[Paper Review] Neural Approaches to Conversational AI

Jianfeng Gao, Michel Galley+1 more|arXiv (Cornell University)|Sep 21, 2018

Topic Modeling137 citations

TL;DR

This paper surveys neural methods for conversational AI across three categories—question answering, task-oriented dialogues, and social chatbots—and discusses a unified RL-based perspective, training approaches, and industry applications.

ABSTRACT

The present paper surveys neural approaches to conversational AI that have been developed in the last few years. We group conversational systems into three categories: (1) question answering agents, (2) task-oriented dialogue agents, and (3) chatbots. For each category, we present a review of state-of-the-art neural approaches, draw the connection between them and traditional approaches, and discuss the progress that has been made and challenges still being faced, using specific systems and models as case studies.

Motivation & Objective

• Provide a comprehensive survey of neural approaches to conversational AI across QA, task-oriented dialogues, and social chatbots.
• present a unified view of dialogue as optimal decision making via reinforcement learning.
• connect neural approaches to traditional NLP/IR methods and highlight progress and challenges.
• outline training paradigms (supervised and reinforcement learning) and evaluation techniques.
• illustrate industry trends and case studies of deployed conversational systems.

Proposed method

• Categorize neural dialogue systems into QA, task-oriented, and social chatbot domains.
• Review state-of-the-art neural models and training regimes for each category.
• Connect neural methods with traditional symbolic approaches to reveal evolution and trade-offs.
• Adopt a unified reinforcement learning framework to model hierarchical dialogue decision making.
• Discuss end-to-end versus modular architectures and their implications.
• Present industry case studies to illustrate progress and remaining challenges.

Experimental results

Research questions

• RQ1 What are the state-of-the-art neural approaches for QA, including knowledge-base QA and machine reading comprehension?
• RQ2 How are neural methods applied to task-oriented dialogues and what are effective training and evaluation strategies?
• RQ3 What are the developments and challenges in neural social chatbots, especially regarding end-to-end generation and grounding?
• RQ4 How can dialogue be framed as an optimal decision-making process within reinforcement learning?
• RQ5 What industry systems embody these neural approaches and what lessons do they provide?

Key findings

• Neural approaches have achieved state-of-the-art results on various conversation benchmarks across QA, task-oriented, and social chatbot tasks.
• End-to-end neural architectures are reshaping the dialogue stack, reducing reliance on hand-engineered features and traditional NLP components.
• Reinforcement learning provides a unified framework for learning hierarchical dialogue policies and handling long-term engagement versus short-term task completion.
• Grounding chatbots with world knowledge and multimodal information (e.g., images) enhances contentfulness and usefulness of conversations.
• Industry systems (e.g., Bing QA, Satori QA, XiaoIce) demonstrate practical deployment of neural dialogue models and highlight ongoing challenges such as evaluation, data, and user engagement.