句子語義配對是自然語言處理的重要任務之一，主要被廣泛應用在比較多個句子的語義並獲得他們的相似度以進行篩選或排名，常被用於搜尋引擎、問答系統，以找出最合適的回覆。過去的研究通常考慮不同的文字特徵萃取方法，卻忽略不同語義的句子會提供不同的交互知識，對於句子語義配對任務而言，相似的句子間仍會存在不同構面，傳統的方法中只能表明其相關性，不足以分出更適當的候選，導致系統表現有限。為解決這種難題，我們開發了一種新的強化對比學習（RCL）模型來產生語義特徵，該模型結合了交叉注意力機制和對比學習來輔助判斷相鄰特徵。我們也將 RCL 運用至真實世界的資料集中，並驗證其表現皆優於基準模型。;Sentence Semantic Matching (SSM) is a crucial component in natural language processing (NLP) tasks. It involves comparing the semantics of multiple sentences and ranking their similarities to identify the most similar one. Recently, contrastive learning has been proven to be beneficial in generating complex semantic features and promoting performance. Early research usually considers the different data construction, but ignoring the different semantic sentences will give variant knowledge of the interaction to sentence anchor, which might not be enough to capture the comprehensive observation of semantic features and lead to limited performance. We developed a new Reinforced Contrastive Learning (RCL) model to generate contextual features, which combined a cross-attention mechanism and contrastive learning to assist the adjacent feature. RCL was applied to numerous real-world datasets and it demonstrated state-of-the-art experimental results on the SSM task.
