current position:Home>At 7:30 tonight, senior algorithm engineer of Alibaba Dharma Academy - "semi supervised pre training dialogue model space"

At 7:30 tonight, senior algorithm engineer of Alibaba Dharma Academy - "semi supervised pre training dialogue model space"

2022-05-15 05:17:12Aitime theory

Click on the blue words

85c18433c87556782dd3f2cbf7dc7cce.png

Pay attention to our

AI TIME Welcome to everyone AI Fans join in !

5 month 10 Japan 、11 Japan 、12 Friday night 19:30, In this issue, we invite Dai yinpei, senior algorithm engineer of Alibaba Dharma Academy 、 Wang Benyou, a researcher of EU Marie Curie, and Zhang Lifeng, a lecturer at the school of information of Renmin University of China, brought you wonderful sharing !

d09622f114b368c2a95ea52c3af52a3f.gif

5 month 10 Japan 19:30-20:30

2398967def6451994aee5057d3275b5f.gif

c7f3ae0d534ed0799a6658818b49149e.png

Dai yinpei :

Senior algorithm engineer of Alibaba Dharma Academy , Master graduated from the Department of electronic engineering of Tsinghua University , His research field is natural language processing and dialogue intelligence (Conversational AI), Specific directions include dialogue and understanding 、 Dialogue management and large-scale pre training dialogue model . stay ACL / AAAI / SIGIR/ ICASSP He has published many papers and served as ACL / EMNLP / NAACL / AAAI Wait for the reviewer of the meeting .

Share content :

462377f7685b790c631f78041095a3de.gif

Semi supervised pre training dialogue model SPACE

Introduction to the report :

How to integrate human prior knowledge into the pre training model at low cost has always been a problem NLP The problem of .

In this work , A new training paradigm based on semi supervised pre training is proposed by the dialogue intelligent team of Dharma Academy , A small amount of marked dialogue data and a large amount of non-standard dialogue data are pre trained through semi supervised method , Using the consistent regularization loss function, the dialogue strategy knowledge contained in the labeled data is injected into the pre training model , So as to learn a better model representation .

A new semi supervised pre training dialogue model SPACE(Semi-Supervised Pre-trAined Conversation ModEl) Firstly, it focuses on the knowledge of dialogue strategies .

Experiments show that ,SPACE1.0 Model in Cambridge MultiWOZ2.0, Amazon MultiWOZ2.1 And other classic conversation data sets 5%+ Significantly improve , And under various low resource settings ,SPACE1.0 Better than existing sota The models have stronger small sample learning ability .

5 month 11 Japan 19:30-20:30

d09aaa4d9e26214f7389fac9fd759cc4.gif

e7bf576261cebae63d45c1639af1ccb5.png

Ben you wang :

Doctoral student at the University of Padua, Italy , Researcher Marie Curie of the European Union . Will be in 2022 year 6 He joined the Chinese University of Hong Kong as an assistant professor ( Shenzhen ) School of data science . Under the guidance of professors song Dawei and Zhang Peng, he obtained a master's degree from Tianjin University , At the University of Copenhagen in Denmark , University of Montreal, Canada , University of Amsterdam in the Netherlands , Huawei Noah Ark Laboratory , Institute of theoretical physics, Chinese Academy of Sciences , Exchange visit to the Language Institute of the Chinese Academy of social sciences . In industrial applications , He 2017 He worked full-time at Tencent since , As the main algorithm designer , Built a robust intelligent customer service system from zero on Tencent cloud . In a relatively short academic career , He is committed to building more robust and intelligent natural language processing systems , Give consideration to technical rationality and linguistic motivation . So far, he and his collaborators have won the top international information retrieval conference SIGIR 2017 Nomination for best paper and international top conference on natural language processing NAACL 2019 Best interpretable paper , Published, including international top conferences ICLR/SIGIR/WWW/NAACL/AAAI/IJCAI/CIKM etc. 20 Yu Wen .

Share content :

9f8dedcf7a758122b6b6daba4dd630e0.gif

On position embeddings

Introduction to the report :

transformer Widely used in nlp Mission ( Especially the pre training model ), Even started using cv The field is coming to the fore .transformer Its structure will not be able to model the order of input without position coding , So location coding is very important .

At present, many pre training models use different location codes ( For example, fully learnable location coding , Trigonometric function fixed position coding , Relative position coding ), Good experience , However, there is a lack of a unified framework to understand and evaluate these location codes .

Let's first explain why the motivation of trigonometric function position coding , In short, it is to replace the translation by rotation , To inject position information into the word vector .

Our latest work formalizes some principled properties of location coding ( Translation invariant , Monotonicity and symmetry ), And evaluate the extent to which the existing location codes meet these attributes , Finally, quantitatively evaluate how these attributes benefit / Damage downstream tasks .

We find that fully learnable location coding works well in whole sentence classification scenarios , Thanks to its flexibility CLS special token And normal position ; The relative position code is in span prediction Better effect .

Student Recruitment Information

Chinese University of Hong Kong ( Shenzhen ) Wang Benyou, School of Data Science / Professor Li Haizhou's team recruited natural language processing / Voice Processing / In the direction of machine learning 3 Full award doctoral students ( Sure 2022FALL, 2022WINTER, 2023 Admission ),3 Research Assistant ,6 Famous blogger .

The team has strong links between industry and academia , It has rich computing resources , There are enough resources to train the super large pre training language model , Give full play to the scientific research creativity of team members .

Doctoral applications for admission this fall should start as early as June , Both undergraduate and master students can apply for a doctorate , Need IELTS or TOEFL scores ( A foreign degree can be exempted ), The doctor is awarded a degree certificate issued by the Chinese University of Hong Kong ;RA and post-doc Anytime , Until it's full .

For details, please see https://wabyking.github.io/files/JD4PhD-CUHKSZ.pdf   perhaps https://zhuanlan.zhihu.com/p/500582441 .

Details can also be consulted [email protected]

5 month 12 Japan 19:30-20:30

2ea373193746bf92e6cd073581cb0f1c.gif

6dee8384b22364e24be7c3ef2be7dc35.png

Zhang Lifeng :

Lecturer, School of information, Renmin University of China . The main research interests include two parts :1) Research on the methodology of intelligent optimization algorithms such as evolutionary computing , And the application of management operation research algorithm and decision support system in production practice ;2) Research on the theory and methodology of system identification and machine learning , And the application of statistical methods in various fields of data analysis .

Share content :

373b295223c4654e3ca2d20721aae20f.gif

Quickly detect the complex correlation between data

Introduction to the report :

Detecting and distinguishing the relationship between variables is a basic work of data analysis , Quickly finding and measuring the variables with correlation not only saves researchers' time , It also provides valuable direction guidance for subsequent analysis and modeling .

This study proposes a new kind of statistical tools , Nearest neighbor correlation coefficient (nCor), From a new perspective , Can effectively detect continuous 、 discrete , And categorical variables .

And all kinds of mutual information (MI) Our valuation algorithm 、MIC、dCor、RDC、HSIC Compared with the hot methods in recent years , New methods for various data types 、 Complex relationships are more applicable , Stronger detection ability and robustness .

The new method can also better distinguish predictable 、 Heteroscedastic 、 The interaction of , And all kinds of complex data relationships that overlap , Provide more in-depth and effective guidance for follow-up analysis and research .

This study is based on three papers published in recent years , This paper expounds the principle and specific implementation methods of the new statistics in different application situations .

After the live broadcast, you can ask questions in the group , Please add “AI TIME Little helper ( WeChat ID :AITIME_HY)”, reply “PhD-4”, Will pull you into “AI TIME PhD Communication group -4”!

84873d6f3ef91297f9381375d6b12509.gif

AI TIME Wechat assistant

465f01be2c8d5e9dae77e0cdf7b77b7f.png

Lord         do :AI TIME 

Associated Media :AI Data pie

partners : Wisdom spectrum ·AI、 Chinese Academy of Engineering Zhiling live 、 School Online 、 Kou enjoys academic 、AMiner、 Ever Chain action 、 Scientific research cloud 、 An endless stream of Science

Excellent articles in the past are recommended

dae458431624cdb756f13bf7190f79e6.png

Remember to pay attention to us ! There is new knowledge every day !

  About AI TIME 

AI TIME From 2019 year , It aims to carry forward the spirit of scientific speculation , Invite people from all walks of life to the theory of artificial intelligence 、 Explore the essence of algorithm and scenario application , Strengthen the collision of ideas , Link the world AI scholars 、 Industry experts and enthusiasts , I hope in the form of debate , Explore the contradiction between artificial intelligence and human future , Explore the future of artificial intelligence .

so far ,AI TIME Has invited 600 Many speakers at home and abroad , Held more than 300 An event , super 150 10000 people watch .

7e840496938094db6744e270f7956c58.png

I know you.

Looking at

Oh

~

4ee596deea91b7264d12b2adcc3fd41d.gif

Click on Read the original   Reservation live broadcast !

copyright notice
author[Aitime theory],Please bring the original link to reprint, thank you.
https://en.cdmana.com/2022/131/202205111246309527.html

Random recommended