Scholarlink

<p>Organizations seeking external funding often face the labor-intensive challenge of reviewing complex calls for proposals, which typically include legal, financial, and geographic eligibility criteria. This paper presents a hybrid classification approach that integrates generative AI (specifically the GPT-o4-mini model) with semantic embeddings to automate the classification of funding calls as viable or non-viable, and relevant or not, based on both formal criteria and thematic alignment with institutional experience. The model leverages prompt engineering to extract explicit and implicit eligibility information from unstructured call documents and uses sentence embeddings (all-MiniLM-L6-v2) to measure semantic similarity between call objectives and a university’s project history. Two prompting strategies are evaluated: a single general prompt and a modular, piecewise prompting approach. The latter, though more computationally intensive, achieves significantly higher performance (F1 score of 0.84 vs. 0.66). This study demonstrates that combining large language models with semantic similarity analysis enhances classification accuracy and supports more informed, efficient decision-making in funding opportunity assessment.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

From Text to Decision: A GenAI Framework for Strategic Evaluation of Funding Opportunities

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年9月13日

查看全文
  • Yeersainth Figueroa-Gómez,
  • Ixent Galpin

摘要

Organizations seeking external funding often face the labor-intensive challenge of reviewing complex calls for proposals, which typically include legal, financial, and geographic eligibility criteria. This paper presents a hybrid classification approach that integrates generative AI (specifically the GPT-o4-mini model) with semantic embeddings to automate the classification of funding calls as viable or non-viable, and relevant or not, based on both formal criteria and thematic alignment with institutional experience. The model leverages prompt engineering to extract explicit and implicit eligibility information from unstructured call documents and uses sentence embeddings (all-MiniLM-L6-v2) to measure semantic similarity between call objectives and a university’s project history. Two prompting strategies are evaluated: a single general prompt and a modular, piecewise prompting approach. The latter, though more computationally intensive, achieves significantly higher performance (F1 score of 0.84 vs. 0.66). This study demonstrates that combining large language models with semantic similarity analysis enhances classification accuracy and supports more informed, efficient decision-making in funding opportunity assessment.