Knowledge graph construction for intelligent cockpits based on large language models

Area of Science:

• Human-Computer Interaction
• Artificial Intelligence
• Knowledge Representation

Background:

• Intelligent cockpits require advanced user behavior inference for proactive interaction.
• Traditional methods face scalability, generalization, and accuracy issues, leading to pseudo-demand functions.
• Knowledge graphs (KGs) offer a solution for organizing complex information and understanding user needs.

Purpose of the Study:

• To develop a novel knowledge graph construction (KGC) model, GLM-TripleGen, for analyzing intelligent cockpit states and behaviors.
• To precisely mine latent relationships between cockpit states and user behaviors.
• To address challenges in entity recognition and relationship extraction within cockpit data.

Main Methods:

• Introduced GLM-TripleGen, a domain-specific KGC model for intelligent cockpits.
• Constructed an instruction-following dataset tailored to vehicle states and in-cockpit interactions.
• Employed the Low-Rank Adaptation (LoRA) method for efficient model fine-tuning and parameter optimization.

Main Results:

• GLM-TripleGen demonstrated superior performance compared to state-of-the-art KGC methods.
• The model accurately generated normalized cockpit triple units.
• Achieved robust performance and strong generalization capabilities with unknown entities and relationships.

Conclusions:

• GLM-TripleGen effectively addresses the challenges of KGC in intelligent cockpit domains.
• The model enhances the system's ability to understand user behavior and contextual information.
• GLM-TripleGen offers a robust and adaptable solution for knowledge graph construction in evolving intelligent systems.