Extracting events from texts using neural networks has gained increasing research focus in recent years. However, existing methods prepare candidate arguments in a separate classifier suffering from the error propagation problem and fail to model correlations between entity mentions and event structures. To improve the performance of both entity recognition and event extraction, we propose a transition-based joint neural model for the tasks by converting graph structures to a set of transition actions. In particular, we design ten types of novel actions and introduce a global normalization strategy to alleviate the label bias issue. We conduct experiments based on the widely used Automatic Content Extraction (ACE) corpora and the results show that our model achieves 88.7% F1-score on entities and 75.3% F1-score on event triggers, outperforming the baseline neural networks by a large margin. Further in-depth analysis shows the effectiveness of our model in capturing structural dependencies in long sentences. The proposed model can be used for facilitating a range of downstream tasks.

近年来，使用神经网络从文本中提取事件已成为越来越多的研究重点。但是，现有方法在遭受错误传播问题的单独分类器中准备候选参数，并且无法对实体提及和事件结构之间的相关性进行建模。为了提高实体识别和事件提取的性能，我们通过将图结构转换为一组过渡动作，为任务提出了基于过渡的联合神经模型。特别是，我们设计了十种新颖的动作，并引入了一种全局标准化策略来缓解标签偏差问题。我们根据广泛使用的自动内容提取（ACE）语料库进行了实验，结果表明，我们的模型在实体上达到88.7％的F1-分数，在事件触发条件下达到75.3％的F1-分数，在很大程度上优于基线神经网络。进一步的深入分析显示了我们的模型在捕获长句中的结构依存关系方面的有效性。所提出的模型可用于促进一系列下游任务。