基于LERT和BiTCN的金融领域命名实体识别-《计算机技术与发展》

文章信息/Info

作者:: 陈雪松1; 王璐瑶1; 王浩畅2; 1. 东北石油大学电气信息工程学院,黑龙江大庆 163318;
2. 东北石油大学计算机与信息技术学院,黑龙江大庆 163318

Author(s):: CHEN Xue-song1; WANG Lu-yao1; WANG Hao-chang2; 1. School of Electrical and Information Engineering,Northeast Petroleum University,Daqing 163318,China;
2. School of Computer and Information Technology,Northeast Petroleum University,Daqing 163318,China

Keywords:: LERT model; financial field; named entity recognition; bi-directional temporal convolutional network (BiTCN); conditional random field (CRF)

摘要:: 针对传统的命名实体识别方法难以解决金融文本中一词多义且文本的语义特征提取不够充分的问题,提出了一种基于 LERT-BiTCN-CRF 的金融领域命名实体识别模型。首先,使用 LERT 模型对输入的金融文本进行预训练以生成相对应字符向量;然后,通过在 TCN 内部增加反向卷积层将其改进为 BiTCN,采用 BiTCN 对字符向量进行编码以提取字符向量的全局语义特征;最后,通过 CRF 进行解码以得到最佳的预测标签序列。在公开数据集 ChFinAnn 和自制数据集 FinanceNER 两个金融领域数据集上进行对比实验,该模型在两个数据集上的 F1 值分别达到了 84. 16% 和 92. 17% 。相较于其它模型,该模型在金融领域的命名实体识别任务中效果更好,表明该模型具有一定的有效性。同时又在公开的 Resume 数据集上进行对比实验,该模型 F1 值相较于基线模型 BiGRU-CRF 提升 2. 31% ,表明该模型具有一定的泛化性。

Abstract:: In order to solve the problem that the traditional named entity recognition method is difficult to solve the problem of multiple meanings of words in financial texts and insufficient semantic feature extraction of texts,a named entity recognition model in the financial field based on LERT-BiTCN-CRF was proposed. Firstly,the LERT model was used to pre-train the input financial text to generate the corresponding character vectors. Then,by adding a reverse convolutional layer inside the TCN,it was improved into BiTCN,and the BiTCN was used to encode the character vector to extract the global semantic features of the character vector. Finally,CRF was used to decode to obtain the best predicted label sequence. Comparative experiments were carried out on two financial domain datasets,the public dataset ChFinAnn and the self-made dataset FinanceNER,and the F1 values of the model on the two datasets reached 84. 16% and 92.17% ,respectively. Compared with other models,the proposed model has better effect in the named entity recognition task in the financial field,indicating that the model has certain effectiveness. At the same time,comparative experiments were carried out on the public Resume dataset,and the F1 value of the model was increased by 2. 31% compared with the baseline model BiGRU - CRF,indicating that the model has a certain generalization.