基于BERT-CBG-BiLSTM-CRF的羊养殖命名实体识别

Sheep breeding named entity identification based on BERT-CBG-BiLSTM-CRF

羊养殖知识多以文本的形式记录存储,知识量大、碎片化程度严重。为了改善构建羊养殖知识图谱时命名实体识别效果不佳的问题,本文的羊养殖文本命名实体识别模型将双向门控循环单元与卷积神经网络相结合,模型通过BERT预处理进行文本向量化处理,处理结果在CBG层通过训练字词向量,得到初步提取的上下文语义和词语语义,连接双向长短期记忆网络;条件随机场最终得到最大概率的输出序列。实验对特征、产地、建设、经济价值、品种、产区环境6类实体进行识别,最高F1值为95.86%。

Sheep breeding knowledge is mostly recorded and stored in the form of texts,which has the characteristics of large amount of knowledge and serious degree of fragmentation.In order to improve the problem of poor recognition of named entities when constructing sheep breeding knowledge graphs,the named entity recognition model of sheep breeding text in this paper is an optimization model that combines two-way gated circular units with convolutional neural networks.The model performs text vectorization processing through BERT preprocessing,and the processing results are trained in the CBG layer to obtain the contextual semantics and word semantics of the initial extraction,and then connect the two-way long-term short-term memory network;the conditional output sequence with the airport finally obtains the maximum probability.In this paper,six types of entities were identified experimentally for characteristics,origin,construction,economic value,varieties,and production area environment,and the highest F1 value was 95.86%.