基于局部实体特征的事件触发词抽取

Event Trigger Extraction Based on Local Entity Characteristics

基于神经网络的触发词抽取模型利用实体信息判别触发词,但大量无关实体会影响触发词抽取效果。提出一种借助局部实体特征的事件触发词抽取方法,该方法先初步过滤无关实体,并将保留实体分为核心与非核心2类分别进行建模。利用卷积神经网络(CNN)抽取局部特征的特性,从众多实体中定位有助于触发词识别的局部重要实体,采用注意力机制提高其权重,同时利用有效非核心实体的语义排除干扰实体,从而借助重要实体的特征信息判别触发词。在特定和通用领域事件语料库上的实验结果均表明,该方法能够减少无关实体对触发词抽取的干扰,其触发词抽取性能的F1值比基准系统最高可提升0.017。

The current neural network-based event trigger extraction model distinguishes triggers by entity information,but lots of unrelated entities have negative interference on trigger extraction.Therefore,we propose an event trigger extraction method based on local entity characteristics.First,we filter unrelated entities,separate the reserved entities into core and non-core types,and build the reserved entity model.Then,with the utilization of CNN network,which is superior in the extraction of local features,the model manages to locate the entities that are important for trigger recognition from all the entities.Subsequently,we adopt the Attention mechanism to improve the weights of important entities,so as to identify triggers by their characteristic information.Experimental results on domain-specific and domain-general event corpus show that the proposed method can reduce the interference of unrelated entities on trigger extraction,and its F1-score of trigger extraction performance,if at the highest level,can exceed the baseline by 0.017.