基于BERT的高校图书馆微信信息服务的命名实体识别方法

Named Entity Recognition Method of the University Library Wechat Information Service Based on BERT

[目的/意义]微信服务推文信息的命名实体识别是高校图书馆智慧服务领域的重要研究方向。在目前微信服务推文的命名实体识别领域中,存在高校图书馆微信服务领域的实体名称多样、特征信息提取不明确等问题。[方法/过程]针对以上问题,本文通过挖掘高校图书馆服务领域的微信服务推文文本,采用现有知识库和词典规则扩充句子语义信息,并在领域专家的指导下构建了一种高校图书馆微信服务推文文本的命名实体语料集ULICNER,实验数据集包含36035条文本语料数据,7大类16个小类。采用基于BERT-BiLSTM-CRF的模型研究高校图书馆微信推文文本的命名实体识别方法。该模型利用BERT网络将大规模文本生成具有语义特征的字符向量,接着对输入文本序列通过BiLSTM获取文本特征,最后连接CRF获得最佳的序列标记输出。[结果/结论]实验结果表明,本文模型在构建的数据集上取得良好的效果,其F1值为98.75%,准确率值为98.59%,召回率值为98.91%,模型可以运用于高校图书馆信息服务的实体识别任务。

[Purpose/Significance]Named entity Recognition(NER)of Wechat service information is an important research direction in the field of intelligent service of university libraries.At present,in the field of NER of Wechat service s tweets,there are insufficient problems such as entity name and feature information extraction in the field of the university library Wechat service.[Method/Process]For the above problems,this paper used the knowledge base and dictionary rules to expand the sentence semantic information by mining the wechat service tweet text in the field of university library service.Under the guidance of the experts,which constructed a corpus set of university library wechat service s NER of the ULICNER.The data set contained 36035 pieces of text corpus data,including 7 categories and 16 subcategories.Then,the model used BERT network to generate character vectors with semantic features from large-scale text,then obtained text features from the input text sequence through BiLSTM,and finally connected CRF to obtain the optimal output of sequence tags.[Result/Conclusion]The experimental results show that the model in this paper achieves good results on the constructed data set.The F1 value is 98.75%,the accuracy value is 98.59%and the recall rate is 98.91%.Finally,the model in this paper can be applied to the named entity recognition of university library information service.