【摘要】 针对目前大多数事件要素识别模型未考虑词级别的语义信息，及模型鲁棒性不高的问题，提出一种融合词信息和对抗训练的事件要素识别方法。将Bert(bidirectional encode representations from transformers)预训练语言模型生成的字向量与分词信息进行融合，在得到的融合向量中添加扰动因子产生对抗样本，将对抗样本与融合向量表示作为编码层的输入；采用BiGRU(bidirectional gating recurrent unit)网络对输入的文本进行编码，丰富文本的上下文语义信息；采用CRF(conditional random field)函数计算完成事件要素的识别任务。在CEC(Chinese emergency corpus)中文突发事件语料库上的实验结果表明，该方法能够取得较好的效果。更多还原

【Abstract】 To solve the problems that most of the current event element recognition models do not consider the semantic information of the word level, and the robustness of the models is not high, an event element recognition method that combined word information and adversarial training was presented. The word vectors generated using the Bert(bidirectional encode representations from transformers) pre-trained language model were fused with word segmentation information. Disturbance factors were added to the obtained fusion vectors to generate adversarial samples. The adversarial samples and fusion vectors were represented as inputs to the encoding layer. BiGRU(bidirectional gating recurrent unit) network was used to encode the input text to enrich the context semantic information of the text. The CRF(conditional random field) function was used to calculate and complete the task of identifying event elements. Experimental results on the CEC(Chinese emergency corpus) show that the method can achieve acceptable results.更多还原