基于多示例学习的长文档检索  

Long Document Retrieval Based on Multi-instance Learning

作　　者：田媛 郝文宁 靳大尉 陈刚 邹傲 TIAN Yuan;HAO Wenning;JIN Dawei;CHEN Gang;ZOU Ao(Command&Control Engineering College,Army Engineering University of PLA,Nanjing 210000,China)

机构地区：[1]陆军工程大学指挥控制工程学院,江苏南京210000

出　　处：《无线电工程》2021年第9期886-892,共7页Radio Engineering

基　　金：国家自然科学基金资助项目(61806221)。

摘　　要：随着互联网信息的爆炸式增长,文档检索已经成为自然语言处理的热点问题。对于长文本检索,使用传统的基于词频的表示方法往往忽略了文本的语义信息,而使用嵌入模型进行文本表示,受输入长度的影响,长文本通常会被截断,此外,一些相似度计算方法会受到文本长度的影响。针对上述问题,提出将多示例学习框架用于文档检索中,以语义相对完整的句子为单位对文本进行切分,将文本表示成包,句子作为示例,通过示例之间的相关性来计算包之间的相关性得分,并将该得分与使用传统文档级检索即将整个文档作为一个单示例计算出的相似度得分相结合,从而检索出相关文档。在Med数据集上的实验结果表明,基于多示例的检索方法能在一定程度上提高文档检索的性能。With the explosive growth of Internet information,document retrieval has become a hot issue in natural language processing.For long text retrieval,the traditional representation method based on word frequency tends to ignore the semantic information of the text,and if the embedded model is used for text representation,the long text will usually be truncated due to the influence of the input length.In addition,some similarity calculation methods will be affected by the text length.A multi-instance learning framework is proposed to be used in document retrieval.The text is segmented by sentences with relatively complete semantics,and the text is represented as a bag.The sentences are taken as instances,and the similarity score between bags is calculated by the similarity between the instances,so as to retrieve the relevant documents.Experimental results on Med dataset show that the proposed method can improve the performance of document retrieval to a certain extent.

关 键 词：文档检索 多示例学习 相关性得分 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]