基于混合深度学习模型的科技文献自动综述模型构建研究  

Research on the Construction of Model for Automatic Review of Scientific Literatures Based on Hybrid Deep Learning Model

作　　者：马浩 崔运鹏[1] Ma Hao

机构地区：[1]中国农业科学院农业信息研究所,北京100081

出　　处：《情报理论与实践》2021年第9期176-182,168,共8页Information Studies:Theory & Application

基　　金：NSTL文献专项任务“下一代开放知识服务平台关键技术优化集成与系统研发”子任务“农业领域专业词表构建及跨文档自动综述生成算法研究”的成果,项目编号:2020XM05。

摘　　要：[目的/意义]在大数据时代,如何高效地进行科技文献知识组织与服务已成为图情领域的研究热点。因此,开展科技文献综述自动生成模型研究具有重要意义。[方法/过程]文章中,综述的自动构建分为两步:首先基于语步理论对输入文本的句子按语步类别进行识别与抽取,然后以各类语步的句子集为输入通过生成模型进行综述生成。研究分别基于SciBERT深度学习模型和Transformer网络构建了语步识别模型和综述文本生成模型。[结果/结论]语步抽取模型整体识别效果评价F1值达到87.12%,生成模型的生成效果与TextRank模型和BiLSTM模型相比在ROUGE-1、ROUGE-2和ROUGE-L三项评价指标上分别提高了4.5%、2.9%和3.3%。研究完成了科研文献综述自动生成任务的整体模型构建与实现。[Purpose/significance]In the era of big data,how to efficiently provide knowledge organization and service of scientific literatures has become a research hot spot in the field of library and information science.Therefore,it is of great significance to study the automatic generation model of scientific literatures review.[Method/process]In this paper,the review is constructed in two steps:firstly,the model identifies and extracts the sentences in the input text according to the rhetorical move theory.Then,with all kinds of sentence sets as input,the review is generated through the generation model.Based on SciBERT and Transformer,this paper constructs a move recognition model and text generation model.[Result/conclusion]The F1 value of the move recognition model reached 87.12%.Compared with TextRank model and BiLSTM model,the evaluation results of the generated model are improved by 4.5%,2.9%and 3.3%in the three evaluation indexes of ROUGE-1,ROUGE-2 and ROUGE-L,respectively.This paper constructs and implements the overall model of automatic generation task of scientific literatures review.

关 键 词：科技文献 自动综述 语步识别 深度学习模型 SciBERT 

分 类 号：G434[文化科学—教育学] G353.1[文化科学—教育技术学]