實驗結果顯示，不同檢索範式能捕捉不同類型的事件對應關係。字面方法較適合處理具有明確文字重疊的案例，而語意與結構表示則能為較隱性的事件對應提供互補訊號。融合系統在不同歷史時期之間亦能維持有效的檢索能力，但敘事改寫仍是最具挑戰性的類型。整體而言，結合字面、語意、時間與關係結構等多種證據，比單一檢索範式更適合處理跨視角歷史事件對齊。;The Ming Shilu and Joseon Sillok record many of the same events of the Imjin War (1592–1598) from different historical perspectives. However， differences in vocabulary， personal designations， recording dates， and narrative emphasis make automatic alignment between the two classical Chinese corpora challenging. This thesis formulates the problem as Cross-Perspective Event Alignment (CPEA)， a multi-label retrieval task that retrieves Joseon-side records describing the same underlying event as a given Ming-side record.
To address this task， a manually verified dual-year dataset is constructed and a temporally separated development–test setting is adopted to evaluate cross-period generalization. The proposed framework combines three complementary retrieval paradigms: sparse lexical retrieval， dense semantic retrieval， and knowledge-graph structural matching. Representative signals from these paradigms are integrated through Bayesian-optimized weighted score fusion to construct a high-recall candidate pool for downstream semantic judgment.
Experimental results show that the three retrieval paradigms capture different types of event correspondence. Lexical methods are most effective for explicit textual overlap， while semantic and structural representations provide complementary evidence for more implicit correspondences. The fused system also maintains effective retrieval performance across historical periods， although narrative paraphrases remain the most challenging category. Overall， the results demonstrate that combining lexical， semantic， temporal， and relational evidence provides a more robust approach to cross-perspective historical event alignment than relying on a single retrieval paradigm.
