| Academic conferences are essential platforms for scholars to communicate and learn.Large offline academic conferences usually contain many academic reports in a short period,which need to be conducted simultaneously in multiple venues.Considering the constraints of report time and movement time between venues,it takes a lot of effort to select a group of reports that are interesting and do not conflict with each other.The existing paper recommendation methods can hardly meet these requirements.Therefore,this paper designs a set of report recommendation methods for large offline academic conferences considering the temporal and spatial constraints,including keyword extraction,report matching ranking and recommendation under the temporal and spatial constraints,and so on,to quickly recommend the reports for the participants.The main contents of the paper are as follows.Report title keyword extraction method: keywords are the primary basis for assisting conference participants in finding reports quickly.To extract common keywords from conference report titles more effectively,this paper proposes a keyword extraction model Ro BERTa-CRF based on BERT and conditional random fields.A web crawler is used to crawl the data of journal papers in a certain publication time range from several Chinese journal websites.The crawled data includes information such as titles,abstracts and keywords provided by the authors of the papers; the forward and backward matching algorithms are designed to match the keywords provided by the authors in the titles of the papers; the matching results are used as the accurate annotation of the title keywords,so as to generate the annotated corpus automatically.After model training and testing,the F1 value of Ro BERTa-CRF is 0.470,which is better than the current mainstream keyword extraction models.Report title matching and sorting method: based on the keywords extracted from the conference report titles or custom keywords,the matching degree of each title in the conference report title database is calculated,and all conference reports are ranked according to the matching degree condition.At present,the calculation methods related to report title matching mainly include methods based on string matching,methods based on word vector and methods based on feature interaction.The typical models of the three methods are selected for model training and comparative analysis,and the feature interaction-based method Interact-BERT significantly outperforms the other two methods in terms of experimental results evaluation indexes,thanks to its pre-training on large-scale text corpus and the underlying interaction mechanism of text features when calculating the matching degree.Therefore,the Interact-BERT model is selected as the report title matching method.A report recommendation model with temporal constraints: to avoid temporal conflicts within the report recommendation results,a Spatio-Temporal Constraints-aware Report Recommender(STCRR)model is proposed by combining Interact-BERT and a first-order Markov chain model.The spatio-temporal information of the reports is converted into parameters of the probability transfer matrix in the Markov chain,and the first-order Markov chain model is used to remove and adjust the report ranking results to form the final report recommendation results.The experiments show that STCRR can better handle the conflict between report time and location.The paper takes the 2021 China Academic Conference on Theory and Methodology of Geographic Information Science as an example,and calculates the title matching ranking and recommendation for 481 academic reports belonging to 15 venues over two days,eliminating the spatio-temporal conflicts and obtaining more reasonable recommendation results,which proves the effectiveness of the group recommendation method,which can enhance the experience of the participants and provide technical support for the smooth holding of large-scale offline conferences. |
PDF Full Text Request