스포츠 지식 문서 기계학습을 통한 융합 스포츠 특허 토픽 모델링 및 토픽 영역 결정 시스템 통합 개발(특허출원 : 10-2023-0124901, 10-2023-0124899)

초록

Purpose: The purpose of this study is to develop an integrated framework that combines deep learning based topic modeling with expert validated topic area identification to analyze convergence sports patents and identify emerging technology trends. Method: SportsBERT, a domain adaptive language model pretrained on approximately eight million sports news articles, was applied to convergence sports patents retrieved from WIPO PATENTSCOPE. BERTopic with UMAP and hierarchical clustering was used for topic extraction, and topic areas were identified through visualization based analysis followed by expert validation. Technology trends were classified as Hot, Warm, Cool, or Cold on the basis of linear regression of annual topic ratios. Results: SportsBERT-BERTopic outperformed latent Dirichlet allocation (LDA) in topic coherence (C_V: 0.647) and expert interpretability (4.3/5.0). Twenty three topics were organized into five topic areas, and expert validation improved topic area matching accuracy by 0.9 to 1.1 points following document map based reallocation. Among the extracted topics, 13 were classified as Hot, with information_processing_space_apparatus (β=0.878), operation_scene_control_game (β=0.851), and exercise_training_fitness_user (β=0.847) exhibiting the strongest growth trends. Conclusion: The proposed framework effectively integrates domain specific language modeling with expert validation to improve topic identification and technology trend analysis in convergence sports patents. These findings provide a practical patent intelligence framework

키워드