当前位置:网站首页>Wenxin Ernie 3.0 blessing! Small samples can also achieve 99% effect of full data!
Wenxin Ernie 3.0 blessing! Small samples can also achieve 99% effect of full data!
2022-06-09 21:58:00 【kaiyuan_ sjtu】
In recent years , With AI Technological development ,NLP Technology has been gradually “ Mount guard ” To various industrial application scenarios , Automatically handle complex and repetitive work , Such as automatic classification of news content 、 Intelligent customer service automatically replies 、 Review of sensitive words 、 User comments, emotion analysis, etc .

One 、 Why? NLP It has become a sharp tool for enterprises to reduce costs and increase efficiency ?
In Finance 、 government affairs 、 law 、 In many industries such as medical treatment , A large amount of document information is generated every day and needs to be processed . Take text information processing as an example :
An auditor can audit at most in a day 5000 A text message , And it is difficult to guarantee the accuracy , And online UGC Information websites often receive more than one million text messages in an hour on average .
This shows that in a real business scenario , All through manpower to achieve information production 、 Handle 、 It is almost impossible to release the whole process tracking , urgent need NLP Technology to achieve intelligent processing of information , Reduce cost and increase efficiency for enterprises .
Two 、NLP “ Mount guard ” To several major problems in the business
NLP The application of technology in the actual scene is not as smooth as expected , Mainly from NLP In the process of customizing the scenario application model , There are many problems :
Data annotation is time-consuming and laborious : Self built models require manual annotation of business data , laborious , The marking cost is very high , Greatly affect business progress ;
There is no clue about the scheme selection : Do not know the optimal model solution , Uncertain model evaluation indicators , I don't know how to tune the model effect ;
It is difficult to deploy the model : The deployment scheme suitable for the business scenario is not clear , It is difficult to implement deployment and development 、 The high cost .
3、 ... and 、 How to zero code 、 High quality AI demand ?
To address these issues , Baidu PaddlePaddle EasyDL It provides you with 「 "One-stop" work style NLP Task development services 」, The data 、 Training and deployment were all taken over , It's also achieved Full process automation , Users only need to drag and drop according to the platform prompts , Don't understand the algorithm 、 Not being able to write code is not a problem .
at present ,EasyDL Text categorization is already supported 、 Text creation 、 Analysis of emotional tendencies 、 Short text similarity matching 、 Entity extraction 、 Entity relation extraction 、 Comment opinion extraction and other task types , comprehensive 、 Efficient 、 Conveniently solve the actual business needs of small and medium-sized enterprises .

Click to read the original text GET
EasyDL - Text experience link
https://ai.baidu.com/easydl/nlp/
Data phase :「AI staff 」 Help with efficient labeling
To help enterprises reduce costs in data preparation , Improve the dimensioning effect ,EasyDL Provided by the platform “AI staff ” Smart label service . in application , Only a small amount of manual annotation is required , smart “AI staff ” You can intelligently label other unlabeled data , Effectively solve the big problem of data annotation , Help enterprises reduce costs in data preparation , Improve the dimensioning effect .
Training phase : only 20% Small sample data to achieve high-precision model effect
EasyDL NLP Recently, I will write a letter ERNIE Big model 「 base 」 Upgrade to 3.0. What does that mean? ? Let's start from the following aspects , See how the big model of literary mind makes EasyDL Text ability is more powerful :
Massive Chinese data knowledge reserves : The big model of literary mind ERNIE 3.0 As a large model for knowledge enhancement of tens of billions of parameters , In addition to learning vocabulary from massive text data 、 structure 、 In addition to semantic knowledge , And learn from large-scale knowledge maps . therefore EasyDL NLP The task shows better effect on Chinese model training , Don't worry that the model can't understand Chinese anymore ~
Small sample quick training : Wenxin big model ERNIE 3.0 Handle both language understanding and language generation tasks , Good training effect can be achieved through a small amount of training data . At present EasyDL NLP The marking amount can be reduced to the original 20%.
The effect of the task is leading : Wenxin big model ERNIE 3.0 Refresh at one stroke 54 Chinese NLP Mission benchmark , Including emotional analysis 、 Take out ideas 、 reading comprehension 、 Text in this paper, 、 Dialogue generation 、 Mathematical operations and other tasks . Verified by authoritative public data sets , Various types NLP The average accuracy of the task is as high as 90% above .

Multi scene creation ability : Wenxin big model ERNIE 3.0 In terms of literary creation ability, it has been significantly improved , Through the study of massive texts and knowledge , Give Way EasyDL Of “ Text creation ” The task does not require special training , You can start a novel 、 The lyrics 、 poetry 、 Couplets and other literary creation . If you don't talk much, it will be effective .


Slide down to see everything
See here , Little buddy must have found out ,EasyDL Text ability is in the big model of literary mind ERNIE 3.0「 base 」 Have strong general knowledge ability with the support of , It's like a Wulin expert who has practiced internal skill for many years . With this general knowledge , Only a small amount of specific business scenario data is required 「 Grasp a typical example and you will grasp the whole category 」, Realization NLP Business landing .
Deployment phase : Free choice in many ways
Public cloud API: Users can directly call the... Provided by Baidu cloud API To use , Fast and easy .
Local server deployment : For some localized 、 Requirements for privatized deployment , Users can flexibly access the deployment mode of localized services , High performance can also be achieved serving Ability .
Four 、 Extensive and mature practical application Help all walks of life AI upgrade
at present EasyDL Zero threshold of 、 Strong professionalism and other characteristics have been widely accepted by small and medium-sized enterprises , Use EasyDL The number of users has exceeded 100 ten thousand , Cover 20 Multiple industry scenarios , Including the Internet 、 Industry 、 Agriculture 、 Medical care 、 logistics 、 retail 、 education 、 Transportation, etc .
Enterprise service : Hancai headhunter uses text classification model , It is realized through intelligent annotation function on the platform 199 Ten thousand pieces of data are automatically marked , Finally, the training accuracy reached 95%+ Of “ Candidate functions ”、“ Candidate rank ” Wait for the model , Intellectualization solves the problem of the company 200 The problem of selecting resumes from the talent pool .
Financial field : Jiniu technology has the customization function based on the text entity extraction model , The key information extraction based on insurance agent visit log is realized autonomously , Effectively improve the efficiency of customer intelligent operation .
Logistics field : An Internet moving platform uses EasyDL Text classification filter high-quality user messages , Judge whether users place orders effectively , Accurate positioning of target users , Recognition accuracy reaches 97% above , Effectively improve the overall operation efficiency of the platform .
E-commerce : Pigeon book number card ERP The order management system is connected to the propeller EasyDL Text processing technology , It realizes the automatic classification and similarity matching of tens of thousands of error messages returned by dozens of upstream operators , With high accuracy 87% about , Greatly simplifies labor costs , Reduce the time cost in the order production process .
Live class preview
6 month 9 Friday night 20:00, Baidu NLP The product manager will bring a wonderful live broadcast , analyse NLP Three pitfalls and corresponding solutions that cannot be ignored in industrial application development , Reading EasyDL How to achieve NLP Industrial application landing , And take you to the actual combat of the project .
Welcome to scan the code into the group
Get Course Links !

Group entry benefits
obtain 6 month 9 Links to daily live classes
Participate in 「 Classification of news and information 」「 Analysis of e-commerce comments 」 Combat camp ,15 Minutes of easy training with high accuracy NLP Model , More exquisite gifts and certificates are distributed free of charge

Read more
Thesis link :
https://arxiv.org/pdf/2107.02137.pdf
Demo link :
https://wenxin.baidu.com/younger/apiDetail?id=20006
边栏推荐
- 快递单信息抽取【三】--五条标注数据提高准确率,仅需五条标注样本,快速完成快递单信息任务
- Inconsistency between the model on swagger and the returned data fields
- Deploy MySQL based on statefulset in kubernetes (Part 2)
- Tke builds efk log service
- Paddlenlp--uie (II) -- fast performance improvement with small samples (including doccona label)
- Pychart always displays the collecting data solution after entering the debug mode
- 数据库每日一题---第7天:订单最多的客户
- What is wave field TRX wallet development
- js 强制类型转换 和 隐式类型转换 和 Unicode编码
- Upper computer development (opening)
猜你喜欢
![[flow analysis] Buu_ [an Xun cup 2019]attack](/img/6a/e8bd90e931ef8dfdca1cd9c66f0695.png)
[flow analysis] Buu_ [an Xun cup 2019]attack

深入理解 Go Modules 的 go.mod 與 go.sum

leetcode:547、朋友圈

spider pi 智能视觉六足机器人 开箱介绍 0602

MFC connection database shows no data source name found and no default driver specified

保存和复制绘图时保留最少的空白

Understand go modules' go Mod and go sum

es5中构造函数的属性继承 借用父构造函数 方法继承 原型对象

易买网开发 趣买买 数据库的导入与数据库结构一览表 0605

Spider PI intelligent vision hexapod robot color recognition function 0603
随机推荐
Modbus protocol and serialport port read / write
Mqtt graphical client-mqttx installation and use tutorial
A thorough understanding of the very important ticket lock in concurrent programming -- stampedlock
Do your filial duty to make an old people's fall prevention alarm system for your family
js 自增和自减(一元运算符)
PostgreSQL近期常用的錶結構查詢語句
JS auto increment and auto decrement (unary operator)
spider pi 智能视觉六足机器人 标签识别 ApirlTag标签 0604
How to locate the cause of high CPU on the server
spider pi 智能视觉六足机器人 开箱介绍 0602
BLE链路层空中包格式
TL,你是如何管理项目风险的?
Highly recommended: data annotation platform doccano - introduction, installation, use and pit stepping records
An RS485 serial interface current sensor snap on type mutual inductor supports Modbus communication protocol
数据库每日一题---第7天:订单最多的客户
保存和复制绘图时保留最少的空白
剑指offer1-32题思路
Matlab implementation of Pettitt mutation test
发电厂企业的关口表参数里的组合无功1和组合无功2的含义--抄表数采问题
Huawei announced the establishment of three corps and two system departments, and has established 20 Corps in total