關係抽取是信息抽取任務中非常重要的一項子任務，核心目的是從句子中尋找存在關係的實體對，並為實體對匹配對應的關係類型。本論文在對關係抽取任務的研究中提出了應用於關係抽取任務的一種類代理特徵融合模型。該方法在一種級聯二進制標註的關係抽取框架基礎上，加入了一個類代理注意力特徵融合模塊，以提高模型對句子中關係實體的抽取效率。該模塊通過深度學習與訓練一個包含句子上下文語義信息與句子類別信息的代理向量，接著模塊會對代理向量中的有效信息進行再次提煉，幫助模型增強訓練數據中提取關係實體和關係類別中隱含的語義信息能力。藉由實驗結果證實，類代理注意特徵融合模型的關係抽取性能相比CasRel有明顯提升，因此，本論文所提出的模型可以有效地提高關係三元組的抽取效率。;Relation extraction is a very important subtask in the information extraction task. The core purpose is to find entity pairs with relations in sentences and match the corresponding relation types for the entity pairs. In the study of relation extraction tasks, this thesis proposes an agent-like attention feature fusion model for relation extraction tasks. Based on a cascade binary tagging relation extraction framework, this method adds an agent-like attention feature fusion module to improve the model’s extraction efficiency of relational entities in sentences. This module trains an agent vector containing sentence context semantic information and sentence category information. Then, the module will further refine the effective information in the agent vector to help the model enhance its ability to extract implicit semantic information in relational entities and relation categories from training data. Experimental results show that the relation extraction performance of the agent-like attention feature fusion model is significantly improved compared with CasRel. Therefore, the model proposed in this thesis effectively improves the extraction efficiency of relational triples.
