类型感知的汉越跨语言事件检测方法

针对汉越跨语言事件检测缺少平行语料，越南语标注困难，需要统一跨语言语义空间，且触发词存在较大的歧义和局限性等问题，提出基于事件类型感知的汉越跨语言事件检测方法。构造类型感知的注意力机制突显事件特征，融入汉越的词位置、词性和命名实体信息，并通过梯度反转（gradient reversal layer，GRL），实现有标注汉语和无标注越南语之间的对抗训练，将从大量汉语新闻文本中学到的语言无关的事件类型特征融入到联合特征提取器中，进行汉越跨语言的无触发词事件检测，缓解越南语的数据稀缺和触发词的局限性。实验中提出的方法较最好的基线模型在准确率上提升了4.32%。

Chinese-Vietnamese cross-language event detection method based on event type awareness

In view of the lack of parallel corpus for Chinese-Vietnamese cross-language event detection, the difficulty of Vietnamese labeling, the need to unify the cross-language semantic space, and the large ambiguity and limitations of trigger words, this paper proposes an event-type-aware-based Chinese-Vietnamese cross-language event detection method. First, an event-type-aware attention mechanism is constructed to highlight event features, integrate Chinese and Vietnamese language differences into position, part of speech, and named entity information, and achieve confrontation training between tagged Chinese and unlabeled Vietnamese through gradient reverse layer (GRL). The language independent event type features learned from a large number of Chinese news texts are integrated into the joint feature extractor to detect Chinese Vietnamese cross language non-trigger word events. It alleviates the data scarcity and the limitation of trigger words in Vietnamese. The accuracy of the proposed method is 4.32% higher than that of the best baseline model.