metadata
license: apache-2.0
construct_text_correction
使用程序自动构造文本纠错数据集,包含拼写和语法纠错数据,可用于中文校对模型的训练。
Data Fields
| Field | Type | Description |
|---|---|---|
| source | string | 可能包含拼写/语法错误的源句子 |
| target | string | 纠错后的目标句子 |
| label | int | 源句子中是否包含错误,若为1,则包含错误,否则不包含错误。 |
{
"source": "健全国有林区经营管理体制,完散集体林权制度改革。",
"target": "健全国有林区经营管理体制,完善集体林权制度改革。",
"label": 1
}
Construction
安装 ltp:
pip install ltp ltp-core ltp-extension
生成 4k 条纠错句子对 4k.jsonl:
python finetune_data.py \
--input sentences/4k.txt \
--output 4k.jsonl \
--ltp_model LTP/legacy \
--basic_hanzi confusion/basic_hanzi_2500.txt \
--shape_confusion confusion/shape_confusion.txt \
--sound_confusion confusion/pinyin.txt \
--same_ratio 0.1 \
--repeat_ratio 0.15 \
--delete_ratio 0.15 \
--sound_ratio 0.5 \
--shape_ratio 0.1
生成 20k 条纠错句子对 20k.jsonl:
python finetune_data.py \
--input sentences/20k.txt \
--output 20k.jsonl \
--ltp_model LTP/legacy \
--basic_hanzi confusion/basic_hanzi_2500.txt \
--shape_confusion confusion/shape_confusion.txt \
--sound_confusion confusion/pinyin_expand.txt \
--same_ratio 0.1 \
--repeat_ratio 0.15 \
--delete_ratio 0.15 \
--sound_ratio 0.5 \
--shape_ratio 0.1