentence-similarity
|
||
tags:
|
||
- sentence-transformers
|
||
- feature-extraction
|
||
- sentence-similarity
|
||
widget:
|
||
source_sentence: "亚丝娜"
|
||
sentences:
|
||
- "火影忍者"
|
||
- "Sword Art Online"
|
||
- "结城明日奈"
|
||
- "アスナ"
|
||
|
||
---
|
||
|
||
# ACGVoc2vec
|
||
|
||
结构为[sentence-transformers](https://github.com/UKPLab/sentence-transformers)，使用其**distiluse-base-multilingual-cased-v2**预训练权重，以5e-5的学习率在动漫相关语句对数据集下进行微调，损失函数为MultipleNegativesRankingLoss。
|
||
|
||
数据集主要包括：
|
||
|
||
* Bangumi
|
||
|
||
* 动画日文名-动画中文名
|
||
* 动画日文名-简介
|
||
* 动画中文名-简介
|
||
* 动画中文名-标签
|
||
* 动画日文名-角色
|
||
* 动画中文名-角色
|
||
* 声优日文名-声优中文名
|
||
|
||
* pixiv
|
||
|
||
* 标签日文名-标签中文名
|
||
* AnimeList
|
||
|
||
* 动画日文名-动画英文名
|
||
|
||
* 维基百科
|
||
|
||
* 动画日文名-动画中文名
|
||
* 动画日文名-动画英文名
|
||
* 中英日详情页h2标题及其对应文本
|
||
* 简介多语言对照（中日英）
|
||
* 动画名-简介（中日英）
|
||
|
||
* moegirl
|
||
|
||
* 动画中文名的简介-简介
|
||
* 动画中文名+小标题-对应内容
|
||
|
||
在进行爬取，清洗，处理后得到8000w对文本对（还在持续增加），batchzise=80训练了20个epoch，使st的权重能够适应该问题空间，生成融合了领域知识的文本特征向量（体现为有关的文本距离更加接近，例如作品与登场人物，或者来自同一作品的登场人物）。
|
||
|
||
## Usage
|
||
|
||
Using this model becomes easy when you have [sentence-transformers](https://www.SBERT.net) installed:
|
||
|
||
```
|
||
pip install -U sentence-transformers
|
||
```
|
||
|
||
Then you can use the model like this:
|
||
|
||
```python
|
||
from sentence_transformers import SentenceTransformer
|
||
sentences = ["This is an example sentence", "Each sentence is converted"]
|
||
|
||
model = SentenceTransformer('OysterQAQ/ACGVoc2vec')
|
||
embeddings = model.encode(sentences)
|
||
print(embeddings)
|
||
```
|
||
|
||
|
||
## Full Model Architecture
|
||
|
||
```
|
||
SentenceTransformer(
|
||
(0): Transformer({'max_seq_length': 128, 'do_lower_case': False}) with Transformer model: DistilBertModel
|
||
(1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': True, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False})
|
||
(2): Dense({'in_features': 768, 'out_features': 512, 'bias': True, 'activation_function': 'torch.nn.modules.activation.Tanh'})
|
||
)
|
||
```
|
||
|
||
## Citing & Authors
|
||
|
||
<!--- Describe where people can find more information --> |