In recent years, the increasing complexity and volume of information in industry-academia collaboration research have posed substantial challenges for effective knowledge discovery. To address this, the study proposes a semantic retrieval-based natural language processing (NLP) system designed to enhance the retrieval and analytical capabilities of textual data by integrating three main modules (retrieval, filtering, and analysis) and implementing both document-level and sentence-level retrieval strategies. Experimental results demonstrate that sentence-level retrieval outperforms document-level retrieval as well as the traditional lexical matching method BM25. Additionally, through word cloud visualization and analysis of temporal trends, this study highlights the evolution of topics in collaborative research, including research outcomes management and innovation of cross-organizational collaboration mechanisms. Although industry-academia collaboration is used as a case example, the proposed system is broadly applicable to other domains, facilitating the identification of emerging trends and overlooked areas, and providing valuable support for researchers and policymakers across various fields.