WangZeJun commited on
Commit
b6149ac
·
verified ·
1 Parent(s): 741abf7

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +62 -0
README.md CHANGED
@@ -1,3 +1,65 @@
1
  ---
2
  license: apache-2.0
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: apache-2.0
3
  ---
4
+
5
+ # construct_text_correction
6
+
7
+ 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。
8
+
9
+ ## Data Fields
10
+
11
+ | Field | Type | Description |
12
+ | ------ | ------ | ----------------------------- |
13
+ | source | string | 可能包含拼写/语法错误的源句子 |
14
+ | target | string | 纠错后的目标句子 |
15
+ | 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。 |
16
+
17
+ ```json
18
+ {
19
+ "source": "健全国有林区经营管理体制,完散集体林权制度改革。",
20
+ "target": "健全国有林区经营管理体制,完善集体林权制度改革。",
21
+ "label": 1
22
+ }
23
+ ```
24
+
25
+ ## Construction
26
+
27
+ 安装 [ltp](https://github.com/HIT-SCIR/ltp):
28
+
29
+ ```shell
30
+ pip install ltp ltp-core ltp-extension
31
+ ```
32
+
33
+ 生成 4k 条纠错句子对 `4k.jsonl`:
34
+
35
+ ```shell
36
+ python finetune_data.py \
37
+ --input sentences/4k.txt \
38
+ --output 4k.jsonl \
39
+ --ltp_model LTP/legacy \
40
+ --basic_hanzi confusion/basic_hanzi_2500.txt \
41
+ --shape_confusion confusion/shape_confusion.txt \
42
+ --sound_confusion confusion/pinyin.txt \
43
+ --same_ratio 0.1 \
44
+ --repeat_ratio 0.15 \
45
+ --delete_ratio 0.15 \
46
+ --sound_ratio 0.5 \
47
+ --shape_ratio 0.1
48
+ ```
49
+
50
+ 生成 20k 条纠错句子对 `20k.jsonl`:
51
+
52
+ ```shell
53
+ python finetune_data.py \
54
+ --input sentences/20k.txt \
55
+ --output 20k.jsonl \
56
+ --ltp_model LTP/legacy \
57
+ --basic_hanzi confusion/basic_hanzi_2500.txt \
58
+ --shape_confusion confusion/shape_confusion.txt \
59
+ --sound_confusion confusion/pinyin_expand.txt \
60
+ --same_ratio 0.1 \
61
+ --repeat_ratio 0.15 \
62
+ --delete_ratio 0.15 \
63
+ --sound_ratio 0.5 \
64
+ --shape_ratio 0.1
65
+ ```