融合多级语义特征的双通道GAN事件检测方法

目前事件检测方法往往将句中事件视为独立个体，忽视了句子或文档内事件间的相关关系，且某些触发词在不同语境下可能触发不同事件，而多种语境下训练的词向量会引入与当前语境无语义关联的噪声.针对此问题，本文提出一种融合多级语义特征的双通道GAN事件检测方法，使用多级门限注意力机制获取句子级和文档级事件间的语义相关性，并利用双通道GAN及其自调节学习能力减轻噪声信息的影响，进而提高事件特征表示的准确性.在公开数据ACE2005英文语料上进行实验，F₁值达到了77%，结果表明该方法能够有效获取事件间的语义相关性，并提高语境判定的准确性. 

Double-Channel GAN with Multi-Level Semantic Correlation for Event Detection

Event detection is an important task of information extraction. In recent years, it has been widely used in the fields of knowledge graph construction, information retrieval and intelligence research. For current event detection methods, events within one sentence are often identified as independent individuals, while the correlation among the events within one sentence or document is ignored. Besides, some triggers may trigger different events in different contexts, and the word vectors training in multiple contexts can introduce noise that is not semantically related to the current context. To solve the problems, a double-channel GAN with multi-level semantic correlation was proposed for event detection. Firstly, a multi-level gated attention mechanism was utilized to capture the semantic correlation among sentence-level events and document-level events. And then, a double-channel GAN with self-regulation learning was used to reduce noise and improve accuracy of the representation of event. Finally, some experiments on ACE2005 English corpus were carried out. The results show that, F₁ score can achieve 77%, and the method can effectively obtain semantic correlation among multi-level events, and improve accuracy of context determination.