MonteXiaofeng
commited on
Commit
•
a61f0c7
1
Parent(s):
f2fd2d4
Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,70 @@
|
|
1 |
-
---
|
2 |
-
license: apache-2.0
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: apache-2.0
|
3 |
+
language:
|
4 |
+
- en
|
5 |
+
base_model:
|
6 |
+
- BAAI/bge-m3
|
7 |
+
---
|
8 |
+
|
9 |
+
该模型是数据集[BAAI/IndustryCorpus2](https://huggingface.co/datasets/BAAI/IndustryCorpus2)用于31个行业分类
|
10 |
+
|
11 |
+
## 模型细节:
|
12 |
+
|
13 |
+
为了提升数据集中行业划分对实际行业的覆盖,并对齐国家标准中定义的行业目录,我们参考国家统计局制定的国民经济行业分类体系和世界知识体系,进行类目的合并和整合,设计了覆盖中英文的最终的31个行业类目。类目表名称如下所示
|
14 |
+
|
15 |
+
```
|
16 |
+
{
|
17 |
+
"数学_统计": {"zh": "数学与统计", "en": "Math & Statistics"},
|
18 |
+
"体育": {"zh": "体育", "en": "Sports"},
|
19 |
+
"农林牧渔": {"zh": "农业与渔业", "en": "Agriculture & Fisheries"},
|
20 |
+
"房地产_建筑": {"zh": "房地产与建筑", "en": "Real Estate & Construction"},
|
21 |
+
"时政_政务_行政": {"zh": "政治与行政", "en": "Politics & Administration"},
|
22 |
+
"消防安全_食品安全": {"zh": "安全管理", "en": "Safety Management"},
|
23 |
+
"石油化工": {"zh": "石油化工", "en": "Petrochemicals"},
|
24 |
+
"计算机_通信": {"zh": "计算机与通信", "en": "Computing & Telecommunications"},
|
25 |
+
"交通运输": {"zh": "交通运输", "en": "Transportation"},
|
26 |
+
"其他": {"zh": "其他", "en": "Others"},
|
27 |
+
"医学_健康_心理_中医": {"zh": "健康与医学", "en": "Health & Medicine"},
|
28 |
+
"文学_情感": {"zh": "文学与情感", "en": "Literature & Emotions"},
|
29 |
+
"水利_海洋": {"zh": "水利与海洋", "en": "Water Resources & Marine"},
|
30 |
+
"游戏": {"zh": "游戏", "en": "Gaming"},
|
31 |
+
"科技_科学研究": {"zh": "科技与研究", "en": "Technology & Research"},
|
32 |
+
"采矿": {"zh": "采矿", "en": "Mining"},
|
33 |
+
"人工智能_机器学习": {"zh": "人工智能", "en": "Artificial Intelligence"},
|
34 |
+
"其他信息服务_信息安全": {"zh": "信息服务", "en": "Information Services"},
|
35 |
+
"学科教育_教育": {"zh": "学科教育", "en": "Subject Education"},
|
36 |
+
"新闻传媒": {"zh": "新闻传媒", "en": "Media & Journalism"},
|
37 |
+
"汽车": {"zh": "汽车", "en": "Automobiles"},
|
38 |
+
"生物医药": {"zh": "生物医药", "en": "Biopharmaceuticals"},
|
39 |
+
"航空航天": {"zh": "航空航天", "en": "Aerospace"},
|
40 |
+
"金融_经济": {"zh": "金融与经济", "en": "Finance & Economics"},
|
41 |
+
"住宿_餐饮_酒店": {"zh": "住宿与餐饮", "en": "Hospitality & Catering"},
|
42 |
+
"其他制造": {"zh": "制造业", "en": "Manufacturing"},
|
43 |
+
"影视_娱乐": {"zh": "影视与娱乐", "en": "Film & Entertainment"},
|
44 |
+
"旅游_地理": {"zh": "旅游与地理", "en": "Travel & Geography"},
|
45 |
+
"法律_司法": {"zh": "法律与司法", "en": "Law & Justice"},
|
46 |
+
"电力能源": {"zh": "电力与能源", "en": "Power & Energy"},
|
47 |
+
"计算机编程_代码": {"zh": "编程", "en": "Programming"},
|
48 |
+
}
|
49 |
+
```
|
50 |
+
|
51 |
+
## 行业分类模型的数据构造
|
52 |
+
|
53 |
+
- 数据构建
|
54 |
+
|
55 |
+
数据来源:预训练预训练语料抽样和开源文本分类数据,其中预训练语料占比90%,通过数据采样,保证中英文数据占比为1:1
|
56 |
+
标签构造:使用LLM模型对数据进行多次分类判定,筛选多次判定一致的数据作为训练数据
|
57 |
+
数据规模:36K
|
58 |
+
数据构造的整体流程如下:
|
59 |
+
|
60 |
+
![image-20240919140307205](./img/classify.png)
|
61 |
+
|
62 |
+
## 模型训练:
|
63 |
+
|
64 |
+
参数更新:在预训练的bert模型上添加分类头进行文本分类模型训练
|
65 |
+
|
66 |
+
模型选型:考虑的模型性能和推理效率,我们选用了0.5b规模的模型,通过对比实验最终最终选择了bge-m3并全参数训练的方式,作为我们的基座模型
|
67 |
+
|
68 |
+
训练超参:全参数训练,max_length = 2048,lr=1e-5,batch_size=64,,验证集评估acc:86%
|
69 |
+
|
70 |
+
![image-20240919141408659](./img/classify_exp.png)
|