国家重点研发计划(2018YFC0407105); 江苏省重点研发计划(BE2020729); 江苏省研究生科研创新项目(B200203130); 华能集团总部科技项目(HNKJ19-H12)
针对事件抽取存在未充分利用句法关系、论元角色缺失的情况, 提出了基于双重注意力机制的事件抽取(event extraction based on dual attention mechanism, EEDAM)方法, 有助于提高事件抽取的精确率和召回率. 首先, 基于4种嵌入向量进行句子编码, 引入依赖关系, 构建依赖关系图, 使深度神经网络可以充分利用句法关系. 然后, 通过图转换注意网络生成新的依赖弧和聚合节点信息, 捕获长程依赖关系和潜在交互, 加权融合注意力网络, 捕捉句中关键的语义信息, 抽取句子级事件论元, 提升模型预测能力. 最后, 利用关键句检测和相似性排序, 进行文档级论元填充. 实验结果表明, 采用基于双重注意力机制的事件抽取方法, 在ACE2005数据集上, 较最佳基线联合多中文事件抽取器(joint multiple Chinese event extractor, JMCEE)在精确率、召回率和F1-score分别提高17.82%、4.61%、9.80%; 在大坝安全运行日志数据集上, 较最佳基线JMCEE在精确率、召回率和F1-score分别提高18.08%、4.41%、9.93%.
In view of the fact that the syntactic relationship is not fully utilized and the argument role is missing in event extraction, an event extraction based on dual attention mechanism (EEDAM) method is proposed to improve the accuracy and recall rate of event extraction. Firstly, sentence coding is based on four embedded vectors and dependency relation is introduced to construct dependency relation graph, so that deep neural network can make full use of syntactic relation. Then, through graph transformation attention network, new dependency arcs and aggregate node information are generated to capture long-range dependencies and potential interactions, weighted attention network is integrated to capture key semantic information in sentences, and sentence level event arguments are extracted to improve the prediction ability of the model. Finally, the key sentence detection and similarity ranking are used to fill in the document level arguments. The experimental results show that the event extraction method based on dual attention mechanism can improve the accuracy rate, recall rate, and F1-score by 17.82%, 4.61%, and 9.80% respectively compared with the optimal baseline joint multiple Chinese event extractor (JMCEE) on ACE2005 data set. On the data set of dam safety operation records, the accuracy, recall rate, and F1 score are 18.08%, 4.41%, and 9.93% higher than the optimal baseline JMCEE, respectively.