Reproduce “Think Before You Speak: Explicitly Generating Implicit Commonsense Knowledge for Response Generation” on OpenResearch

About this paper

Pei Zhou, Karthik Gopalakrishnan, Behnam Hedayatnia, Seokhwan Kim, Jay Pujara, Xiang Ren, Yang Liu, Dilek Hakkani-Tur

The Think-Before-Speaking (TBS) framework enhances neural response generation models by explicitly generating implicit commonsense knowledge before formulating a response. This two-step approach yields more informative, specific, and explainable dialogue, consistently outperforming existing baselines and demonstrating the ability to generate novel, high-quality knowledge.

Key insights

• Explicitly generating implicit commonsense knowledge as an intermediate step leads to more informative, specific, and explainable responses in dialogue systems.
• A unified auto-regressive generative model can effectively learn to both infer relevant commonsense knowledge and leverage it for high-quality response generation.
• The framework demonstrates the ability to generate novel, high-quality commonsense knowledge not explicitly present in its training knowledge base, suggesting strong generalization capabilities.

Summary by alphaXiv.