Source-linked AI summary
A Survey on Dialogue Systems: Recent Advances and New Frontiers
Hongshen Chen, Xiaorui Liu, Dawei Yin, Jiliang Tang
TL;DR
Dialogue-system research seeks capable human-computer conversation while reducing reliance on handcrafting and addressing limitations of conventional task-oriented pipelines. This survey organizes task-oriented and non-task-oriented systems, reviews deep-learning methods and representative algorithms, and discusses research directions. It concludes that deep learning has advanced both system types, while response evaluation and deeper understanding remain open challenges.
Problem
Dialogue systems need to support capable human-computer conversation, but conventional task-oriented systems rely on domain-specific handcrafting and non-task-oriented response quality remains difficult to evaluate automatically.
Method
The survey divides dialogue systems into task-oriented and non-task-oriented models, reviews deep-learning methods and representative algorithms, and discusses research directions.
Results
Deep learning has advanced both task-oriented and non-task-oriented dialogue systems by leveraging large amounts of data for representations and response-generation strategies with limited handcrafting.
Takeaways & Limitations
The survey identifies deep understanding, use of unstructured knowledge, and improved response evaluation as directions for further dialogue-system research.
Takeaways & Limitations
Current neural dialogue systems rely heavily on annotated data, structured knowledge bases, and conversation data, while generated responses can lack diversity or meaningfulness.
Abstract
from arXiv · showhide
Dialogue systems have attracted more and more attention. Recent advances on dialogue systems are overwhelmingly contributed by deep learning techniques, which have been employed to enhance a wide range of big data applications such as computer vision, natural language processing, and recommender systems. For dialogue systems, deep learning can leverage a massive amount of data to learn meaningful feature representations and response generation strategies, while requiring a minimum amount of hand-crafting. In this article, we give an overview to these recent advances on dialogue systems from various perspectives and discuss some possible research directions. In particular, we generally divide existing dialogue systems into task-oriented and non-task-oriented models, then detail how deep learning techniques help them with representative algorithms and finally discuss some appealing research directions that can bring the dialogue system research into a new frontier.
1. INTRODUCTION
Dialogue systems have gained attention as big data and deep learning make data-driven human-computer conversation increasingly feasible. The survey reviews recent deep-learning advances and research directions across task-oriented and non-task-oriented systems.
- Motivation: Big data provides conversational material for learning responses in data-driven, open-domain human-computer dialogue systems.Deep learning is described as effective for capturing complex patterns in large datasets.
- Dialogue-system scope: Task-oriented systems assist with activities such as finding products or booking accommodations and restaurants.Traditional systems process messages through understanding, state representation, policy actions, and response realization.
- Survey scope: The survey covers recent deep-learning advances in dialogue systems and discusses possible research directions.Its organization includes task-oriented pipeline and end-to-end methods, neural generative and retrieval-based methods, and concluding research directions.
2. TASK-ORIENTED DIALOGUESYSTEMS
Task-oriented dialogue systems are reviewed as pipelines whose components interpret utterances, track dialogue state, choose actions, and generate responses. The survey describes deep-learning applications across language understanding and state tracking alongside supervised or reinforcement-learning policy methods.
- Section scope: The survey reviews pipeline and end-to-end methods for task-oriented dialogue systems.It presents task-oriented dialogue as an important branch of spoken dialogue systems.
- Pipeline methods: A traditional task-oriented pipeline contains language understanding, dialogue state tracking, policy learning, and natural language generation.The components respectively parse utterances, maintain state, select actions, and render responses.
- Language understanding: Natural language understanding maps utterances into scenario-specific semantic slots, including intent, domain, and word-level labels.Intent detection classifies utterances into predefined intents, while slot filling assigns semantic labels to words.
- Dialogue state tracking: Dialogue state tracking represents the dialogue history and estimates the user’s goal at each turn.Statistical systems maintain distributions over multiple state hypotheses to handle noise and ambiguity.
- Policy learning: Policy learning generates the next system action from the tracked state using supervised learning or reinforcement learning.A rule-based agent may warm-start the system before supervised or end-to-end reinforcement-learning optimization.
2.2 End-to-End Methods
The conventional task-oriented pipeline is difficult to adapt because feedback cannot readily reach upstream modules and component dependencies complicate coordinated changes. End-to-end neural methods are presented as a response to these limitations.
- Limitations: Domain-specific handcrafting makes traditional task-oriented dialogue systems difficult to adapt to new domains.This is identified as one limitation of conventional pipeline systems.
- Limitations: User feedback is hard to propagate to each upstream module, creating a credit assignment problem in conventional pipelines.The pipeline’s modular structure limits how end-user feedback reaches earlier components.
- Limitations: Process interdependence requires other components to be adapted when one component or its data changes.Changing slots or features can require substantial human effort to maintain global optimization.
3. NON-TASK-ORIENTED DIALOGUESYS-TEM
Non-task-oriented dialogue systems, or chatbots, converse in open domains using generative or retrieval-based methods. Recent neural approaches use encoder-decoder models, attention, dialogue context, and latent variables to improve response generation.
- Non-task-oriented dialogue systems, also called chatbots, focus on open-domain conversation rather than completing specific user tasks.
- Chatbots generally use generative models to create novel responses or retrieval-based models to select fluent, informative responses from a repository.
- 3.1.1 Sequence-to-Sequence Models: Seq2Seq models encode a message into a context vector with an RNN, then decode that vector to estimate the response-generation probability.
- 3.1.1 Sequence-to-Sequence Models: Attention conditions each response word on a different weighted context vector because words may relate to different parts of the message.
- 3.1.1 Sequence-to-Sequence Models: Neural encoder-decoder models learn dialogue representations and response-generation strategies from large datasets while requiring little domain knowledge or handcrafting.
- 3.1.2 Dialogue Context: Dialogue-context studies report that hierarchical RNNs generally outperform non-hierarchical models, while context produces longer, more meaningful, and diverse replies.
- 3.1.3 Latent Variable Models: Latent-variable models address trivial, noncommittal responses by sampling latent assignments before deterministic decoding and modeling factors such as topic or sentiment.
3.1.5 Outside Knowledge Base
Outside knowledge bases help dialogue systems address gaps between system and human background knowledge, while dialogue research also develops retrieval, generation, hybrid response selection, and evaluation methods.
- Outside Knowledge Base: Knowledge bases are used to bridge gaps between a dialogue system’s background knowledge and human knowledge.Memory networks provide a classic approach for question answering with knowledge bases.
- Outside Knowledge Base: Knowledge-grounded generation can combine knowledge-base words with common words instead of relying only on tuple retrieval.Empirical studies found that this approach generated natural and correct answers by referring to knowledge-base facts.
- Evaluation: Automatic evaluation remains difficult for non-task-oriented dialogue because response diversity weakens the relationship between common metrics and human judgments.BLEU, METEOR, ROUGE, and some embedding metrics show weak or no correlation with human judgments, despite distinguishing model baselines in some settings.
- Retrieval-Based Methods: Retrieval-based systems select candidate responses by matching messages and responses, but must overcome semantic gaps between them.Single-turn models typically encode the message and response as vectors before computing a matching score.
- Retrieval-Based Methods: Multi-turn retrieval models incorporate the current message and previous utterances to select responses relevant to the whole conversational context.Researchers use context and response representations, utterance selection, and word- or utterance-level matching.
- Hybrid Models: Hybrid systems combine retrieval and generation because retrieved responses are precise but blunt, whereas generated responses are fluent but can be meaningless.One ensemble feeds a retrieved candidate and the original message to an RNN generator before post-reranking the final response.
4. DISCUSSION AND CONCLUSION
Deep learning has become foundational across task-oriented and non-task-oriented dialogue systems, but current end-to-end models remain imperfect. The paper identifies faster domain warm-up, deeper understanding, and privacy protection as major research directions.
- Discussion and Conclusion: Deep learning now supports components of traditional task-oriented systems and end-to-end frameworks for both task-oriented and chit-chat dialogue.These developments blur the boundary between the two system categories, while current end-to-end models remain far from perfect.
- Swift Warm-Up: Practical dialogue engineering still relies on traditional pipelines during new-domain warm-up because domain-specific dialogue data is limited and costly to construct.The paper calls for methods that bridge this warm-up stage.
- Swift Warm-Up: Dialogue agents should learn from interaction to reduce dependence on limited domain-specific data during system construction.The paper identifies self-directed learning through human interactions as a promising direction.
- Deep Understanding: Current neural dialogue systems depend heavily on annotated data, structured knowledge bases, and conversation data, while their responses can lack diversity and meaning.Future agents should learn from instruction, use unstructured knowledge, make inferences, and share knowledge across domains.
- Privacy Protection: Dialogue assistants that learn through interaction, comprehension, and inference may inadvertently store sensitive information, making privacy protection necessary.The paper treats privacy protection as a requirement for building better widely used dialogue systems.