Joint Knowledge Base Completion and Question Answering by Combining Large Language Models and Small Language Models
Summary: arXiv:2604.05875v1 Announce Type: new
Knowledge Bases (KBs) play a crucial role in various applications, serving as the foundation for intelligent systems that require structured information. Among the significant tasks related to KBs are Knowledge Base Completion (KBC) and Knowledge Base Question Answering (KBQA). These tasks are closely interconnected and inherently complementary, making it advantageous to approach them in a unified manner. The integration of KBC and KBQA not only enhances their effectiveness but also allows them to reinforce each other.
Despite the potential benefits of a joint approach, existing studies predominantly utilize Small Language Models (SLMs) to enhance these tasks. This method overlooks the significant reasoning capabilities offered by Large Language Models (LLMs), which can provide more robust solutions. To address this gap, a novel framework named JCQL has been proposed, designed to synergistically combine the strengths of both LLMs and SLMs.
The JCQL Framework
The JCQL framework focuses on the iterative enhancement of KBC and KBQA. The design aims to leverage the capabilities of LLMs to improve SLM performance and vice versa. The following key strategies are employed within the framework:
- Enhancing KBQA with KBC: The framework augments the reasoning paths of the LLM-based KBQA model by incorporating an SLM-trained KBC model. This integration serves as an action taken by the agent, addressing challenges related to LLM hallucinations and high computational costs typically associated with KBQA tasks.
- Improving KBC through KBQA: The KBC model is incrementally fine-tuned using reasoning paths generated from the KBQA task. This approach provides supplementary training data that enhances the SLM’s capabilities in KBC, ultimately leading to improved performance.
Experimental Validation
The effectiveness of the JCQL framework was validated through extensive experiments on two public benchmark datasets. The results indicated that JCQL significantly outperforms all baseline models in both KBC and KBQA tasks. The experiments demonstrate how the iterative learning process between the two models creates a more effective system for managing and querying knowledge bases.
Conclusion
In conclusion, the integration of Large Language Models and Small Language Models within the JCQL framework presents a promising direction for advancing knowledge base related tasks. By enabling KBC and KBQA to enhance each other iteratively, this approach not only improves the accuracy and efficiency of these tasks but also paves the way for future research in the field. The potential applications of this framework are vast, spanning various domains that rely on effective knowledge management and retrieval.