【摘要】 深度学习技术虽已广泛运用在金融命名实体识别中，但在现有的金融命名实体识别研究中，模型的训练仍然缺乏足够的数据支持。并且金融文本内容冗余，专业名词较多，难以有效识别出实体。针对上述问题，构建了金融命名实体识别数据集Ann-Finance，并提出一种基于RoBERTa-BGRU-CRF和对抗训练模型（AT-RoBERTa-BGRU-CRF）来提取实体。首先通过RoBERTa预训练模型得到词嵌入向量；然后通过对抗训练生成对抗样本；接着将词向量和扰动样本输入到双向门控循环单元网络，获得上下文的深层语义信息；最后通过条件随机场得到全局最优序列，输出结果。实验结果表明，在构建的数据集Ann-Finance上模型的F1值达到67.92%；而在Resume数据集和Weibo数据集上，模型在实验中的F1值也达到95.54%和71.72%，证明了该命名实体识别模型的有效性。更多还原

【Abstract】 Although deep learning technology has been widely used in financial named entity recognition,in the existing research,the training of the model still lacks sufficient data support. In addition,the content of financial texts is redundant,and there are many professional terms,so it is difficult to effectively identify entities. In response to the above problems,a financial named entity recognition dataset Ann-Finance is constructed,and an entity extraction based on RoBERTa-BGRU-CRF and an adversarial training model(AT-RoBERTa-BGRU-CRF)is proposed. First,the word embedding vector is obtained through the RoBERTa,then the adversarial sample is generated through adversarial training. Then the word vector and the perturbed sample are input into the bidirectional gated recurrent unit network to obtain the deep semantics of the context. Finally,the global optimal sequence is obtained through the conditional random field,and the result is output. The experimental results show that the F1 value of the model on the constructed dataset Ann-Finance reaches 67.92%. On the resume dataset and Weibo dataset,the F1 value also reaches95.54% and 71.72%,which proves that the effectiveness of named entity recognition models.更多还原