基于多粒度信息融合的气象知识命名实体识别

Meteorological Knowledge Named Entity Recognition Based on Multi-granularity Information Fusion

气象与人们的生活息息相关,运用命名实体识别算法抽取相关实体信息,对于构建知识图谱、问答系统等具有重要意义。由于气象科普知识存在大量专业词汇,普通的实体识别模型并不能很好完成识别任务。为此,论文构建了气象科普知识数据集,并提出了基于多粒度信息融合的气象科普知识命名实体识别模型MGTNER的算法。模型利用预训练模型、SoftLexicon结构的BiLSTM网络和键值记忆网络从数据集中以不同粒度提取语义特征信息,取得了很好的实体识别效果。在对气象科普知识数据集和公开Resume数据集实施的命名实体识别实验中,与几种基线模型进行了比较,结果表明论文提出的模型具有更好的识别效果。

Meteorology is closely related to people's lives.Named entity recognition algorithm is used to extract relevant entity information is of great significance for building knowledge graphs and question answering systems.Due to the existence of a large number of specialized vocabulary in meteorological science knowledge,ordinary entity recognition models cannot complete the rec⁃ognition task well.To this end,this paper constructs a meteorological popular science knowledge dataset,and proposes an algorithm of the meteorological popular science knowledge named entity recognition model MGTNER based on multi-granularity information fusion.The model uses the pre-trained model,the BiLSTM network with SoftLexicon structure and the key-value memory network to extract semantic feature information from the dataset with different granularities,and achieves good entity recognition results.In the named entity recognition experiments implemented on the meteorological science knowledge data set and the public Resume data set,the comparison with several baseline models is carried out,and the results show that the model proposed in this paper has a bet⁃ter recognition effect.