一种基于循环神经网络的电网客服语音文本实体识别算法

针对电力领域语音转写文本质量差,不能很好解决电网领域命名实体识别问题,以电网信息通信(information and communications technology,ICT)系统语音转写文本数据为研究对象,构建了一种基于双向长短期记忆(bi-directional long short-term memory,BiLSTM)神经网络融合条件随机场(conditional random field,CRF)面向电力文本特征的实体识别算法。通过与循环神经网络(recurrent neural network,RNN)等神经网络算法的对比验证:BiLSTM-CRF在电网ICT领域实体识别准确率达79%,F1值达80%,优于LSTM(long short-term memory)和其他RNN算法,并能较好地识别转写错误实体。该算法有效提升了领域语音转写文本的实体识别准确率,同时降低了领域语音识别技术成本,为电网客服领域信息检索、智能问答、个性化推荐等自然语言处理应用提供了高质量非结构化样本数据。

A Text Entity Recognition Algorithm Based on Recurrent Neural Network for Customer Service Voice of State Grid

In view of the poor quality of voice transcribed text in the field of power grid,which can not solve the problem of named entity recognition,this paper takes information and communications technology(ICT)voice transcribed text data in power grid as the research object.A domain voice transcribed text entity recognition algorithm based on bi-directional long shortterm memory(BiLSTM)is constructed,combining with conditional random field(CRF),which is verified by comparison with recurrent neural network(RNN)and other neural network algorithms.Results:In the field of ICT,the accuracy of entity recognition is 80%,and the F1 value is 79.56%,better than long short-term memory(LSTM)and other RNN algorithms.Conclusion:It effectively improves the accuracy of entity recognition of domain voice transcribed text,reduces the cost of domain voice recognition technology,and provides a high-quality unstructured sample base for natural language processing tasks such as ICT customer service domain information retrieval,intelligent Q&A,personalized recommendation,etc.