Term Function-based Citation Recommendation Of Academic Literature

In recent years,with the rapid growth of academic literatures,it becomes more difficult for researchers to find related works and do literature review.For this reason,citation recommendation research has gradually been concerned by scholars.Citation recommendation can automatically recommend suitable citations and references for a given paper or topic.With the aid of citation recommendation,authors can improve the efficiency of writing academic literature and avoid missing important relevant literatures.Traditional citation recommendation researches only take limited factors into consideration,such as text similarity,citation calculation,citation networks and so on.However,in this article,we make use of the semantic information of full text and citation content,aiming at recommending a higher quality and diversity paper list.Citation recommendation is a new research area,but some works has been done in recent years.This paper firstly provides a comprehensive literature review on its related research,and gives an overview on the motivation,classification,methods,evaluation,systems,hotspots and difficulties of citation recommendation.For exploring citation recommendation research from the perspective of term function,this paper firstly divides citation function of scientific articles into nine aspects based on previous works,they are Problem,Method,Problem+Method,Method+Problem,Datasets,Evaluation,Tools,Applications and Topic-irrelevant.After that,531 articles from information extraction,sentiment analysis,recommender system areas were labeled,result showing the first four aspects account for 90 percent,which make up the term function classification in this article.Meanwhile,a citation recommendation framework is presented,and recommendation experiment is carried out on the ACL Anthology dataset.To the best of our knowledge,this paper is the first work involving term function into citation recommendation as well as recommending articles at paragraph level,so the most commonly BM25 is used as baseline method in our experiment.Our experiment results showing that BM25 achieved 18.5%on Macro-F when the candidate recommendation list is 10 articles,compared to the baseline method,the term function weighted BM25 achieved nearly 6%improve on Macro-F,demonstrating the effectiveness of term function on citation recommendation.Meanwhile,when ? = 0.7 the ratio of content-based score and term function based score is 7:3,the experiment achieved the highest F1 score 24.5%.Our proposed method provides an innovative direction on how to improve the efficiency of citation recommendation.What's more,the citation recommendation at paragraph level also contributes to the automatic generation of literature review.