A Survey on Dialogue Systems: Recent Advances and New Frontiers

This is a reading note on dialogue systems, covering recent advances and new frontiers in the field.

Generative Approaches

Generative approaches are more proper for open-domain conversations, as they can generate responses that never appeared in training data.

Key papers:

• Neural Responding Machine for Short-Text Conversation
• A Neural Conversational Model

Dialogue Context

How to make context more useful? An empirical study on context-aware neural conversational models:

• How to make context more useful?

Response Diversity

PaperWeekly 第十八期 — 提高seq2seq方法所生成对话的流畅度和多样性

Objective Function:

• A Diversity-Promoting Objective Function for Neural Conversation Models
• An attentional neural conversation model with improved specificity

Beam Search:

• Batra. Diverse beam search: Decoding diverse solutions from neural sequence models
• Generating long and diverse responses with neural conversation models
• A simple, fast diverse decoding algorithm for neural generation

Rerank:

• A diversity-promoting objective function for neural conversation models
• A neural network approach to context-sensitive generation of conversational response

Latent Variable: Latent variable is designed to make high-level decisions like topic or sentiment.

Other Topics (TODO)

• Topics and personalities
• Outside knowledge base
• Interactive learning
• Evaluation (another deep learning model)

Retrieval Approaches

Retrieval-based approaches are informative and fluent.

• An information retrieval approach to short text conversation

Single-turn Response Matching

Multi-turn Response Matching

Hybrid Approaches

• A sequence to sequence and rerank based chatbot engine
• An ensemble of retrieval-and generation-based dialog systems

This survey was compiled in 2018 as a reading note on dialogue systems.