Generative commonsense knowledge subgraph retrieval for open-domain dialogue response generation.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Knowledge Representation

Background:

• Dialogue systems benefit from grounding in commonsense knowledge for more informative and diverse responses.
• Existing methods rely on external knowledge bases (eKB) or generative retrieval, which have limitations in knowledge scope and quality.

Purpose of the Study:

• To propose a novel method, Traverse in Language Model (TiLM), for generating high-quality knowledge subgraphs internally.
• To improve dialogue response generation by grounding it in internally constructed knowledge subgraphs without relying on eKB during inference.

Main Methods:

• TiLM utilizes three 'Chain-of-Thought' sub-tasks: Query Entity Production, Topic Entity Prediction, and Knowledge Subgraph Completion.
• These sub-tasks collaboratively build a knowledge subgraph directly within the language model.

Main Results:

• TiLM successfully generates high-quality knowledge subgraphs for grounding dialogue responses.
• Experimental results on Chinese and English datasets show TiLM's outstanding performance, even with fewer parameters.

Conclusions:

• TiLM offers an effective approach to internal knowledge subgraph generation for enhanced dialogue systems.
• The method overcomes the limitations of explicit eKB retrieval and one-pass generative retrieval, leading to better response quality.