텍스트 의미와 사건 구조 융합 기반 게임 시나리오 유사성 탐지에 관한 연구

초록

This study proposes a Late Fusion framework for game scenario similarity analysis by combining textual semantics with SVO-based event structure. The text channel uses SBERT, Smith-Waterman, DTW, LMSS, and coverage features, while the graph channel analyzes SVO event Tri-Graphs with a GIN encoder. Experiments on about 130 English visual novel scripts and their controlled variants achieved ROC-AUC 0.9966 and Top-1 accuracy 1.0. The results show that the framework can support similarity screening and source candidate narrowing in a controlled benchmark.

키워드