摘要
针对当前中文地址要素解析方法复杂度高,且经常忽略细粒度地址要素语义特征的问题,该文提出一种顾及地理实体知识的中文地址要素解析方法。该方法结合现有地址要素分类分级框架:(1)扩展中文地址表达模型,丰富细粒度地址要素的语义特征表示;(2)设计顾及语义信息的五词位标注方法对地址进行标注,并使用小规模标注语料对ERNIE模型参数进行微调;(3)基于条件随机场模型计算标签组合概率并进行分析推理,利用两种学习策略进一步提升模型的解析性能。通过实验验证与对比分析,该文提出的方法获得的F1值超过97%,同时也大幅度减少了训练时间,其综合性能显著优于现有地址要素解析方法。
Aiming at the problem that the current Chinese address element parsing method is highly complex and often ignores the semantic features of fine-grained address elements,this paper proposed a Chinese address element parsing method taking into account the knowledge of geographic entities.This method combined the existing address element classification and classification framework,firstly expanded the Chinese address expression model,enriched the semantic feature representation of fine-grained address elements;then designed a five-lexeme tagging method that takes semantic information into account to tag addresses,and used a small-scale tagged corpus to fine-tune the parameters of the ERNIE model;finally,calculated the label combination probability based on the conditional random field model and conducted analysis and reasoning,and used two learning strategies to further improve the analytical performance of the model.Through experimental verification and comparative analysis,the F1 value obtained by the method proposed in this paper exceeded 97%,and the training time was also greatly reduced.Its comprehensive performance was significantly better than the existing address element analysis methods.
作者
杨彬
罗安
王勇
李朋朋
YANG Bin;LUO An;WANG Yong;LI Pengpeng(Faculty of Geomatics,Lanzhou Jiaotong University,Lanzhou 730070,China;Chinese Academy of Surveying and Mapping,Beijing 100036,China;Faculty of Geosciences and Environmental Engineering,Southwest Jiaotong University,Chengdu 611756,China)
出处
《测绘科学》
CSCD
北大核心
2023年第9期202-211,共10页
Science of Surveying and Mapping
基金
国家重点研发计划项目(2022YFB3904202)。