A Multi-Task Framework for Intelligent Review and Semantic Consistency Detection in Scientific Project Documentation Using Large Language Models

Abstract

With the intensive and active global technological innovation, the quality of technology project document review is crucial. However, traditional manual review is time-consuming and biased, and rule driven tools are difficult to handle multimodal and dynamic terms. Large Language Models (LLMs) provide a new path for this. This study proposes a multi-dimensional intelligent evaluation multi task architecture based on LLM, with "pre training fusion knowledge enhancement fine tuning adaptation multi task reinforcement learning continuous optimization" as the core: in the pre training stage, the domain knowledge graph (DKG) is fused to construct a terminology library and form a mixed reasoning ability, and in the fine tuning stage, conflict detection, logic evaluation, and consistency analysis sub task heads are designed based on open source models such as LLaMA (Large Language Model Meta AI), while introducing reinforcement learning optimization strategies and achieving model lightweighting. The experiment selected 800 documents from a national research institution as the test set, and the results showed that the accuracy of technical parameter conflict detection was 92.7% (18.3 percentage points higher than traditional rule engines), the F1 value of logical consistency evaluation was 89.1% (13.7 percentage points better than keyword matching methods), the average recall rate of cross document semantic consistency analysis was 87.6%, and processing 100000 words of text only took 2.5 hours (6 times more efficient than manual labor). After integrating reinforcement learning, the coverage of implicit association recognition was increased to 89.7%, the false alarm rate was reduced to 4.3%, and the parameter compression of 35% still maintained 91.2% of core performance. This architecture breaks through traditional limitations and provides technical support for the digitization of technology project management.