IG-LLM 기반의 핵심어를 활용한 국가 핵심기술 연관도 분석 연구

초록

With intensifying global competition for technological supremacy and recurring cases of national core technology leakage, strengthening the protection system for national core technologies has become critical to safeguarding technological sovereignty. However, the current system has been criticized for ambiguously defining the scope of national core technologies and lacking clear classification criteria. Although approaches such as a CPC-based classification system, topic modeling, and statistical keyword analysis have been proposed, research that simultaneously addresses quantitative classification at the keyword and keyphrase level together with qualitative interpretation remains limited. To overcome these shortcomings, this study proposes an IG-LLM hybrid keyword derivation method and a quantitative evaluation framework for analyzing the relevance of national core technologies. Integrated Gradients was applied on KorSciBERT to derive high-salience keywords, which were then compared with and refined using contextual verification results from a GPT-4.1-mini-based large language model to select the final core keywords. The selected keywords were subsequently used as input to a KorSciBERT-based classification model, and four quantitative evaluation schemes-binary classification, weighted-sum classification, machine-learning-based classification, and deep-learning-based classification were constructed for comparative evaluation against existing methods. Experiments conducted on patent documents in the display domain show that the proposed method outperforms CPC-based classification systems and TF-IDF, LDA, and Word2Vec-based single and integrated keyword identification techniques in terms of accuracy and F1-score. Moreover, by combining IG-based importance with LLM-based contextual verification, the proposed approach simultaneously enhances interpretability and discriminative capability at the keyphrase level, enabling clearer presentation of the classification rationale. In addition, through in-depth analyses of the appropriateness of the base learning model, the suitability of the keyphrase representation, and the performance advantages of the hybrid model over single-model approaches, the validity of the proposed framework was verified. This study can serve as foundational evidence to supplement the criteria used to identify national core technologies and holds academic and practical significance as an alternative keyword-based identification and relevance analysis methodology that addresses the limitations of CPC-based classification systems.

키워드