[2017 - 2018 ACL] 对话系统论文研究点整理
（论文编号及摘要见 [2017 ACL] 对话系统. [2018 ACL Long] 对话系统. 论文标题[]中最后的数字表示截止2019.1.21 google被引次数）
1. Domain Adaptation:
challenges:
(a) data shifts (syn -> live user data; stale -> current) cause distribution mismatch bet train and eval. -> 2017.1
(b) reestimate a global model from scratch each time a new domain with potentially new intents and slots is added. -> 2017.4
papers:
2017.1 adversarial training
[Adversarial Adaptation of Synthetic or Stale Data. Young-Bum Kim. 14]
2017.4 model(k + 1) = weighted_combination[model(1),...,model(k)]
[Domain Attention with an Ensemble of Experts. Young-Bum Kim. 17]
2. NLG:
challenges:
(a) integrate LM + Affect. -> 2017.2
(b) refering expression misunderstand -> 2017.5
(c) neural encoder-decoder models in open-domain: generate dull and generic responses. -> 2017.8
(d) multi-turn: lose relationships among utterances or important contextual information. -> 2017.11
(e) automatically evaluating the quality of dialogue responses for unstructured domains: biased and correlate very poorly with human judgements of response quality. -> 2017.12
(f) deep latent variable models used in open-domain: highly randomized, leading to uncontrollable generated responses. -> 2017.14
(g) does not employ knowledge to guide the generation -> tends to generate short, general, and meaningless responses. -> 2018.L1
(h) encoder-decoder dialog model is limited because it cannot output interpretable actions as in traditional systems, which hinders(阻碍) humans from understanding its generation process. -> 2018.L6
(i) translate natural language questions ->structured queries: further improvement hard. -> 2018.L8
papers:
2017.2 language model + affect info
[Affect-LM: A Neural Language Model for Customizable Affective Text Generation. 27]
2017.5 refering expression misunderstand correction - alg: contrastive focus
[Generating Contrastive Referring Expressions. 0]
2017.8 open-domain - Framework - conditional variaional autoencoders
pre: word-level decoder
cur: discourse-level encoder
[Learning Discourse-level Diversity for Neural Dialog Models using Conditional Variational Autoencoders. CMU. 69]
2017.11 muli-turn response selection - sequential matching network (SMN)
pre: concatenates utterances in context
matches a response with a highly abstract context vector
=> lose relationships among utterances or important contextual information
current: matches a response with each utterance on multiple levels of granularity
distills important matching information -> vector -> conv + pooling
accumulate vector -> RNN (models relationships among utterances)
final matching score (calcu with hid of rnn)
[Sequential Matching Network: A New Architecture for Multi-turn Response Selection in Retrieval-based Chatbots. 北航. 南开.微软. 48]
2017.12 auto eval Metric - ADEM
[Towards an Automatic Turing Test: Learning to Evaluate Dialogue Responses. 47]
2017.14 Framework - generation based on specific attributes(manually + auto detected) - both speakers diag states modeled -> personal features
[A Conditional Variational Framework for Dialog Generation. 20]
2017.16 Open-domain - Engine - generation (info retrieval + Seq2Seq) - AliMe chat
[AliMe Chat: A Sequence to Sequence and Rerank based Chatbot Engine. 27]
2018.L1 knowledge guide generation - neural knowledge diffusion (NKD) model - both fact + chi-chats
match the relevant facts for the input utterance + diffuse them to similar entities
[Knowledge Diffusion for Neural Dialogue Generation. 3]
2018.L6 encoder-decoder model - interprete- unsup discrete sent representation learning
DI-VAE + DI-VST - discover interpretable semantics via either auto encoding or context predicting
[Unsupervised Discrete Sentence Representation Learning for Interpretable Neural Dialog Generation. 8]
2018.L8 Framework - DialSQL + human intelligence
identify potential error of SQL -> ask for validation -> feedback to revise query
[ DialSQL: Dialogue Based Structured Query Generation. 4]
3. Task + Non-task hybrid
2017.3 whether to have a chat - dataset
[Chat Detection in Intelligent Assistant: Combining Task-oriented and Non-task-oriented Spoken Dialogue Systems. McGill University. Montreal. 7]
4. E2E
challenges:
(a) data-intensive -> 2017.6
(b) task - interact with KB -> pre: issuing a symbolic query to the KB to retrieve entries based on their attributes. -> 2017.13
disadvantages:
(1) such symbolic operations break the differentiability(可辨性) of the system
(2) prevent end-to-end training of neural dialogue agents
(c) only consider user semantic inputs and under-utilize other user info. -> 2018.L4
(b) incorporating knowledge bases. -> 2018.L7
papers:
2017.6 Framework - HCNs : RNN + knowledge(software/sys action templates) - reduce train data - opt (sup + RL) - bAbI dialog dataset - 2 commercial diag sys
[Hybrid Code Networks: practical and efficient end-to-end dialog control with supervised and reinforcement learning. Microsoft Research. 87]
2017.13 KB-InfoBot - E2E - task -multi-turn - interact with KB - present a agent
replacing symbolic queries -> induced "soft" posterior distribution over the KB
integrate soft retrival process + RL
[Towards End-to-End Reinforcement Learning of Dialogue Agents for Information Access. CMU. MS. 国立台北. 82]
2018.L4 multimodel info (sup + RL) - user adaptive - reduce diag length + improve success rate
[Sentiment Adaptive End-to-End Dialog Systems. 2]
2018.L7 Mem2Seq - first neural generative model: combines [ multi-hop attention over memories + idea of pointer network]
[Mem2Seq: Effectively Incorporating Knowledge Bases into End-to-End Task-Oriented Dialog Systems. UST.8]
5. NLU
challenges:
(a) no systematic comparison to analyze how to use context effectively. -> 2017.15
papers:
2017.7 identity discussion points + discourse relations
[Joint Modeling of Content and Discourse Relations in Dialogues. 7]
2017.15 context utiliztion eval -empirical study and compare models - variant: weights context vectors by context-query relevance
[How to Make Contexts More Useful? An Empirical Study to Context-Aware Neural Conversation Models. 18]
6. Dialogue state tracking
challenges:
(a) have difficulty scaling to larger, more complex dialogue domains. -> 2017.10
(1) Spoken Language Understanding models that require large amounts of annotated training data
(2) hand-crafted lexicons for capturing some of the linguistic variation in users' language.
(b) handling unknown slot values -> Pre: assume predefined candidate lists and thus are not designed to output unknown values. especially in E2E, SLU is absent. -> 2018.L10
papers:
2017.10 Framework - Neural Belief Tracking (NBT) - representation learning (compose pre-trained word vector -> utterances and context)
[Neural Belief Tracker: Data-Driven Dialogue State Tracking. 63]
2018.L9 Global-Locally Self-Attentive Dialogue State Tracker (GLAD)
global modules: shares parameters between estimators for different types (called slots) of dialogue states
local modules: learn slot-specific features
[Global-Locally Self-Attentive Encoder for Dialogue State Tracking. 0]
2018.L10 E2E + pointer nerwork (PtrNet)
[An End-to-end Approach for Handling Unknown Slot Values in Dialogue State Tracking. 2]
7. Framework
challenges:
(a) pipeline: introduces architectural complexity and fragility. -> 2018.L2
papers:
2018.L2 Seq2Seq + opt (sup / RL) - Task
design text spans named belief spans -> track dialogue believes -> allow task-oriented sys be modeled in Seq2Seq
Two Stage CopyNet instantiation -> reduce para, train time + better than pipeline on large dataset + OOV
[Sequicity: Simplifying Task-oriented Dialogue Systems with Single Sequence-to-Sequence Architectures. 新加坡国立. 复旦. 京东. 9]
8. RL
challenges:
(a) Training a task-completion dialogue agent via reinforcement learning (RL) is costly: requires many interactions with real users.
(b) use a user simulator: lacks the language complexity + biases
papers:
2018.L3 RL - policy learning - Deep Dyna-Q
first deep RL framework that integrates planning for task-completion dialogue policy learning
world model update with real user experience + agent opt using real and simulated experience
[Deep Dyna-Q: Integrating Planning for Task-Completion Dialogue Policy Learning. 3]
8. Chi-chat
challenges:
(a) lack specificity
(b) do not display a consistent personality. -> 2018.L5
papers:
2018.L5 add profile info[i. given + ii.partner] train to engage ii with personal topics -> used to predict profile
[Personalizing Dialogue Agents: I have a dog, do you have pets too? 31.]
9. Others:
challenges:
(a) open-ended dialogue state. -> 2017.9
papers:
2017.9 Symmetric Collaborative Dialogue - two agents to achieve a common goal
[Learning Symmetric Collaborative Dialogue Agents with Dynamic Knowledge Graph Embeddings. 21]
