[논문 리뷰] Recipes for building an open-domain chatbot ( feat. blenderbot1, parlai, facebook, chatbot, open-domain, bb1, 블렌더봇, 블렌더봇1, 페이스북)

 

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 
• 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

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.