|
|
--- |
|
|
library_name: pytorch, transformers |
|
|
base_model: KikKoh/Hok2Han |
|
|
license: apache-2.0 |
|
|
tags: |
|
|
- seq2seq |
|
|
- transformer |
|
|
- taiwanese |
|
|
- pinyin-to-chinese |
|
|
--- |
|
|
|
|
|
# Model Card for Hok2Han |
|
|
|
|
|
本模型為基於 PyTorch 實作的 Seq2Seq Transformer，用於將台語拼音轉換成台語漢字。 |
|
|
|
|
|
## Model Details |
|
|
|
|
|
### Model Description |
|
|
|
|
|
本模型利用自訂架構 Seq2Seq Transformer，學習從台語拼音序列映射到對應的台語漢字序列。訓練資料包含大量台語語料及對應拼音標註，模型架構包含6層編碼器與解碼器，採用512維嵌入與8頭注意力機制。 |
|
|
|
|
|
* **Developed by:** KikKoh |
|
|
* **Model type:** Seq2Seq Transformer |
|
|
* **Language(s):** Taiwanese (Hokkien) |
|
|
* **License:** Apache-2.0 |
|
|
* **Finetuned from model:** 自訂架構，非標準預訓練模型 |
|
|
|
|
|
### Model Sources |
|
|
|
|
|
* **Repository:** [https://huggingface.co/KikKoh/Hok2Han](https://huggingface.co/KikKoh/Hok2Han) |
|
|
* **Config and weights:** Hugging Face Hub |
|
|
|
|
|
## Uses |
|
|
|
|
|
### Direct Use |
|
|
|
|
|
可用於台語拼音轉漢字的自動翻譯、語音識別後處理等應用場景。 |
|
|
|
|
|
### Out-of-Scope Use |
|
|
|
|
|
不適用於非台語拼音輸入、其他語言翻譯或語音直接識別。 |
|
|
|
|
|
## Bias, Risks, and Limitations |
|
|
|
|
|
模型僅訓練於台語拼音資料，對其他方言、口音或非標準拼音可能表現不佳。使用時應注意語料多樣性限制及可能產生誤翻譯。 |
|
|
|
|
|
## How to Get Started with the Model |
|
|
|
|
|
```python |
|
|
from hok2han_model import Seq2SeqTransformer |
|
|
model = Seq2SeqTransformer.from_pretrained("KikKoh/Hok2Han") |
|
|
model.eval() |
|
|
|
|
|
from transformers import Wav2Vec2Processor, BertTokenizer |
|
|
input_processor = Wav2Vec2Processor.from_pretrained("你的輸入tokenizer路徑或repo") |
|
|
output_tokenizer = BertTokenizer.from_pretrained("你的輸出tokenizer路徑或repo") |
|
|
|
|
|
# 範例推論 |
|
|
output = model(src=input_ids, tgt=tgt_ids, |
|
|
src_pad_idx=input_processor.tokenizer.pad_token_id, |
|
|
tgt_pad_idx=output_tokenizer.pad_token_id) |
|
|
|
|
|
pred_ids = output.argmax(dim=-1) |
|
|
pred_text = output_tokenizer.decode(pred_ids[0], skip_special_tokens=True) |
|
|
print(pred_text) |
|
|
``` |
|
|
|
|
|
## Training Details |
|
|
|
|
|
### Training Data |
|
|
|
|
|
使用台語拼音與漢字對照語料，包含公開及自建資料。 |
|
|
|
|
|
### Training Procedure |
|
|
|
|
|
使用標準Seq2Seq Transformer訓練方法，採用交叉熵損失，AdamW優化器。 |
|
|
|
|
|
## Evaluation |
|
|
|
|
|
評估主要依據拼音到漢字的轉換準確率及語句流暢度。 |
|
|
|
|
|
## Environmental Impact |
|
|
|
|
|
訓練過程使用標準GPU伺服器，耗電與碳排放量中等。 |
|
|
|
|
|
## Technical Specifications |
|
|
|
|
|
### Model Architecture and Objective |
|
|
|
|
|
6層編碼器與解碼器，512維嵌入，8頭多頭注意力。 |
|
|
|
|
|
--- |
|
|
|
|
|
歡迎聯絡 KikKoh |
|
|
Facebook: [https://www.facebook.com/kikkoh.2024](https://www.facebook.com/kikkoh.2024) |
|
|
|
|
|
--- |
