AirDialogue: An Environment for Goal-Oriented Dialogue Research

构造了一个比较有效的数据集

Deep Learning for Dialogue Systems

Deep Learning for Dialogue Systems

BERT for Joint Intent Classification and Slot Filling

Supervised Semantic Indexing

TOWARDS AI-COMPLETE QUESTION ANSWERING A SET OF PREREQUISITE TOY TASKS

Enhanced LSTM for Natural Language Inference

A retrieval-based dialogue system utilizing utteranceand context embeddings

EVALUATING PREREQUISITE QUALITIES FOR LEARNING END-TO-END DIALOG SYSTEMS

本文的目标不是来提升在狭窄的酒店预定领域的效果，而是用一个传统的手工系统就有较好系统来对比没有领域知识的end-to-end系统的优劣。

• https://research.fb.com/downloads/babi/

MEMORYNETWORKS

• https://arxiv.org/pdf/1503.08895.pdf

最终结果说明，在给出的任务中基于规则的毫无疑问效果比模型的好，但是对于在真实场景的真实问题来说，MemNN效果更好

和现在的架构很像

先用word匹配和正则等制定一个规则系统来作为baseline.

LEARNING END-TO-END GOAL-ORIENTED DIALOG

ParlAI: A Dialog Research Software Platform

A Network-based End-to-End Trainable Task-oriented Dialogue System

• https://github.com/shawnwun/NNDIAL

这个end-to-end的系统，在意图识别的阶段用的是cnn+LSTM 在状态管理(belief state tracking)也用的LSTM，在policy的时候自定义了一套算法，将前面的几个输出向量做了个线性模型，输出。

策略生成是用前面的特征向量加和乘积

造出来的一个belief state的概念：

由intent网络生成的分布式表示和对slot-value组的概率表示叫做belief stat。

Learning Task-Oriented Dialog with Neural Network Methods

An End-to-End Trainable Neural Network Model withBelief Tracking for Task-Oriented Dialog

模型训练思路：

• 1 首先离线有监督学习 人和人的对话数据
• 2 然后让模型和人交互，基于反馈和模仿用强化学习来学习

为了解决样本效率问题，提出了learning-from-user and learning-from-simulationl两个方案。

基于神经网络的端到端的网络模型能够健壮的跟踪对话状态，和知识库交互，用结构化的信息来成功的完成任务驱动型对话。

End-to-End Learning of Task-Oriented Dialogs

端到端的task类型对话的鼻祖

传统pipeline方案的问题点： 1 流程比较复杂，每步骤独立训练，但是流程输入和输出有依赖，错误放大，难以跟进。

Dialogue Learning with Human Teaching and Feedback in End-to-End Trainable Task-Oriented Dialogue Systems

一个混合学习过程，在人类的指导教育和反馈下增强强化学习的过程

Attention-Based Recurrent Neural Network Models for Joint Intent Detection and Slot Filling

用一个模型来解决两个不同类型的问题，intent detect是分类，填槽是序列标注。都用基于attention机制的RNN来搞定了

本文为了对intent和slot-filling联合建模，额外添加了一个decoder来进行意图检测。

本文中用到的attention-RNN算法。

D. Bahdanau, K. Cho, and Y. Bengio, “Neural machine trans-lation by jointly learning to align and translate,”arXiv preprintarXiv:1409.0473, 2014

整体来说，无法验证，没有任何实质的创新点。

说是构建了一个第一个中文电商机器人对话系统(really?)

M = (I, C, A)

I是intent，C是product category， A是商品attribute。 M是根据用户Query得到的信息的表示。

意图分类：PhraseLDA 1000个topic

产品分类： a CNN-based approach that resembles (Huang et al. 2013)and (Shen et al. 2014

在购物场景中主要的行为有：

• Recommendation
• Comparison
• Opinion Summary
• Question Answering
• Proactive Questioning
• Chit-chat

为了解决所谓的问题：

• 1 结构化商品信息
• 2 用户的搜索日志
• 3 社区网站，挖掘购买意图和产品相关的词

The Sogou Spoken Language Understanding System for the NLPCC 2018 Evaluation

SLU 模型做法：

• 1 第一步是词汇分析，也就是分词，然后词性标注。本文用的是HanLP做词性分析。

• 2 第二步是槽位边界检测。这个任务看成一个用BILOU进行序列标注的。我们用了基于字和词的序列标注。基于字的 版本是用一个window为7的CRF，用此法特征和词典特征，另外基于词的的CRF模型是window size为5的词法特征，词性特征和词典特征。词典特征是指“当前字词是否 prefix/infix/suffix 在实体词典中某个条目关系。”每个CRF输出n(3)个输出，这整个2n个输出用到下一步。用基于字的序列标注是为了弥补分词效果差带来的可能影响。

• 3 第三部是槽位类型识别。用的是LR+L正则分类器，预测出的slot，上下文的字词，上下文的词性标注作为特征。

• 4 第四步是槽位纠正。这个是为了解决因为ASR导致的错误识别造成的结果。用的是一个基于搜索的方法。鉴于已经有各种槽位类型的词典，如果一个预测出来的槽位s类型T没有在对应的槽位词典中，那么就用s作为查询词来在根据最小编辑距离来查询槽位词典中的记录。这个操作会进行两次，一个是s作为中文字符，另一个是s作为拼音来查询。最好的结果是从这两个查处的结果中重新排序后得到的。

• 5 最后一步是意图分类。用的是XGBoost及其默认参数。用到的特征是单词token，query length，以及前面步骤预测出来的槽位。

规则的具体形式是,"如果query q被列在了一个特定的词汇表L，并且其前面的queries和它们预测出来的领域标签满足特定条件，那么q就可以被打上一个特定的意图标签，并且对于短的命令来说，整个query q是当作对应于L的一个槽位类型".所有规则按照优先级顺序组织的。

一个query首先经过基于规则的无明显意图词的判定过程，如果是空的话那就意味含有IISPs基于模型的组件会继续来处理，否则的话也就意味着query被看作是只有实体的，那么规则的结果就作为最终结果直接返回。

可以把预料文本分成2类，一类是有明显的预示意图的词语，另一类是没有。

基于检索的是从候选的回复中选出一个。检索式的关键是message-response的匹配。

B. Hu, Z. Lu, H. Li, and Q. Chen. Convolutional neu-ral network architectures for matching natural lan-guage sentences. InAdvances in neural informationprocessing systems, pages 2042–2050, 2014.

单轮的匹配 match(X,Y) = X^TAy

X：message的向量表示， y：回复的向量表示。

H. Wang, Z. Lu, H. Li, and E. Chen. A dataset for re-search on short-text conversations. InProceedings ofthe 2013 Conference on Empirical Methods in NaturalLanguage Processing, pages 935–945, Seattle, Wash-ington, USA, October 2013. Association for Compu-tational Linguistics

Z. Lu and H. Li. A deep architecture for matchingshort texts. InInternational Conference on Neural In-formation Processing Systems, pages 1367–1375, 2013.

B. Hu, Z. Lu, H. Li, and Q. Chen. Convolutional neu-ral network architectures for matching natural lan-guage sentences. InAdvances in neural informationprocessing systems, pages 2042–2050, 2014

M. Wang, Z. Lu, H. Li, and Q. Liu. Syntax-based deepmatching of short texts.InIJCAI, 03 2015

Y. Wu, W. Wu, Z. Li, and M. Zhou. Topic augmentedneural network for short text conversation.CoRR,2016

多轮匹配

任务型对话系统整体来说可以分为两类：

• 1 pipeline，也就是包含SLU+DST+PL+NLG
• 2 end-to-end

在传统的task-oriented对话系统中，尽管有很多特定领域的人工定制，很难推广其他领域，更进一步的是pipeline的方法有两个局限。

• 一个是信用分配问题，一个用户的反馈很难传播到上游每个组件中。
• 另一个是问题流程的相互依赖。一个组件的输入依赖上一个组件的输出。一部分变动其他都得动。（这个真的是问题么？）

这俩文章介绍来一种基于网络的end-to-end的可训练的task-oriented对话系统，方法是把对话系统学习看成从对话历史到回复响应的mapping，并用encoder-decoder模型来训练整个模型。不过这个系统是以有监督的方式训练的，不仅需要大量的训练数据，并且由于在训练数据中缺乏对对话控制的探索也不能找到一个鲁棒的好策略。

• A network-based end-to-end trainable task-oriented di-alogue system
• Learningend-to-end goal-oriented dialog.

下文中，首次提出了一个联合训练dialogue state tracking和policy learning来优化得到更鲁棒的系统行为。

• Towards end-to-end learn-ing for dialog state tracking and management us-ing deep reinforcement learning

task-oriented系统经常需要query外部知识库，前面的系统是通过发出一个符号请求到知识库基于属性来获得条目。

一个典型的pipeline方法构建的task-oriented对话系统包含四部分：

• Language understanding.NLU/SLU,目标是解析理解用户输入为intent，slot

• Dialogue state tracker. 根据当前对话输入信息结合历史信息给出当前会话状态。

• Dialogue policy learning.基于当前对话状态给出接下来要采取的行动

• Natural language generation(NLG). 将映射的选择的动作行为转换生成对应的输出回复。

策略学习 基于前面state tracker的状态表示，策略学习(policy learning)是来生成下一个可用的系统行动。无论是监督学习或者强化学习都可以用来优化策略学习。 H. Cuayhuitl, S. Keizer, and O. Lemon. Strategic di-alogue management via deep reinforcement learning.arxiv.org, 2015.

通常都用一个基于规则的agent来初始化系统。 Z. Yan, N. Duan, P. Chen, M. Zhou, J. Zhou, andZ. Li. Building task-oriented dialogue systems for on-line shopping. InAAAI Conference on Artificial Intel-ligence, 2017

然后用监督学习来基于规则生成的规则来学习。Building task-oriented dialogue systems for on-line shopping. 强化学习，Strategic di-alogue management via deep reinforcement learning.结果据说比很多系统，rule based，superviesed都好

状态管理。

统计对话系统维护了一个对真实状态基于多重假设来描述的分布，以应对噪声场景和歧义。

• S. Young, M. Gai, S. Keizer, F. Mairesse, J. Schatz-mann, B. Thomson, and K. Yu. The hidden informa-tion state model: A practical framework for pomdp-based spoken dialogue management. 在DSTC比赛中结果形式是每轮对话中每个slot的一个概率分布。各种统计学方法如下：
• 规则集， Z. Wang and O. Lemon. A simple and generic belieftracking mechanism for the dialog state tracking chal-lenge: On the believability of observed information. InSIGDIAL Conference, pages 423–432, 2013

• CRF S. Lee and M. Eskenazi. Recipe for building robustspoken dialog state trackers: Dialog state trackingchallenge system description. InSIGDIAL Conference,pages 414–422, 2013

  S. Lee. Structured discriminative model for dialogstate tracking. InSIGDIAL Conference, pages 442–451, 2013

H. Ren, W. Xu, Y. Zhang, and Y. Yan. Dialog statetracking using conditional random fields. InSIGDIALConference, pages 457–461, 2013.

• maximum entropy model J. Williams. Multi-domain learning and generaliza-tion in dialog state tracking. InSIGDIAL Conference,pages 433–441, 2013.

• web-style ranking J. D. Williams. Web-style ranking and slu combina-tion for dialog state tracking

深度学习的状态管理。用一个滑动窗口来在任意数量可能值上输出一个概率序列。 M. Henderson, B. Thomson, and S. Young. Deep neu-ral network approach for the dialog state tracking chal-lenge. InProceedings of the SIGDIAL 2013 Confer-ence, pages 467–471, 2013

多领域的RNN状态跟进模型： B. Thomson, M. Gasic, P.-H. Su, D. Vandyke, T.-H. Wen, and S. Young. Multi-domain dialog state tracking using recurrent neuralnetworks.

基于neural belief tracker(NBT)来检测slot-value对。 Neural belief tracker: Data-driven dia-logue state tracking.

跟进对话状态是保障dialog system的robust的核心。主要目标是预测每轮对话的用户目标。经典的状态结构通常叫做slot-filling 或者 sematic frame.

传统用手工规则的方法： D. Goddeau, H. Meng, J. Polifroni, S. Seneff, andS. Busayapongchai. A form-based dialogue managerfor spoken language applications. InSpoken Language,1996. ICSLP 96. Proceedings., Fourth InternationalConference on, volume 2, pages 701–704. IEEE, 1996

基于规则的方法倾向于常见的错误，然后很多结果并不是想要的。 J. D. Williams. Web-style ranking and slu combina-tion for dialog state tracking. InSIGDIAL Conference,pages 282–291, 2014

填槽这个问题更多的是看成一个序列标注的问题。句子中的每个词都打上一个语义标签。输入是由词组成的句子，输出是每个词对应的slot/concept IDs.

DBN 类的处理：

• A Deoras and R. Sarikaya. Deep belief network basedsemantic taggers for spoken language understanding.

  L. Deng, G. Tur, X. He, and D. Hakkani-Tur. Use ofkernel deep convex networks and end-to-end learningfor spoken language understanding

RNN：

• G. Mesnil, X. He, L. Deng, and Y. Bengio. Investi-gation of recurrent-neural-network architectures andlearning methods for spoken language understanding.Interspeech, 2013.
• K. Yao, G. Zweig, M. Y. Hwang, Y. Shi, and D. Yu.Recurrent neural networks for language understand-ing. InInterspeech, 2013
• R. Sarikaya, G. E. Hinton, and B. Ramabhadran.Deep belief nets for natural language call-routing
• K. Yao, B. Peng, Y. Zhang, D. Yu, G. Zweig, andY. Shi. Spoken language understanding using longshort-term memory neural networks. InIEEE Insti-tute of Electrical & Electronics Engineers, pages 189 –194, 2014

目标是根据一个用户utterance/query 得到其对应的语义slot。slots是预先根据场景定于的。通常来说有两种类型的表示，一个是句子级别的类别，例如用户的意图和utterance的类别。另外一个是单词级别的信息抽取，例如命名实体和槽位填充。

意图识别是根据一句话来检测用户的意图。 基于深度学习的意图识别： L. Deng, G. Tur, X. He, and D. Hakkani-Tur. Use ofkernel deep convex networks and end-to-end learningfor spoken language understanding. InSpoken Lan-guage Technology Workshop (SLT), 2012 IEEE, pages210–215. IEEE, 2012

G. Tur, L. Deng, D. Hakkani-T ̈ur, and X. He. Towardsdeeper understanding: Deep convex networks for se-mantic utterance classification. InAcoustics, Speechand Signal Processing (ICASSP), 2012 IEEE Interna-tional Conference on, pages 5045–5048. IEEE, 2012.

D. Yann, G. Tur, D. Hakkani-Tur, and L. Heck. Zero-shot learning and clustering for semantic utteranceclassification using deep learning. 2014.

尤其是这个用CNN来抽取query vector进行query分类。 H. B. Hashemi, A. Asiaee, and R. Kraft. Query intentdetection using convolutional neural networks. InIn-ternational Conference on Web Search and Data Min-ing, Workshop on Query Understanding, 2016

P.-S. Huang, X. He, J. Gao, L. Deng, A. Acero, andL. Heck. Learning deep structured semantic modelsfor web search using clickthrough data. InProceedingsof the 22nd ACM international conference on Confer-ence on information & knowledge management, pages2333–2338. ACM, 2013

Y. Shen, X. He, J. Gao, L. Deng, and G. Mesnil.Learning semantic representations using convolutionalneural networks for web search. InProceedings of the23rd International Conference on World Wide Web,pages 373–374. ACM, 2014.

SLU有俩事，意图识别+填槽。

实践中的困难：

• 1 意图分类的复杂性
• 2 世界知识
• 3 用户状态

任务型人机对话系统中的认知技术

用一个带有copy机制的简单seq2seq框架超过现有最好的真实DSTC2 7个点。

A Copy-Augmented Sequence-to-Sequence Architecture GivesGood Performance on Task-Oriented Dialogue

Task-oriented dialogue focuses on conversational agents that participate in dialogues with user goals on domain-specific topics. In contrast to chatbots, which simply seek to sustain open-ended meaningful discourse, existing task-oriented agents usually explicitly model user intent and belief states. This paper examines bypassing such an explicit representation by depending on a latent neural embedding of state and learning selective attention to dialogue history together with copying to incorporate relevant prior context. We complement recent work by showing the effectiveness of simple sequence-to-sequence neural architectures with a copy mechanism. Our model outperforms more complex memory-augmented models by 7% in per-response generation and is on par with the current state-of-the-art on DSTC2, a real-world task-oriented dialogue datase

这个地方的NLU和NLG都是用基于模版的模型。

症状数据抽取，BIO格式。每个中文字符标注为“B","I","O".每个抽取出的症状根据病人真实情况打标为“True","False"。3人2个都标过的才有效，第三人评判。Cohhen kappa 相关性来作为标注标准。

In this paper, we make a move to builda dialogue system for automatic diagno-sis. We first build a dataset collected froman online medical forum by extractingsymptoms from both patients’ self-reportsand conversational data between patientsand doctors. Then we propose a task-oriented dialogue system framework tomake the diagnosis for patients automat-ically, which can converse with patients tocollect additional symptoms beyond theirself-reports. Experimental results on ourdataset show that additional symptoms ex-tracted from conversation can greatly im-prove the accuracy for disease identifica-tion and our dialogue system is able tocollect these symptoms automatically andmake a better diagnosis

从一个在线医疗论坛抽取来病人的病情自述以及和医生的对话过程作为训练数据，结果表明从对话过程获得的病情描述能大幅提高医生对疾病的诊断，并且论文的对话系统能够有效的收集到这些信息帮助诊断。

在生物医药领域很缺数据，为了解决这个问题，常识了基于模版，术语大的数据扩展技术。先在大的数据集上用ELMo来构建词向量。把任务评估分成两个子任务来进行，slot-filling和意图分类。

偏应用的一篇文章，结果也说明不了什么

Natural language understanding for task oriented dialog in the biomedical domain in a low resources context

PyDial: A Multi-domain Statistical Dialogue System Toolkit

一个开源的端到端的统计对话系统工具。

• http://pydial.org
• https://github.com/claycollier/pydial
• https://bitbucket.org/dialoguesystems/pydial.git

其总的架构包含Sematic Decode,Belief Tracker,Policy Reply System,Language generator. 整体来说整个系统都支持了基于规则的判断过程，也融合了模型的支持。源码值得一看的。

对于填槽任务，输入是一个有一系列词组成的语句，输出是每个词对应的slot/concept IDs。在统计SLU系统里，这个任务可以看作是：给定词序列L,SLU的目标是找到一个slot 序列来最大化后验概率P(S/L).

Using Recurrent Neural Networks for Slot Filling in Spoken Language Understanding

双向的 Jordan-type网络对槽最好用

Using Recurrent Neural Networksfor Slot Filling in Spoken Language Understanding

非常简单的介绍，可以看作是一个文献阅读索引

detail :

• http://opendialogue.miulab.tw

• https://www.csie.ntu.edu.tw/~yvchen/doc/OpenDialogue_Tutorial_ISCSLP18.pdf

• https://www.csie.ntu.edu.tw/~yvchen/doc/OpenDialogue_Tutorial_IJCNLP.pdf

Open-Domain Neural Dialogue Systems

Open-Domain Neural Dialogue Systems

Open-Domain Neural Dialogue Systems

用到的数据集是一个coffee ordering的对话过程的数据，31567通对话，142412对会话。数据只有用正则匹配出来的填充的标签信息。

agent model 构建前先预训练。网络结构和user model一样，但是输入和输出反转，同时也把之前对话中已经填充的槽位信息作为输入。但是这俩部分信息并不是简单的直接拼接在一起，而是来学习适合的attention 权重来更好的利用注意力机制。此外任何其他额外的语义意图标签都不必用。

User model.直接用双向的LSTM，以agent的utterance作为X，对应的用户的utterance作为Y。

这个认识是说在通常场景中用户先表达出意图然后回答agent的一个个问题来具体自己的需求。

用户通常是被动的，偶尔的有一轮问题。换句话说用户基本上都是在一轮中回答由agent提出的问题。所以可以基于一个用户只需要考虑一轮回答来给出回复这样的假设来构建user model,,让agent model来处理多轮对话。

吐槽现有机器人比较low都是手工规则，强化学习只适用有限的几个场景。所以受用户和agent角色的不对称关系造了samia。首先是用户模型不是规则或者排序而是seq2seq，然后基于用户模型来用强化学习构建agent。

ntegrating User and Agent Models: A Deep Task-Oriented Dialogue System Weiyan Wang, Yuxiang WU, Yu Zhang, Zhongqi Lu, Kaixiang Mo, Qiang Yang (Submitted on 10 Nov 2017) Task-oriented dialogue systems can efficiently serve a large number of customers and relieve people from tedious works. However, existing task-oriented dialogue systems depend on handcrafted actions and states or extra semantic labels, which sometimes degrades user experience despite the intensive human intervention. Moreover, current user simulators have limited expressive ability so that deep reinforcement Seq2Seq models have to rely on selfplay and only work in some special cases. To address those problems, we propose a uSer and Agent Model IntegrAtion (SAMIA) framework inspired by an observation that the roles of the user and agent models are asymmetric. Firstly, this SAMIA framework model the user model as a Seq2Seq learning problem instead of ranking or designing rules. Then the built user model is used as a leverage to train the agent model by deep reinforcement learning. In the test phase, the output of the agent model is filtered by the user model to enhance the stability and robustness. Experiments on a real-world coffee ordering dataset verify the effectiveness of the proposed SAMIA framework.

Deep Reinforcement Learning for Dialogue Generation

Recent neural models of dialogue generationoffer great promise for generating responsesfor conversational agents, but tend to be short-sighted, predicting utterances one at a timewhile ignoring their influence on future out-comes. Modeling the future direction of a di-alogue is crucial to generating coherent, inter-esting dialogues, a need which led traditionalNLP models of dialogue to draw on reinforce-ment learning. In this paper, we show how tointegrate these goals, applying deep reinforce-ment learning to model future reward in chat-bot dialogue. The model simulates dialoguesbetween two virtual agents, using policy gradi-ent methods to reward sequences that displaythree useful conversational properties: infor-mativity, coherence, and ease of answering (re-lated to forward-looking function). We evalu-ate our model on diversity, length as well aswith human judges, showing that the proposedalgorithm generates more interactive responsesand manages to foster a more sustained conver-sation in dialogue simulation. This work marksa first step towards learning a neural conversa-tional model based on the long-term success ofdialogues.

Dialog System & Technology Challenge 6 Overview of Track 1 - End-to-End Goal-Oriented Dialog learning

End-to-end dialog learning is an important research subject inthe domain of conversational systems. The primary task consistsin learning a dialog policy from transactional dialogs of a givendomain. In this context, usable datasets are needed to evaluatelearning approaches, yet remain scarce. For this challenge, atransaction dialog dataset has been produced using a dialogsimulation framework developed and released by Facebook AIResearch. Overall, nine teams participated in the challenge. Inthis report, we describe the task and the dataset. Then, we specifythe evaluation metrics for the challenge. Finally, the results ofthe submitted runs of the participants are detailed.

A COMPARATIVE STUDY OF RECURRENT NEURAL NETWORK MODELS FOR LEXICAL DOMAIN CLASSIFICATION

Intent Detection for code-mix utterances in task oriented dialogue systems

Intent detection is an essential component of taskoriented dialogue systems. Over the years, extensiveresearch has been conducted resultingin many state of the art modelsdirected towards resolving user’sintents indialogue. A variety of vector representation for user utterances have been explored for the same. However, these models and vectorization approaches have more so been evaluated in a single language environment. Dialogue systems generally have to deal with queries in different languages and most importantly Code-Mix form of writing. Since Code-Mix texts are not bounded by a formal structure they are difficult to handle. We thus conduct experiments across combinations of models and various vector representations for Code-Mix as well as multi-language utterancesand evaluate how these models scale to a multi-languageenvironment. Our aim is to find the best suitable combination of vector representation and models for the process of intent detection for code-mix utterances. We have evaluated the experiments on two different dataset consisting of only Code-Mix utterances and the otherdataset consisting of English, Hindi, and Code-Mix( English-Hindi) utterances

Sequence-to-Sequence Learning for Task-oriented Dialogue with Dialogue State Representation

Mem2Seq: Effectively Incorporating Knowledge Bases into End-to-End Task-Oriented Dialog Systems

Improving Semantic Parsing for Task Oriented Dialog

Semantic parsing using hierarchical representations has recently been proposedfor task oriented dialog with promising results. In this paper, we present three dif-ferent improvements to the model: contextualized embeddings, ensembling, andpairwise re-ranking based on a language model. We taxonomize the errors pos-sible for the hierarchical representation, such as wrong top intent, missing spansor split spans, and show that the three approaches correct different kinds of errors.The best model combines the three techniques and gives 6.4% better exact matchaccuracy than the state-of-the-art, with an error reduction of 33%, resulting in anew state-of-the-art result on the Task Oriented Parsing (TOP) dataset

对话系统任务综述与基于POMDP的对话系统