일 | 월 | 화 | 수 | 목 | 금 | 토 |
---|---|---|---|---|---|---|
1 | 2 | 3 | 4 | |||
5 | 6 | 7 | 8 | 9 | 10 | 11 |
12 | 13 | 14 | 15 | 16 | 17 | 18 |
19 | 20 | 21 | 22 | 23 | 24 | 25 |
26 | 27 | 28 | 29 | 30 | 31 |
- 딥러닝
- 빅데이터
- 2021
- 노마드코더
- SQL
- MYSQL
- 리뷰
- 성실히
- 부트캠프
- 열심히
- 기초통계
- yolo
- 파이썬
- 자료구조
- 선형회귀
- leetcode
- 재미져
- selenium
- 매일매일
- 꾸준히
- 독서
- 주간보고
- Codestates
- 코딩테스트
- bootcamp
- Ai
- python
- pandas
- JavaScript
- 코드스테이츠
- Today
- Total
코딩일기
[논문 리뷰] Recipes for building an open-domain chatbot ( feat. blenderbot1, parlai, facebook, chatbot, open-domain, bb1, 블렌더봇, 블렌더봇1, 페이스북) 본문
[논문 리뷰] Recipes for building an open-domain chatbot ( feat. blenderbot1, parlai, facebook, chatbot, open-domain, bb1, 블렌더봇, 블렌더봇1, 페이스북)
daje 2022. 8. 1. 20:42
사전에 숙지해야할 사항
1. Transformers
-. Title : Attention all you need
-. link : https://arxiv.org/pdf/1706.03762.pdf
-. review : 2021.10.04 - [Paper Reviews] - [논문리뷰] Attention is all you need (feat. Transformer)
2. BERT
-. Title : Pre-training of Deep Bidirectional Transformers for Language Understanding
-. link : https://arxiv.org/pdf/1810.04805.pdf
-. review : 진행 예정
3. Poly-encoders
-. Title : architectures and pre-training strategies for fast and accurate multi-sentence scoring
-. link : https://arxiv.org/pdf/1905.01969.pdf
-. review : 진행 예정
4. BART
-. Title : Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension
-. link : https://arxiv.org/pdf/1910.13461.pdf
-. review : 진행 예정
목차
- 0. Abstract
- 1. Introduction
- 2. Model architectures
- 2.1 Retriever
- 2.2 Generator
- 2.3 Retriveve and Refine
- Dialogue Retrieval
- Knowledge Retrieval
- 3. Training Objectives
- 3.1 Ranking for Retrieval
- 3.2 Likelihood Training for Generation
- 3.3 a-blending for Retrieve and Refine
- 3.4 Unlikelihood trainin for generation
- 4. Decoding
- 4.1 Beam Search
- 4.2 Sampling
- 4.3 Response Length
- Minimum length
- Predictive length
- 4.4 Subsequence Blocking
- 5. Training Details
- Pre-training Ranking models
- Pre-training Generative models
- Fine-tuning
- 6. Training Data
- 6.1 Pre-training
- pushshift.io Reddit
- 6.2 Fine-tuning
- ConvAI2
- Empathetic Dialogues(ED)
- Wizard of Wikipedia(WoW)
- Blended Skill Talk
- 6.1 Pre-training
- 7. Safety Characteristics
- 8. Evaluation Methods
- ACUTE-Eval
- Self-Chat ACUTE-Eval
- 9. Related Work
- 10. Results & Analysis
- 10.1 Automatic Evaluations
- Retriever
- Generator
- Retrieve and Refine(RetNRef)
- Safety
- 10.2 Self-Chat Evaluations
- Retrieval vs. Generator vs. RetNRef
- Generator Decoding choices
- Small vs. Large models
- Pre-trainin vs. Fine-Tuning
- Persona context vs. No context given
- Likelihood vs. Unlikehood
- 10.3 Full(Human-Bot Chat) Evaluations
- Retrieval vs. Generator vs. RetNRef
- Comparison to Meena
- Model vs. Human-human Chat Comparisons
- Response Length
- 10.4 Example Successful Conversations
- 10.5 Failure Cases and Mpdel Extensions
- Vocabulary Usage
- Nontrivial Repetition
- Contradiction and Forgetfulness
- Knowledge and Factual Correctness
- Conversation Length and Memory
- Deeper Understanding
- Further Notes onEvaluation
- 11. Released code and models
- 12. Discussion
- 10.1 Automatic Evaluations
0. Abstract
- prior works has shown that scaling neural models in the number of parameters and the size of the data.
- Good conversation requires a number of skills that an expert conversationalist blends in a seamless way.
- We show that large scale models can learn these skills when given appropriate training data and choice of generation strategy and then we build new models.
- Human evaluations show our best models are superior to extisting approaches in multi-turn dialogue.
- 기존 연구들은 더 많은 데이터와 파라미터를 가지는 인공신경망을 만드는데 연구를 해왔습니다.
- 좋은 대화는 전문가적인 스킬과 공감, 개성 등을 반영한 매력적인 대화여야한다고 이야기하고 있습니다.
- 이러한 좋은 대화를 하는 모델을 만들기 위해서는 적절한 훈련데이터와 generation strategy가 필요하다고 이야기하고 있습니다.
- 또한, 새로운 평가 지표를 만들어서 평가를 했다고 이야기하고 있습니다.
1. Introduction
- the pre-training on large corpora is important.
- Beyond simply scaling models the two main takeaways from our study are Blending Skills and Generation Strategies
디코딩 전략을 어떻게 설정하냐에 따라서 perplexity가 같은 두 모델도 엄청나게 다른 결과를 내놓을 수 있습니다.
특히, bot의 utterances의 길이가 사람이 봇의 응답을 판단할 때 큰 영향을 준다고 언급합니다.
이전 연구에서는 beam search가 별로라고 이야기 했지만, minimum beam length를 설정하여 좋은 응답을 뽑아 낼 수 있다고 이야기하고 있습니다.