TL;DR
This paper provides a comprehensive systematic survey of recent advancements in deep learning-based dialogue systems, categorizing them by model and system type. It aims to equip researchers with a thorough understanding of the state-of-the-art, evaluation methods, and datasets, paving the way for future research in this rapidly evolving field.
What Problem Does This Paper Address?
Dialogue systems, or chatbots, are increasingly prevalent in various applications, from customer service to personal assistants. The field has seen a surge in deep learning-based approaches due to their superior performance. However, the rapid pace of research has led to a proliferation of novel works, making it challenging for researchers to stay abreast of the latest developments, understand the underlying principles of different models, and grasp the nuances of various system architectures. This paper addresses this challenge by offering a structured and comprehensive overview of the deep learning landscape in dialogue systems, analyzing them from multiple perspectives.
Methods
The paper systematically reviews and categorizes deep learning models commonly used in dialogue systems. These include:
- Neural Models: Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs) (including Jordan-type, Elman-type, LSTMs, GRUs), Vanilla Sequence-to-sequence Models (Encoder-Decoder), Hierarchical Recurrent Encoder-Decoder (HRED), Memory Networks, Attention Networks, Transformer models, Pointer Nets, CopyNets, Deep Reinforcement Learning (RL) models, Generative Adversarial Networks (GANs), and Knowledge Graph Augmented Neural Networks.
- System Types: Task-oriented dialogue systems (TOD) and open-domain dialogue systems (OOD).
- Evaluation Methods: Discusses metrics for both TOD (e.g., task completion rate, cost) and OOD (e.g., word-overlap metrics, neural metrics) systems.
- Datasets: Reviews datasets used for training and evaluation.
The survey emphasizes a multi-angle analysis, considering both the underlying neural architectures (model type) and the functional goals of the systems (system type).
Datasets
The paper discusses numerous datasets used in dialogue system research. Key datasets mentioned include:
- Task-oriented Dialogue Systems:
- MultiWOZ
- E2E
- MSR-E2E
- YELPNLG
- Clinical Conversation data set
- OOS
- ATIS
- SNIPS-NLU
- bAbI
- JDC
- Ubuntu V2
- MICROSOFT DIALOGUE CHALLENGE data set
- WOZ
- DSTC series
- SimDial
- SMD
- BANKING
- Weather forecast
- MedDialog-(EN,CN)
- CamRest
- Taskmaster
- Frames
- JDDC
- Court Debate Dataset
- TreeDST
- RiSAWOZ
- Cambridge Restaurant
- SB-TOP
- GSIM
- SGD
- cite-8K
- Open-domain Dialogue Systems:
- Large-Scale Corpus for Conversation Disentanglement
- DuConv
- PERSUASION FOR GOOD
- Mutual Friends
- SAMSum
- OpenDialKG
- doc2dial
- DialEdit
- CHART DIALOGS
- CONAN
- Dialogue NLI
- MuTual
- RST-DT
- NLPCC
- MELD
- EMPATHETIC DIALOGUES
- IEMOCAP
- EmoryNLP
- MojiTalk
- CBET
- Stanford Politeness Corpus
- AIT-2018
- EMOTYDA
- Wizard of Wikipedia
- CMU DOG
- Holl-E
- Interview
- Curiosity
- KdConv
- ELI5
- Topical Chat
- WHERE ARE YOU?
- MMD
- OpenSubtitles
- NTCIR
- Douban Conversation Corpus
- E-commerce Dialogue Corpus
- STC-SeFun
- DailyDialog
- PDTB
- Luna
- Edina-DR
- Cornell Movie Dialog Corpus
- Reddit Movie Dialogue Dataset
- LIGHT
- This American Life
- RadioTalk
- French EPAC
- TREC Conversational Assistance
- Search as a Conversation
- Amazon Alexa Prize Competition
- SwitchBoard
- SPOLIN
- CRD3
- Baidu Zhidao
- Webis Gmane Email Corpus 2019
- LibreSpeech Corpus
- Motivational Interviewing
- SubTle Corpus
- TED-LIUM
- ECG NLPCC 2017 Data
- SEMEVAL15
- AMAZONQA
- TGIF-QA
- QuAC
- SQUAD
- LIF
- Yelp
- Debates
- PERSONACHAT
- KvPI
- ConvA12
- PEC
- Guess What?!
- VisDial
- AVSD
- VFD
- PhotoBook
- IGC
- Image-Chat
The paper does not introduce any new datasets.
Key Insights and Ideas
- Deep Learning Dominance: Deep learning models, particularly those based on RNNs, LSTMs, Transformers, and attention mechanisms, have become the de facto standard for state-of-the-art dialogue systems due to their ability to capture complex sequential patterns and semantic nuances.
- Task-Oriented vs. Open-Domain Dichotomy: The survey clearly delineates the distinct challenges and approaches for task-oriented systems (focused on specific goals) and open-domain systems (aiming for general conversation), highlighting the trade-offs between control and flexibility.
- End-to-End vs. Modular Systems: The paper discusses the shift towards end-to-end trainable systems, which offer better optimization and interpretability compared to traditional modular pipelines, although challenges remain in areas like knowledge base integration.
- The Importance of Context: Context awareness is a critical factor for generating coherent and relevant responses, with advancements in hierarchical models and attention mechanisms playing a key role.
- Beyond Text: The survey touches upon multimodal dialogue systems, indicating a growing trend to incorporate visual and other sensory information for richer interactions.
- Data Scarcity and Augmentation: The paper implicitly highlights the ongoing challenge of data scarcity, particularly for specialized tasks, and the importance of techniques like transfer learning, meta-learning, and data augmentation.
Key Results
The paper is a survey and does not present novel experimental results. Its primary contribution is the comprehensive synthesis and categorization of existing research. It provides a structured overview of the field, enabling researchers to understand:
- The principles, characteristics, and applications of various deep learning models used in dialogue systems.
- The current research landscape and hot topics in both task-oriented and open-domain dialogue systems.
- The evolution of evaluation methods and the availability of datasets.
- Potential future research directions.
Limitations and Future Work
Limitations:
- The paper acknowledges that despite its comprehensiveness, it may still omit some relevant works.
- The rapid pace of research means that the field is constantly evolving, and the survey represents a snapshot in time.
Future Work:
The authors identify several promising research trends:
- Multimodal Dialogue Systems: Integrating vision, hearing, and other senses to create more human-like conversational agents.
- Multitask Dialogue Systems: Blending task-oriented and open-domain capabilities into unified systems or models.
- Corpus Exploration on the Internet: Developing methods to effectively access and utilize the vast amount of conversational data available online for training dialogue systems.
- User Modeling: Improving the ability of systems to understand and simulate user behavior for more coherent and personalized interactions.
- Dialogue Generation with Long-Term Goals: Moving beyond simple chit-chat to enable systems to pursue specific objectives over extended conversations, leading to more intelligent applications like negotiation or storytelling chatbots.
Summary updated