气象文本推荐研究  

METEOROLOGICAL TEXT RECOMMENDATION

作　　者：梅钰[1] 唐卫[1] 王慕华[1] 王阔音 Mei Yu;Tang Wei;Wang Muhua;Wang Kuoyin(Public Meteorological Service Center,China Meterological Administration,Beijing 100081,China)

机构地区：[1]中国气象局公共气象服务中心

出　　处：《计算机应用与软件》2019年第8期138-144,共7页Computer Applications and Software

基　　金：国家自然科学基金项目(41871020)

摘　　要：气象文本是国家气象部门面向公众发布的气象信息,具体包括预警、预报、专报、公报、提示等类型。现有文本生产需要人工编写审核,效率不高,而全自动文本生成主要依赖模板、形式比较固定。针对这个现状,提出气象文本推荐思路并给出具体实现方法。气象文本推荐读入用户输入信息,自动推荐后续相关文本供用户选择,提升编写效率及质量。该方法分为两步:进行气象要素抽取,替换得到模板文本;基于模板文本构建邻居子句生成模型。要素抽取使用CRF序列标注模型,文本生成利用Seq2Seq模型。基于公开预警文本的实验结果表明:利用CRF进行要素抽取平均准确率超过90%,基于Seq2Seq模型的生成方法在BLEU值上达到12.2,准确率达到65%。Meteorological text is a kind of meteorological information released by the national meteorological department to the public,including early warning,forecast,special report,bulletins,prompt and so on.Existing text production needs manual editing and verification,which is inefficient,while automatic text generation mainly relies on templates and forms are relatively fixed.In view of this situation,we put forward the idea of meteorological text recommendation and gave the implementation method.Meteorological text recommendation read in user input information,automatically recommended subsequent relevant text for user selection.It improved the efficiency and quality of writing.The method was divided into two steps: one was to extract meteorological elements and replaced them to get template text;the other was to build a neighbor clause generation model based on the template text.Factor extraction used CRF sequence annotation model,and text generation used Seq2Seq model.The experimental results based on open early warning text show that the average accuracy of factor extraction using CRF is over 90%,and the BLEU value of the method based on Seq2Seq model is 12.2,and the accuracy is 65%.

关 键 词：气象文本推荐 自然语言处理 信息检索 机器学习 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]