基于篇章信息和Bi-GRU的中文事件检测

Chinese Event Detection Based on Document Information and Bi-GRU

事件抽取是信息抽取中一个重要的研究方向,其中事件检测是事件抽取的关键。目前,中文神经网络事件检测方法均是基于句子的方法,这种方法获得的局部上下文的信息不足以解决事件触发词的歧义性。针对这个问题,文中探索了篇章信息的作用。首先,以双向门控循环单元网络(Bidirectional Gated Recurrent Units,Bi-GRU)模型为基线,定义3个窗口来学习句子特征;然后,将句子表示进行拼接,利用双向门控循环单元网络学习句子的上下文特征;最后,将句子表示和上下文表示进行融合,以丰富句子的语义信息,并减少候选触发词语义模糊现象,通过Softmax函数进行事件触发词的分类。在ACE2005数据集上的实验结果表明,句子的上下文特征能够有效提升中文事件检测方法的性能,该中文事件检测方法的F1值比当前最好的模型高1.5%。

Event extraction is an important research task in information extraction and event detection is the key to event extraction.Existing Chinese neural network event detection methods are sentence-based and the local context information obtained by this method is not enough to resolve the event triggers semantic ambiguity.In order to solve this problem,this paper studies document information effects.Firstly,based on the bidirectional gated recurrent units network(Bi-GRU),this paper defines three windows to learn sentence features.Then,the sentence-level representation is concatenated and the document features are learned by using the bidirectional gated recurrentunits network.Finally,to enrich the semantic information of sentences and reduce the event-trigger sematic event triggers ambiguity,it merges the sentence-level representation and the document-level representation and then classifies eventtriggers through the Softmax function.Experimental results on the ACE2005 dataset show that the sentences-context representation can improve the Chinese event detection performance and this event detection method outperforms state-of-the-art results by 1.5% on F1.