对话AI的论文列表

论文列表格式
 论文发表年份： 论文题目&论文链接：第一作者（第一作者所属学校/机构），代码链接

Overview

Existing Models of Dialog System

Task-Oriented Dialog

• 13: POMDP-Based Statistical Spoken Dialog Systems: A Review: Steve Young(Cambridge University)
• 11: Spoken Language Understanding: Systems for Extracting Semantic Information from Speech: Book!
• 11:Data-Driven Response Generation in Social Media: Alan Ritter(University of Washington Seattle)

• 15: A Neural Network Approach to Context-Sensitive Generation of Conversational Responses: Alessandro Sordoni(Universite de Montreal)

• 15: A Neural Conversational Model: Oriol Vinyals(Google), code via tensorflow

• 15: Neural Responding Machine for Short-Text Conversation: Lifeng Shang(Noah’s Ark Lab), code via theano and tensorflow

Traditional NLP component stack

Challenge of NLP

• 09: SPEECH and LANGUAGE PROCESSING An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition Second Edition: book

Deep Semantic Similarity Model(DSSM)

application scenarios

1. Web search
   • 13: Learning deep structured semantic models for web search using clickthrough data: Po-Sen Huang(University of Illinois at Urbana-Champaign), code via tensorflow
   • 14: A Latent Semantic Model with Convolutional-Pooling Structure for Information Retrieval: Yelong Shen(Microsoft Research)
   • 16: Deep Sentence Embedding Using Long Short-Term Memory Networks: Analysis and Application to Information Retrieval: Hamid Palangi, code
2. Entity linking
   • 14: Modeling Interestingness with Deep Neural Networks: Jianfeng Gao(Microsoft Research)
3. Image captioning
   • 15: From Captions to Visual Concepts and Back: Hao Fang&Li Deng(Microsoft Research)
4. Machine Translation
   • Learning Continuous Phrase Representations for Translation Modeling: Jianfeng Gao(Microsoft Research)
5. Online recommendation
   • [duplicate] 14: Modeling Interestingness with Deep Neural Networks: Jianfneg Gao(Microsoft Research)

Framework of Model

• [duplicate] 13: Learning deep structured semantic models for web search using clickthrough data: Po-Sen Huang(University of Illinois at Urbana-Champaign), [code]
• [duplicate] 14: A Latent Semantic Model with Convolutional-Pooling Structure for Information Retrieval: Yelong Shen(Microsoft Research)
• 16: Deep Sentence Embedding Using Long Short-Term Memory Networks: Analysis and Application to Information Retrieval: Hamid Palangi, code
• Sent2Vec: software by microsoft

Go beyound DSSM

• [duplicate] 15: From Captions to Visual Concepts and Back: Hao Fang&Li Deng(Microsoft Research)

---

Question answeriing(QA) and Machine Readiing Comprehension(MRC)

Open-Domain Question Answering

Knowledge Base-QA

1. Symbolic approach via Large-scale knowledge graphs
   • [oral] 98: MindNet: acquiring and structuring semantic information from text: Stephen D.Richardson(Microsoft Research)
   • [oral] 13: Semantic Parsing on Freebase from Question-Answer Pairs: Jonathan Berant(Stanford University)
   • 15: Attention with Intention for a Neural Network Conversation Model: Kaisheng Yao(Microsoft Research)
   • 14: Knowledge-Based Question Answering as Machine Translation: Junwei Bao(Harbin Institute of Technology)
   • 15: Semantic Parsing via Staged Query Graph Generation: Question Answering with Knowledge Base:Wen-tau Yih(Microsoft Research)

2. ReasoNet with Shared Memory

   • [oral][duplicate] 16: Link Prediction using Embedded Knowledge Graphs: Yulong Shen（Microsoft&Google Research）
   • 17: ReasoNet: Learning to Stop Reading in Machine Comprehension:Yelong Shen(Microsoft Research)

3. Search Controller in ReasoNet

   • [duplicate] 16: Link Prediction using Embedded Knowledge Graphs: Yulong Shen（Microsoft&Google Research）
4. ReasoNet in symbolic vs neural space
   • Symbolic is comprehensible but not robust
     • 11: Random Walk Inference and Learning in A Large Scale Knowledge Base:Ni Lao(Carnegie Mellon University)
     • 98: MindNet: acquiring and structuring semantic information from text:Stephen D.Richardson(Microsoft Research)
   • Neural is robust but not comprehensible
     • [duplicate] 16: Link Prediction using Embedded Knowledge Graphs: Yulong Shen（Microsoft&Google Research）
     • [oral] 15: EMBEDDING ENTITIES AND RELATIONS FOR LEARNING AND INFERENCE IN KNOWLEDGE BASES:Bishan Yang(Cornell University), TensorFlow code, PyTorch code
   • Hybrid is robust and comprehensible
     • 18: M-Walk: Learning to Walk in Graph with Monte Carlo Tree Search:Yelong Shen(Microsoft Research&Tecent AI Lab)
     • 18: [oral] DeepPath: A Reinforcement Learning Method for Knowledge Graph Reasoning:Wenhan Xiong(University of California,Santa Barbara), code1 code2
     • 18: GO FOR A WALK AND ARRIVE AT THE ANSWER: REASONING OVER PATHS IN KNOWLEDGE BASES USING REINFORCEMENT LEARNING:Rajarshi Das(University of Massachusetts,Amherst),
5. Multi-turn KB-QA
   • Programmed Dialogue policy
     • 15: A Probabilistic Framework for Representing Dialog Systems and Entropy-Based Dialog Management through Dynamic Stochastic State Evolution:Ji Wu(IEEE)
   • Trained via RL Dialogue policy
     • 16: Neural Generative Question Answering :Jun Yin(Noah’s Ark Lab, Huawe) corpus
     • [oral] 16: A Network-based End-to-End Trainable Task-oriented Dialogue System:Tsung-Hsien Wen(Cambridge University), Theano code
     • [oral] 17: Towards End-to-End Reinforcement Learning of Dialogue Agents for Information Access:Bhuwan Dhingra(Carnegie Mellon University), Theano code

Text-QA

1. MS MARCO
   • 16: MS MARCO: A Human Generated MAchine Reading COmprehension Dataset:Tri Nguyan(Microsoft AI&Research)
2. SQuAD
   • 16: SQuAD: 100,000+ Questions for Machine Comprehension of Text:Pranav Rajpurkar(Stanford University)

Neural MRC Models

BiDAF

• 16: BI-DIRECTIONAL ATTENTION FLOW FOR MACHINE COMPREHENSION:Minjoon Seo(University of Washington)
  • code1
  • code2
  • code3
  • code4

SAN

• 18: Stochastic Answer Networks for Machine Reading Comprehension: Xiaodong Liu(Microsoft Research,Redmond), code

Neural MRC Models on SQuAD

1. Encoding: map each text span to a semantic vector

   • Word Embedding
     • 14: GloVe: Global Vectors for Word Representation:Jeffrey Pennington(Stanford University)
       • code:midi-glove
       • code:wiki-glove
     • 13: Distributed Representations of Words and Phrases and their Compositionality:Tomas Mikolov(Google Inc.)
       • code1
       • code2

   • Context Embedding

     1. capture context info for each word

        • 16: context2vec: Learning Generic Context Embedding with Bidirectional LSTM:Oren Melamud(Bar-Ilan University)
        • 18: Deep contextualized word representations:Matthew E.Peters(Allen Institute for Artificial Intelligence), code
        • 18: QANET: COMBINING LOCAL CONVOLUTION WITH GLOBAL SELF-ATTENTION FOR READING COMPREHENSION:Adams Wei Yu(CMU&Google Brain)
          • code1
          • code2

     2. Context Embedding via BiLSTM/ELmo

        • [duplicate] 18: Deep contextualized word representations:Matthew E.Peters(Allen Institute for Artificial Intelligence), code
        • 17: Learned in Translation: Contextualized Word Vectors:Bryan McCann(SalesForce)
        • 16: [duplicate]context2vec: Learning Generic Context Embedding with Bidirectional LSTM:Oren Melamud(Bar-Ilan University)

     3. Context Embedding

        • [duplicate] 18: QANET: COMBINING LOCAL CONVOLUTION WITH GLOBAL SELF-ATTENTION FOR READING COMPREHENSION:Adams Wei Yu(CMU&Google Brain)
          • code1
          • code2
   • Query-context/Content-query attention

2. Reasoning: rank and re-rank semantic vectors

   • Multi-step reasoning for Text-QA

     • [duplicate] 17: ReasoNet: Learning to Stop Reading in Machine Comprehension:Yelong Shen(Microsoft Research)

   • Stochastic Answer Net

     • [duplicate] 18: Stochastic Answer Networks for Machine Reading Comprehension: Xiaodong Liu(Microsoft Research,Redmond), code

---

Task-oriented dialogues

overview

A Example Dialogue with Movie-Bot

• source code

Conversation as Reinforcement Learning

• 00: A Stochastic Model of Human-Machine Interaction for Learning Dialog Strategies: Esther Levin(IEEE)
• 00: Optimizing Dialogue Management with Reinforcement Learning: Experiments with the NJFun System:Satinder Singh(AT&T Labs)
• 07: Partially observable Markov decision processes for spoken dialog systems:Jason D.Williams(AT&T Labs)

Dialogue System Evaluation(Simulated Users)

1. Agenda based
   • 09: The Hidden Agenda User Simulation Model:Jost Schatzmann(IEEE)
   • source code
2. Model based
   • 16: A Sequence-to-Sequence Model for User Simulation in Spoken Dialogue Systems: Layla El Asri(Maluuba Research)
   • 17: End-to-End Task-Completion Neural Dialogue Systems:Xiujun Li(Microsoft Research&National Taiwan University)

traditional approache

Decison-theoretic View of Dialogue Management

• [duplicate] 00: Optimizing Dialogue Management with Reinforcement Learning: Experiments with the NJFun System:Satinder Singh(AT&T Labs)
• 00: A Stochastic Model of Human-Machine Interaction for Learning Dialog Strategies: Esther Levin(IEEE)
• 00: Learning Optimal Dialogue Strategies: A Case Study of a Spoken Dialogue Agent for Email: Marilyn A.Walker(ATT Labs Research)
• 02: Automatic learning of dialogue strategy using dialogue simulation and reinforcement learning:Konrad Scheffler(Cambridge University)

Language Understanding Uncertainty: POMDP as a principled framework

• 00: Spoken Dialogue Management Using Probabilistic Reasoning: Nicholas Roy(Carnegie Mellon University)
• 01: Spoken Dialogue Management as Planning and Acting under Uncertainty:Bo Zhang(Tech. of China)
• 07: Partially observable Markov decision processes for spoken dialog systems:Jason D.Williams(AT&T Labs)

scaling up Dialogue Optimization

1. Use approxmiate POMDP algorithms leveraging problem-specific structure
   • 00: Automatic learning of dialogue strategy using dialogue simulation and reinforcement learning:Konrad Scheffler(Cambridge University)
   • 07: Partially observable Markov decision processes for spoken dialog systems:Jason D.Williams(AT&T Labs)
2. Use Reinforcement Learning algorithms with function approximation
   • 08: Hybrid Reinforcement/Supervised Learning of Dialogue Policies from Fixed Data Sets: James Henderson
   • 09: Reinforcement Learning for Dialog Management using Least-Squares Policy Iteration and Fast Feature Selection: Lihong Li(Rutgers University)
   • 14: Incremental on-line adaptation of POMDP-based dialogue managers to extended domains:M.Gasic[Cambridge University]

Natural language understanding and dialogue state tracking

Language Understanding

1. DNN for Domain/Intent Classification

   • 15: Recurrent Neural Network and LSTM Models for Lexical Utterance Classification: Suman Raviuri(University of California,Berkeley)

2. Slot filling

   • 16: Multi-Domain Joint Semantic Frame Parsing using Bi-directional RNN-LSTM: Dilek Hakkani-Tur(Microsoft Research)

3. Further details on NLU

   • ppt
   • E2E MemNN for Contectual LU: End-to-End Memory Networks with Knowledge Carryover for Multi-Turn Spoken Language Understanding: Yun-Nung Chen(National Taiwan University )
   • [duplicate] LU Importance: 17: End-to-End Task-Completion Neural Dialogue Systems:Xiujun Li(Microsoft Research&National Taiwan University)

Dialogue State Tracking(DST)

1. DSTC(Dialog State Tracking Challenge)
   • DSTC1 official website
   • DSTC2&3 official website
   • DSTC4 official website
   • DSTC5 official website

2. Neural Belief Tracker

   • 16: Neural Belief Tracker: Data-Driven Dialogue State Tracking: Nikola Mrksic(University of Cambridge)

3. NN-Based DST

   • 13: Deep Neural Network Approach for the Dialog State Tracking Challenge: Matthew Henderson(University of Cambridge)
   • 15: Multi-domain Dialog State Tracking using Recurrent Neural Networks: Nikola Mrksic(University of Cambridge)
   • [duplicate] 16: Neural Belief Tracker: Data-Driven Dialogue State Tracking: Nikola Mrksic(University of Cambridge)

Deep RL for dialogue policy learning

Two main classed of RL algorithms

1. Value function based:
   • 15: Human-level control through deep reinforcement learning: Volodymyr Minh
     • code1 by tensorflow
     • code2 by pytorch
   • 16: Towards End-to-End Learning for Dialog State Tracking and Management using Deep Reinforcement Learning: Tiancheng Zhao(Carnegie Mellon University)
2. Policy based:
   • 92: Simple statistical gradient-following algorithms for connectionist reinforcement learning: Ronald J.Williams
   • 17: On-line Active Reward Learning for Policy Optimisation in Spoken Dialogue Systems: Pei-Hao Su(University of Cambridge)

Domain Extension and Exploration(BBQ network)

• 18: BBQ-Networks: Efficient Exploration in Deep Reinforcement Learning for Task-Oriented Dialogue Systems: Zachary Lipton(Carnegir Mellon University)

Composite-task Dialogues

1. A Hierarchical Policy Learner
   • 98: Reinforcement Learning with Hierarchies of Machines: Ronald Parr(UC Berkeley)
   • 17: Composite Task-Completion Dialogue Policy Learning via Hierarchical Deep Reinforcement Learning: Baolin Peng(Microsoft Research)
2. Integrating Planning for Dialogue Policy Learning
   • 18: Deep Dyna-Q: Integrating Planning for Task-Completion Dialogue Policy Learning: Baolin Peng(Microsoft Research) , code

Decision-theoretic View of Dialogue Management

Hybrid Code Networks

• 17: Hybrid Code Networks: practical and efficient end-to-end dialog control with supervised and reinforcement learning: Jason D. Williams(Microsoft Research)

Differentiating KB Accesses

• [duplicate] 17: Towards End-to-End Reinforcement Learning of Dialogue Agents for Information Access:Bhuwan Dhingra(Carnegie Mellon University)

An E2E Neural Dialogue System

• [duplicate] 17: End-to-End Task-Completion Neural Dialogue Systems:Xiujun Li(Microsoft Research&National Taiwan University)

---

Fully data-driven conversation models and chatbots

Historical overview

Response retrival system

• 10: Filter, Rank, and Transfer the Knowledge: Learning to Chat:
  Alan Ritter(University of Washington)

Response generation using Statistical Machine Translation

• 11: Data-Driven Response Generation in Social Media: Alan Ritter(University of Washington)

First neural response generation systems

1. Neural Models for Response Generation
   • 15: A Neural Network Approach to Context-Sensitive Generation of Conversational Responses: Alessandro Sordoni(University de Montreal)
   • 15: A Neural Conversational Model: Oriol Vinyals(Google .Inc)
   • 15: Neural Responding Machine for Short-Text Conversation: Lifeng Shang(Noah’s Ark Lab), code
2. Neural conversation engine:
   • 16: A Diversity-Promoting Objective Function for Neural Conversation Models: Jiwei Li(Stanford University)

challenges and remedies

Challenge: The blandness problem

• [duplicate] 16: A Diversity-Promoting Objective Function for Neural Conversation Models: Jiwei Li(Stanford University)

Challenge: The consistency problem

1. Solution: Personalized Response Generation
   • Microsoft Personality chat:speaker embedding LSTM: A Persona-Based Neural Conversation Model: Jiwei Li(Stanford University), code via Pytorch
2. Personal modeling as multi-task learning
   • 17: Multi-Task Learning for Speaker-Role Adaptation in Neural Conversation Models: Yi Luan(University of Washington)
3. Improving personalization with multiple losses
   • 16: Conversational Contextual Cues: The Case of Personalization and History for Response Ranking: Rami Al-Rfou(Google .Inc)

Challenge: Long conversational context

1. It can be challenging for LSTM/GRU to encode very long context
   • 18: Sharp Nearby, Fuzzy Far Away: How Neural Language Models Use Context: Urvashi Khadelwal(Stanford University)
2. Hierarchical Encoder-Decoder(HRED), code
   • 16: Building End-to-End Dialogue Systems Using Generative Hierarchical Neural Network Models: Iulian V.Serban(University de Montreal), code
3. Hierarchical Latent Variable Encoder-Decoder(VHRED)
   • 17: A Hierarchical Latent Variable Encoder-Decoder Model for Generating Dialogues: Iulian V. Serban

Grounded conversation models

A Knowledge-Grounded Neural Conversation Model

• 15: End-To-End Memory Networks: Sainbayar Sukhbaatar(New York University)
  • code1 via Tensorflow
  • code2 via Tensorflow
  • code3 for bAbI QA tasks
• 17: A Knowledge-Grounded Neural Conversation Model: Marjan Gahzvininejad(USC)

Grounded E2E Dialogue Systems

• 16: Visual Dialog: Abhishek Das(Georgia Institute of Tehhnology)
  • code1 via Lua
  • code2 via Pytorch
• 17: Image-Grounded Conversations: Multimodal Context for Natural Question and Response Generation: Nasrin Mostafazadeh(University of Rochster)
• 18: Emotional Dialogue Generation using Image-Grounded Language Models:Bernd Huber(Harvard University)

Beyond supervised learning(Deep Reinforcement Learning for E2E Dialogue)

• 16: Deep Reinforcement Learning for Dialogue Generation:Jiwei Li(Stanford University)
  • code1 via Tensorflow
  • code2 via keras
  • code3 by Jiwei Li

Data and evaluation

Conversational datasets(for social bots, E2E dialogue research)

• 15: A Survey of Available Corpora for Building Data-Driven Dialogue Systems: Iulian Vlad Serban(Universite de Montreal)

Evaluating E2E Dialogue Systems via Autumatic evaluation

1. Machine-Translation-Based Metric
   • 02: BLEU: a Method for Automatic Evaluation of Machine Translation: Kishore Papineni(IBM), code
   • 02: Automatic Evaluation of Machine Translation Quality Using N-gram Co-Occurrence Statistics: George Doddington
2. Sentence-level correlation of MT metrics:
   • 16: How NOT To Evaluate Your Dialogue System: An Empirical Study of Unsupervised Evaluation Metrics for Dialogue Response Generation: Chia-Wei Liu(McGill University)
   • 15: Accurate Evaluation of Segment-level Machine Translation Metrics: Yvette Graham(The University of Melbourne)

The importance of sample size

• [duplicate] 02: BLEU: a Method for Automatic Evaluation of Machine Translation: Kishore Papineni(IBM), code
• 06: Statistical Significance Tests for Machine Translation Evaluation: Philipp Kowehn(MIT)

Corpus-level Correlation

• [duplicate] 02: BLEU: a Method for Automatic Evaluation of Machine Translation: Kishore Papineni(IBM), code
• [duplicate] 06: Statistical Significance Tests for Machine Translation Evaluation:

Chatbot in public

Social Bots: commercial systems

1. For end users
   • Replika.ai system description: replika_ai: Slides
   • XiaoIce:
     15:Chatting With Xiaoice: News
2. For bot developers
   • [duplicate] Microsoft Personality chat:speaker embedding LSTM: A Persona-Based Neural Conversation Model: Jiwei Li(Stanford University), code via Pytorch
   • Microsoft Personality chat’s API: Project Personality Chat’s url

Open Benchmarks

1. Alexa Challenge

   • website: Alexa Prize Proceedings

2. Dialogue System Technology Challenge(DSTC)

   • DSTC7
   • Visual-Scene: DSTC7-Audio-Visual-Scene-Aware-Dialog-AVSD-Challenge 2018
   • background article:
     DSTC7-End-to-End-Conversation-Modeling 2018
   • Registration Link:
     DSTC7 Registration