Government-funded science and technology innovation projects are vital for driving industrial development and supporting talent cultivation. However, evaluating their outcomes remains a significant challenge, especially when some researchers misattribute unrelated publications to funding projects, raising concerns about research integrity and transparency. This paper focuses on the challenging task of assessing the relevance between project proposals and research outputs, formulated as a long-text matching problem. Due to the fact that even valid research outputs often address only subtopics of the original project objectives, traditional methods, which typically compare entire documents, often fail to provide accurate relevance assessments. To address this, we propose ConceptSplitter, a concept-based chunking method inspired by long-text structuring strategies. As part of a retrieval-augmented generation (RAG) pipeline, ConceptSplitter serves as the chunking module that improves retrieval precision and contextual relevance in large language model inference. To support robust evaluation, we also construct a domain-diverse dataset that mirrors real-world funding scenarios. Experiments on this dataset show that ConceptSplitter outperforms traditional methods by enhancing chunking quality, improving the accuracy of relevance classification, and providing more reliable confidence estimation in large language model outputs.