Advancing faithful KBQA services via prompt-driven knowledge graph-enhanced LLMs in cloud

LLM hallucination and limited factual knowledge: LLMs have constrained factual stores and hallucination issues that impair effectiveness in intricate reasoning scenarios.

Static KG utilization: Current KG-based approaches often treat knowledge graphs as fixed repositories and overlook their relational architecture, causing suboptimal and potentially spurious reasoning.

LLM hallucination and limited factual knowledge: LLMs have constrained factual stores and hallucination issues that impair effectiveness in intricate reasoning scenarios.

Static KG utilization: Current KG-based approaches often treat knowledge graphs as fixed repositories and overlook their relational architecture, causing suboptimal and potentially spurious reasoning.

Prompt-driven KG-enhanced LLM reasoning: PDR combines LLMs and KGs cohesively while refining prompts to improve reasoning reliability and interpretability.

PageRank-based subgraph retrieval with document expansion: PDR retrieves relevant subgraphs using a refined PageRank algorithm aligned with KG structure and integrates document retrieval to extend graph boundaries.

Prompt-guided chain-of-thought KG path reasoning: PDR uses task-specific prompts with subqueries as cues to generate chain-of-thought and candidate KG paths, then filters them by semantic coherence and KG structural alignment.

Prompt-driven KG-enhanced LLM reasoning: PDR combines LLMs and KGs cohesively while refining prompts to improve reasoning reliability and interpretability.

PageRank-based subgraph retrieval with document expansion: PDR retrieves relevant subgraphs using a refined PageRank algorithm aligned with KG structure and integrates document retrieval to extend graph boundaries.

Outperforms state-of-the-art baselines:

Improved answer accuracy and interpretability:

High-coverage relevant subgraphs: