基于协同专家系统的建筑施工大语言模型问答系统

A Construction Question and Answer System Based on a Collaborative Expert System and Large Language Model

摘要:

为解决大型语言模型问答系统在建筑施工场景中存在的生成幻觉与部署成本高的问题，提出了一种基于协同专家机制的大型语言模型施工问答系统。该系统通过共享专家与路由专家的协同工作方式，在保证模型表达能力的同时，显著提升了问答生成的准确性与推理效率，并有效降低了计算开销。此外，设计了一种领域知识库注入的微调策略，在训练阶段引导模型深度学习施工领域专业语义，从而增强其对工程文本的理解能力，确保生成结果更加符合实际工程需求。实验结果表明，在仅激活约1/3模型参数的情况下，所提出系统仍可达到81.1%的生成语义相似度，兼顾了效率与性能，为建筑施工管理提供了一种高效、可靠且具备工程针对性的智能决策支持工具。

Abstract:

To address the issues of hallucinated generation and high deployment costs encountered by large language model （LLM）–based question answering systems in construction scenarios， this paper proposes a construction-oriented question answering system based on a collaborative expert mechanism. The system integrates shared experts and routing experts in a coordinated manner， which significantly improves the accuracy of answer generation and inference efficiency while preserving the model’s expressive capacity and reducing computational overhead. In addition， a domain knowledge base–injected fine-tuning strategy is introduced to guide the model to deeply learn professional semantics in the construction domain during training， thereby enhancing its understanding of engineering-related texts and ensuring that the generated responses better align with practical engineering requirements. Experimental results demonstrate that， with only approximately one-third of the model parameters activated， the proposed system achieves a generation semantic similarity of 81.1%， effectively balancing efficiency and performance and providing an efficient， reliable， and construction-specific intelligent decision-support tool for construction management.