README.md

Introduction

• History
  • Rule-based (e.g., AIML)
  • Statistical-based (e.g., POMDP (Partially Observable Markov Decision Process): the next action is determined by an estimated current state)
  • Deep learning-based
• Categories
  • Task-oriented (in closed domain)
  • Non-task-oriented (i.e., chat bots [8], in open domain) [1]
  • Question answering
    • FAQ based
    • KBQA
• Question answering tells us how to do and dialogue system tells us the steps to do it successfully

1. Task-oriented

1.1 Pipeline (ASR -> NLU -> DM -> NLG)

• NLU: maps the utterance into semantic slots

  • Domain identification: text classification
  • User domain specific intent detection: text classification (We can incorporate preceding text information during perform classification, see [7])
    • Contextual(i.e. multi-turn) SLU with memory network with implementation and tutorial
  • Slot filling: sequence labeling
    • Parameters of the INTENT
    • Since slots are intent-specific, we can Joint learning of intent identification and slot filling, but put more emphasis on slot error
    • Add lexicon feature
  • Industrial
    • Not slot: for example, the user do not like something
    • Interchange of different intents: when the bot ask the slot of one intent, the user ask another intent
    • To avoid OOV
      • Character-based classification and sequence labeling
      • Fasttext sub n-gram
      • Combine CNN or LSTM on character embedding and word embedding
    • New intent or new domain
      • Zero-shot learning
    • For intent conflicts in multiple domain, we can use adversarial multi-task learning

• DM: Dialogue Management

  • Dialogue State Tracking: aka belief state, is a probability distribution over each slot for each turn

    • DSTC
    • TRADE, with implementation

  • Policy Learning: generates next action

    • Supervised learning
    • Reinforcement learning
      • Entity: agent and environment
      • Rule: agent takes action and environment gives reward and state
      • Policy: a rule that the agent should follow to select actions given the current state

  • Models

    • Rule-based is popular in production system [8]
      • Frame based
      • Finite state machine
        • Node is action
        • Edge is slot status
      • Agenda based
    • Markov Decision Process
    • Neural network

• NLG: convert action to utterance

  • Action template
  • Hybrid model [8]
    • If the user dialogue actions can be found in the predefined sentence templates, the **template-based NLG ** is applied; otherwise, the utterance is generated by the model-based NLG

1.2 End-to-end

• 2017_IJCNLP_End-to-End Task-Completion Neural Dialogue Systems.pdf, A_User_Simulator_for_Task-Completion_Dialogues: RL to train DM based on simulation data

  • RL is based on whole dialogue, not each turn. A rule-based agent is used to warm-start the system (i.e., experience replay), via supervised learning. Episode is a whole dialogue. s_t, a_t, r, s_t+1
  • Error control to mimic mistake of real user
  • Based on request_slots (i.e., user want to ask) and inform_slots (i.e., user can provide) to generate simulation data
  • A simplified version, and blog, RECOMMEND!

• 2017_EACL_A Network-based End-to-End Trainable Task-oriented Dialogue System

• 2017_ICLR_Learning_end-to-end_goal_oriented_dialog: memory network

• Report of Dr.Che

1.3 Knowledge grounded

• Mem2Seq
• GLMP

1.4 User simulator

• How to Build User Simulators to Train RL-based Dialog Systems

1.5 Additional concept

• Dialog act: more general intentions, such as statement, question, option

2. Non-task-oriented

2.1 Retrieval-based

• Given pre-defined responses, retrieval based model predict one response given current input context [2]
• Two steps [3]:
  • retrieval top-k response candidates by directly matching
    • Elasticsearch
  • reranking and give best candidate by incorporing context
    • Apply knlowledge distillation from BERT to ESIM
• Multi-turn matching
  • [https://github.com/JasonForJoy/Leaderboards-for-Multi-Turn-Response-Selection](Leaderboard of state-of-the-art models)

2.2 Generative

• Generative new responses from scratch [2]
• Multi-turn
  • HRED transfered from query suggestion, see industrail application
  • VHRED
  • Building End-To-End Dialogue Systems Using Generative Hierarchical Neural Network Models
  • Context-sensitive ACL 2018
• Avoid generic responses
  • Mutual information
• CS224s slides

2.3 Hybrid

• Alime

3. KBQA

to do...

4. Dialogue history mining

4.1 Intent automatic induction

• Alternating view k-means: two encoders (one for user first utterance, another for remaining dialogue), one encoder -> clustering -> update another encoder by prototypical training, and alternating the above step

4.2 FAQ based

• FAQ automatically building
• Industrial experiences

Dataset

• bAbI: book a table at a restaurant
• Ubuntu Dialogue Corpus v2.0
• A New Multi-Turn, Multi-Domain, Task-Oriented Dialogue Dataset
• Dialog State Tracking Challenge
• Douban Conversation Corpus
• 2017第六届全国社会媒体处理大会之中文人机对话技术评测
• MultiWOZ: contains 7 domain and 30 (do-main, slot) pairs
• CrossWOZ: Chinese Cross-Domain WOZ, it contains 5 domains (i.e., hotel, restaurant, attraction, metro, and taxi)

Implementation

• Baidu UNIT and lecture

• Google DialogFlow and lecture

• Rasa

  • Rasa NLU
    • Intent classification
    • Entity recognition
  • Rasa Core
    • Transformer Embedding Policy, refer to arXiv for latest version
  • Rasa UI
  • tensorflow_embedding in NLU from StarSpace: Embed All The Things!
  • Chinese chatbot example

• ConvLab: multi-domain end-to-end dialog system platform, aiming to enable researchers to quickly set up experiments

• ConvLab 2

• wxpy

• wxBot

• WeRoBot: weChat bot

• Awesome-Chatbot gives a list of public repositories about chatbot

• DeepPavlov

• Wit.ai

• ChatterBot

• pyDial: The Cambridge University Python Multi-domain Statistical Dialogue System Toolkit

Products

• IP梦工厂
• 图灵机器人
• 小i机器人
• 海知智能
• XiaoIce
• Pandorabots
• Rokid

Leading Researchers

• MiuLab
• Minlie Huang
• Rui Yan
• Nan Duan

References

[1] A Survey on Dialogue Systems: Recent Advances and New Frontiers
[2] Deep Learning for Chatbots part 1, part 2
[3] https://c.m.163.com/news/l/180148.html?spssid=fba792c95ad60299db5435a91da37e10&spsw=1&spss=other.
[4] A Network-based End-to-End Trainable Task-oriented Dialogue System, and implementation, Ali Xiaomi
[5] End-to-end LSTM-based dialog control optimized with supervised and reinforcement learning, Hybrid Code Networks: practical and efficient end-to-end dialog control with supervised and reinforcement learning, and implementation
[6] Learning End-to-End Goal-Oriented Dialog from Facebook, and implementation
[7] Sequential Short-Text Classification with Recurrent and Convolutional Neural Networks
[8] “Chitty-Chitty-Chat Bot”: Deep Learning for Conversational AI
[9] End-to-End Task-Completion Neural Dialogue Systems, implementation
[10] 任务型对话系统研究综述
[11] 多轮人机对话与对话管理技术探索与实践
[12] 任务导向型对话系统——对话管理模型研究最新进展
[13] 阿里巴巴高级算法专家李永彬：小蜜智能对话开发平台的技术探索与场景实践
[15] How to use question to increase Chatbot's interactivity ?