JCQL: A Joint Framework for Knowledge Base Completion and Question Answering via Collaboration Between Large and Small Models

JCQL (Joint Completion and Query Learning) is an innovative large-small model collaboration framework. Its core uses an iterative collaboration mechanism to enhance Knowledge Base Question Answering (KBQA) with the reasoning capabilities of Large Language Models (LLMs), while using KBQA's reasoning paths to optimize Knowledge Base Completion (KBC) models, achieving mutual enhancement of the two tasks and forming a knowledge loop.

Knowledge bases face two major challenges: incompleteness (missing factual connections) and query complexity (converting natural language to knowledge base queries), corresponding to the KBC and KBQA tasks. Traditional KBC relies on Small Language Models (SLMs) which are efficient but weak in reasoning; KBQA uses LLMs which are strong in reasoning but prone to hallucinations and high in cost. The two are complementary: knowledge completed by KBC can enhance the factual foundation of KBQA, and KBQA's reasoning paths can provide training signals for KBC. However, existing studies mostly handle them in isolation or only use SLMs for joint modeling.

JCQL achieves bidirectional enhancement based on three core concepts: complementary capabilities, iterative enhancement, and action abstraction:

1. KBC enhances KBQA: The KBC model trained by SLMs is used as an action of the LLM Agent, which can be called to complete missing facts, reduce hallucinations, lower costs, and expand coverage.
2. KBQA enhances KBC: Implicit relationships in the successful reasoning paths of KBQA are extracted to incrementally fine-tune the KBC model, enabling it to learn implicit patterns, continuously improve, and enhance generalization. The two form an iterative loop: Initial KBC assists KBQA → KBQA reasoning generates new signals → Enhances KBC → Further improves KBQA.

Evaluated on public datasets, the baselines include independent KBC/KBQA and existing joint methods, with metrics being MRR/Hits@k for KBC and accuracy for KBQA. Results: JCQL outperforms baselines in both tasks; ablation studies show that removing KBC actions or KBQA feedback significantly reduces performance; in cases, the LLM Agent calls KBC to complete key facts and successfully answers multi-hop questions, and the reasoning paths feed back to KBC.

JCQL demonstrates a new paradigm of large-small model collaboration: LLMs are responsible for planning and reasoning, SLMs handle completion and retrieval, with seamless integration; the iterative mechanism is cross-task self-supervision, where the two tasks serve as data sources for each other, reducing reliance on manual annotations; it emphasizes the importance of the knowledge loop (knowledge base → question answering → feedback → knowledge base) for long-term maintenance of knowledge systems.

JCQL is applicable to:

1. Enterprise knowledge management: Completing evolving enterprise knowledge bases and reducing deployment costs;
2. Open-domain question answering: Dynamically expanding knowledge coverage and forming a positive cycle;
3. Multilingual knowledge bases: Cross-language reasoning paths promote knowledge transfer.

Current limitations: High iterative training cost, risk of error accumulation, weak handling of complex long reasoning chains. Future directions: Adaptive iteration to balance performance and cost, introduction of error detection mechanisms, expansion to multimodal knowledge bases, federated learning to achieve distributed collaborative completion under privacy protection.