当前位置：网站首页>Acl-ijcai-sigir top conference paper report meeting (AIS 2022) Note 3: dialogue and generation

Acl-ijcai-sigir top conference paper report meeting (AIS 2022) Note 3: dialogue and generation

2022-07-26 16:19:00 【The gods were silent】

The gods were silent - personal CSDN Blog Directory

Video address ：AIS 2022 丨 ACL-IJCAI-SIGIR The top conference paper report will be played back and the video will be made public
Activity homepage of Zhiyuan official website , Introduction with a detailed agenda ：ACL-IJCAI-SIGIR Top conference papers and presentations （AIS 2022）

This article focuses on 5 month 14 Dialogue and generation part of the day . The main content is screenshots slides, And note some explanations . Than I wrote before cs224w And Li Hongyi's notes will be much simpler , I will not introduce the self-evident contents in the figure .

List of articles

1. Cutting edge trend report 2：AIS2022 Overview of dialogue and generative research

Zhang Weinan, Harbin Institute of technology
Insert picture description here
Selected some representative articles

AIS 2022

Insert picture description here
Insert picture description here
Insert picture description here Insert picture description here

ACL 2022

Insert picture description here
Insert picture description here
Insert picture description here
Insert picture description here
Insert picture description here
Insert picture description here Insert picture description here
state tracking（domain Of slot）
Insert picture description here
Insert picture description here
Connected in series as pipeline
Insert picture description here
Insert picture description here

IJCAI 2022

Insert picture description here
Insert picture description here
Insert picture description here

SIGIR 2022

Insert picture description here
Insert picture description here
Insert picture description here
Current session 、 User history information 、 Similar users
Insert picture description here
Insert picture description here Insert picture description here Insert picture description here Insert picture description here

2. Session 3：Dialog and Generation（ Dialogue and generation ）

1. AdaLoGN

Li Xiao, Nanjing University
Insert picture description here
Insert picture description here Insert picture description here
Implication relation
Insert picture description here
The process of deriving other relationships from known relationships → Extension of the structure on the graph
Insert picture description here
One out of four
Insert picture description here
Initial build diagram
Syntactic rules （ no , For example, verbs are opposites of each other ）
Insert picture description here
nerve + Symbolic reasoning
roberta→token Average
Graph Expansion
Insert picture description here Insert picture description here Insert picture description here Insert picture description here Insert picture description here Insert picture description here Insert picture description here

2. A corpus combining task dialogues and open domain dialogues

Baidu Xujun
Insert picture description here Insert picture description here Insert picture description here Insert picture description here Insert picture description here
Insert picture description here
The yellow part ： Knowledge dialogue
Q & a dialogue
Insert picture description here

Corpus

Insert picture description here
Insert picture description here Insert picture description here

The PLATO-MT Model with Prompt-based Mechanism

Insert picture description here

DuClarifyDial as a New Benchmark

Insert picture description here
Insert picture description here
Insert picture description here
Insert picture description here

3. Problem generation

Fudan University
Insert picture description here Insert picture description here
Color is the answer to the request , Models need to learn to ask questions

This paper aims at multi-hop Problem generation ：
Insert picture description here

The diagram
Entities can guarantee the depth of the problem （ What is the occupation of the wife of the president of the United States ）
Insert picture description here
Whether the current entity must appear in the problem （flag）
Insert picture description here Insert picture description here Insert picture description here Insert picture description here Insert picture description here

4. Implicit variables in the pre training model

Chen Wei, Fudan University
Insert picture description here Insert picture description here Insert picture description here Insert picture description here Insert picture description here
Future prediction strategy ： It can make the decoder have better planning ability
Insert picture description here
Insert picture description here
KL loss： Suppose it turns out to be a normal distribution , Minimize distribution differences
free bits：KL loss Optimization starts only after reaching a certain threshold
BOW loss： Implicit variables predict vocabulary directly and non autoregressively , I hope the implicit variable can contain response Information
Insert picture description here
Insert picture description here
First in reddit-short Use a larger one batch size Do the preliminary training , And then reddit-long Use smaller ones batch size Do the preliminary training
Insert picture description here
At the same time, increase the Correlation BLEU And diversity distinct（ Generally, it is not allowed to ）
Insert picture description here
Insert picture description here

5. Search dialog

Chen Wei, Fudan University
Insert picture description here Insert picture description here This article focuses on the first stage , A large number of response, extract
Insert picture description here
Retrieval efficiency
Insert picture description here Insert picture description here Insert picture description here
Example ： It's the same response
Contrast the loss
Insert picture description here
Insert picture description here
Insert picture description here Insert picture description here Insert picture description here Insert picture description here
Insert picture description here

6. Online companion robot 、 Emotional dialogue

Xiaomi Li Yanran

6.1 MISC

Insert picture description here
Insert picture description here
Insert picture description here
Insert picture description here
It should mean 2021 Tsinghua's article in ：Towards Emotional Support Dialog Systems
Ask about , Reasoning background information , Do counseling
Insert picture description here
Knowledge base
comet Model ： Common sense generation model , Input situation, Infer the emotional state 、 demand
Insert picture description here
knowledge-enhanced context representation
Choose the distribution of strategies , Multiply with the matrix , obtain mixed stragy representation
Insert picture description here
The above paper is this
Insert picture description here
Insert picture description here Insert picture description here Insert picture description here
Mixing strategies for long sentences
Only the original supervision signal is needed

6.2 C 3 K G C^3KG C3KG

common sense + dialogue

Insert picture description here Insert picture description here
What knowledge should be chosen to help understand the dialogue
Insert picture description here
Insert picture description here
Insert picture description here
One round dialogue , Mark emotions and intentions
Insert picture description here
event Turning point
Insert picture description here

Insert picture description here

6.3 HICR

Plain text appeases → Rich media / Multimodal
Insert picture description here
Insert picture description here
Play a game 、 listen to the music
Insert picture description here Insert picture description here
Recommend based on user information
Insert picture description here Insert picture description here Insert picture description here
Insert picture description here

7. COSPLAY

Xu Chen, Beijing University of Technology
Insert picture description here

7.1 background

Insert picture description here
Insert picture description here
Insert picture description here

7.2 motivation

Express more to the other party while maintaining consistency persona The opportunity of
Insert picture description here

7.3 Model

Reinforcement learning
Encourage questions
Insert picture description here
Insert picture description here
Insert picture description here
Force the model to chat around common topics Insert picture description here

7.4 Experiment and analysis

Insert picture description here
Insert picture description here
Insert picture description here
Purple indicates that the model pays attention to itself , Dark color means to the other party
Insert picture description here

8. Neutral statements demonstrate why

Lijiangnan from the Information Industry Office
Insert picture description here Insert picture description here Insert picture description here The disadvantage is the lack of use of emotional information
Insert picture description here Insert picture description here Insert picture description here Insert picture description here
Insert picture description here Insert picture description here Insert picture description here Insert picture description here Insert picture description here Insert picture description here Insert picture description here Insert picture description here Insert picture description here Cluster prediction
Insert picture description here Insert picture description here Insert picture description here Insert picture description here

原网站

版权声明
本文为[The gods were silent]所创，转载请带上原文链接，感谢
https://yzsam.com/2022/207/202207261558531762.html

边栏推荐

猜你喜欢

随机推荐