基于评审意见的科技论文要点识别与利用

Identification and Utilization of Key Points of Scientific Papers Based on Peer Review Texts

科研用户查找文献往往瞄准特定任务,如寻找选题、方法、结论等,面对检索出的大量文献需要甄别要点,判断价值,这一过程既需要熟悉专业知识又耗时费力。论文评审意见中蕴含了同行专家的权威看法,包括对论文要点和参考价值的揭示,可以为满足上述需求提供有效帮助。本文以论文评审意见为对象,围绕科研活动中的典型要素定义评审意见中的要点类型,通过有监督学习方法提取评审意见所述的论文要点,不但为论文提供了结构化要点概括,还能用于辅助文献检索。本文采集《心理学报》期刊2014年年初至2020年年底发表的549篇论文及其对应的评审意见,将其中概括信息划分为概述、方法、结果和亮点4种要点类型,用SVM(support vector machine)、FastText、TextCNN(convolutional neural networks)及BiLSTM(bi-directional long short-term memory)4种方法训练分类模型并比较效果。研究结果表明,BiLSTM方法对要点识别效果最佳,在5次交叉检验中的平均识别准确率达到91%。要点中的亮点进一步分为选题、价值、方法和写作4种类型,采用SVM方法识别,F1值达到85%。作为对研究结果的应用,本文采用识别出的论文要点辅助对论文的理解,并按亮点做检索结果分类,改进了论文检索的组织与服务形式。本文创新之处在于:①提出了从评审意见中挖掘论文要点的研究问题,制定了要点类型的框架和层次;②将识别要点转化为分类问题,通过比较评价找到综合最优的识别方法;③实现了基于要点的检索结果分类组织,帮助用户理解论文并进行价值判断。

Scientific researchers often aim at specific tasks when searching literatures,such as seeking topics,methods,and conclusions.However,distinguishing the numerous key points of scientific papers and judging their value is time consuming and laborious.The task also requires extensive professional knowledge.The peer review contains the disclosure of the paper’s key points and the authoritative evaluation of the reference value of the paper,which can effectively help in meeting the aforementioned needs.This study considers the peer review as the object,defines the key point types in the review on the basis of the typical elements in scientific research activities,and extracts the key points of the paper described in the peer review through supervised learning methods,which not only provide a structured summary of the key points of the paper,but can also be used to assist the literature retrieval.This research collected 549 papers published in Acta Psychologica Sinica between 2014 and 2020 and their corresponding reviews.Four types of key points are defined:general information,methods,results,and highlights.Then,four classification models are trained using SVM(support vector machine),FastText,TextCNN(convolutional neural networks),and BiLSTM(bi-directional long short-term memory)to compare the results.Experiments show that the BiLSTM method has the best efficacy in key points recognition,with an average recognition accuracy of 91%across five tests.The highlights in the key points are further categorized into four:topic selection,value,method,and writing,which is then subdivided by the SVM method,with an F1 value of 85%.Similar to the application of the research results,this study also uses the recognized key points to facilitate in-depth understanding of the scientific paper,classifies the search results on the basis of the highlights,and improves the organization and service form of paper retrieval.This study’s contributions are as follows:(1)introducing the research problems of mining the key points of a scientific paper from the peer review and constructing the framework and hierarchy of the key points;(2)transforming the key points recognition into a classification task and comparing various classification methods to determine the comprehensively optimal method;and(3)achieving the classification organization of retrieval results on the basis of key points and assisting users in understanding and judging the results.