Research On Subject Classification And Interdisciplinary Measurement Of Periodical Literature From The Cognitive Perspective

As scientific problems become more and more complex,interdisciplinary research plays an increasingly prominent role in supporting cutting-edge science and innovation and finding solutions to complex social problems,which has aroused widespread concern in the scientific and technological circles.The most important task in interdisciplinary research is to determine the subject classification system.However,in most databases of English journals,the subject categories of papers are approximately represented by the subject categories of their journals,which leads to errors in the results of quantitative measurement.At present,the quantitative interdisciplinary measurement research can be divided into cognitive perspective and cooperative perspective.In the cognitive perspective,most of the research is based on the reference perspective,which maps the citation relationship between documents to disciplines,so as to conduct interdisciplinary measurement research.This ignores that the text content of periodical literature itself can more intuitively show the different disciplinary attributes it contains,and measuring the interdisciplinary degree from the perspective of text content is the most direct method of interdisciplinary measurement.Therefore,in this study,441,216 academic papers published in 4,727 journals in 2016,which can be retrieved from the Web of Science database,are selected from the cognitive perspective of combining text content and reference information,and natural language processing algorithm is introduced to classify journal documents based on the text content and reference information,so as to measure interdisciplinary,and the results are compared and analyzed with those from the perspective of text content and reference respectively,so as to show that the cognitive perspective of combining text content and reference is different from other single-view research results..This research mainly consists of the following three parts:(1)Collect and process the periodical literature data,label the subject categories of periodical literature according to the ECOOM subject classification system,extract the paper collection number,abstract and reference information of each periodical literature,and construct the experimental data sets of the first-class discipline and two disciplines category respectively.(2)Subject automatic classification of periodical literature is carried out from three perspectives: reference,single perspective of text content and cognitive perspective of combining text content with reference.Nine classical natural language processing algorithms are selected to classify periodical literature based on reference information;On the basis of SCIBERT model,SCIBERT-CNN-Bi LSTM model is constructed to classify disciplines based on text content;Introducing reference information to further deepen and improve,constructing the citation embedding SCIBERT-Attention model,and classifying periodical literature from the cognitive perspective.(3)According to the results of subject classification of periodical literature,the recall rate and probability matrix of subject classification are used to measure the first-level discipline and two disciplines interdisciplinary of text content perspective,the cognitive perspective of text content and reference respectively,and the results are compared with the interdisciplinary measurement results of traditional methods of reference perspective,so as to comprehensively explore the interdisciplinary of different disciplines with different granularities from different perspectives.The results show that:(1)the pre-training language model is the best in the subject classification of periodical literature,followed by deep learning model and machine learning model;(2)Citation embedding SCIBERT-Attention is the best classification method of all periodical literature disciplines,which improves the classification effect by about 3% compared with SCIBERT model.The F1 values of the first-level disciplines and sub-disciplines data sets are 83.93% and 63.68% respectively,and the classification effect is over 90% in individual categories.(3)The granularity of subject classification system will affect the subject classification effect of periodical literature,and the fine granularity is worse than the coarse granularity,and the classification effect of periodical literature under sub-disciplines classification system is lower than that under the first-level disciplines classification system;(4)The highest and lowest interdisciplinary subjects in the first-level disciplines are Biomedical Research and Physics respectively;The highest and lowest interdisciplinary subjects in sub-disciplines are Multidisciplinary Chemistry,Dermatology/Urogenital System respectively.Through empirical analysis and comparative study,this paper verifies the effectiveness of interdisciplinary measurement based on the automatic classification model of periodical literature from the cognitive perspective,which not only broadens the current research system of interdisciplinary measurement,but also provides new research ideas and methods for interdisciplinary measurement,which is helpful to understand interdisciplinary research from multiple angles and in depth,and also provides theoretical basis for the construction and development of disciplines for the government and related researchers..