Boosting expertise and efficiency in LLM: A knowledge-enhanced framework for construction support

Current research on question answering in construction management faces multiple challenges, including insufficient construction knowledge in general large language models (LLMs) and the challenge for traditional LLM-based Q&A systems to balance response accuracy with computational cost. To address these issues, this study proposes a knowledge-enhanced framework that incorporates two key components: (i) a Collaborative Expert Module (CEM), which enhances generation reliability while preserving efficiency through coordinated learning in a sparse network, and (ii) a Knowledge-Injected (KI) training strategy that dynamically augments model training with relevant domain knowledge. Experimental evaluations demonstrate that the proposed method not only outperforms the baseline Qwen1.5-MoE by 8.2% in ROUGE-L and 9.0% in semantic similarity, but also surpasses state-of-the-art models such as GPT-4 Turbo and DeepSeek V3, while operating with only 2.7B activated parameters. This work contributes a data-driven approach that strengthens LLMs with construction-specific knowledge, offering a more reliable and efficient foundation for domain-specific question answering.