Citation Recommendation Based On Citing Time Preference And Argumentative Zoning Of User Query

With the development of science and technology,researchers have published millions of academic papers.In the process of scientific research activities,to better grasp the current research progress,people often need to cost a lot on literature surveys.To alleviate the information overload problem,academic search engines,academic social networking sites and literature management platforms have become the main ways for researchers to obtain information they need from the massive scientific data.Morevoer,academic paper recommendation systems are constantly proposed and put into use by the academia and industry.At present,there are two types of academic paper recommendation: one is based on user modeling to recommend academic papers that users are interested in.Another one is to recommend articles based on user queries,which are often used as references by the users.Relatively speaking,the second type of paper recommendation which is called as citation recommendation plays a more important role and is significant in saving time and energy when writing academic documents and reducing the rate of missing important documents.With the rapid development of big data storage and text processing technologies,metadata and fulltext information of academic papers are becoming rich and avaliable.With the continuous progress of traditional machine learning,deep learning and other technologies,researchers have made achievements in academic text mining,and citation recommendation task has ushered in development momentum.Existing citation recommendation studies have the following problems: First,the text representation method used currently is more traditional,and the neural network model is less applied in representation learning of user query and candidate citation.When using the ranking metrics to resort the recommendation list,the existing researches only applied some single ranking metrics without considering their ranking utility and the possibility of combination among different metrics.In addition,most existing citation recommendation tasks are static recommendations,without adding the time factor.Also,the semantic information mining for user query mostly stays in the stage of text representation learning without in-depth semantic analysis.To further improve the effectiveness of citation recommendation algorithms,this dissertation carries out studies based on query argument structure and citation time preference,which mainly includes the following four aspects:(1)firstly,this dissertation takes the citation recommendation as the task to classify the candidate citations,that is,to judge whether citations are cited or not.Before training the text classification model,recommendation algorithms need to represent the text and extract features.In order to find effective text representation methods in citation recommendation tasks,Chapter 3 compared some popular neural network models and further improved the recommendation performance through the selection of pre-trained language models,construction of neural network input and so on.(2)as more and more researches put forward various reference requirement indicators to sort the recommendation list,this dissertation uses the learing to rank algorithm to analyze the sorting effect of four currently popular indicators and the combinations between different indicators.In addition,Chapter 4 also used the article authority index to reorder the citation recommendation results based on the relevance of text content.(3)in order to solve the problem that the existing recommendation studies ignore the time dimension and fail to consider the preference of people's citation behavior in time,this dissertation proposes the citation recommendation method considering the time factor.Chapter 5 built a neural network model to predict the citing time preference based on user queries and applied citing time preference to effectively reorder the recommended results based on the similarity of text content.In order to further improve the prediction effect,this dissertation also tried a network structure with an attention mechanism.(4)in order to deeply analyze the semantic information contained in user queries,based on the previous research work,Chapter 6 added the argumentation structure information of user queries to citation recommendation task,and proposed a citation recommendation model considering argumentation structure of user queries.In order to train the model,this dissertation carried out the human labeling of papers from the field of biomedicine.In addition,this dissertation also put forward the citing parameter based on the publication time of the citation,and put it into the neural network model as a weight to improve the recommendation effect.After several investigations over citation recommended task,this dissertation makes utilizations of deep learning,machine learning technology to analyze text representation models,reference requirement indexes,user query argument structure types and citing time preference.Different models are put forward and we obtain the related conclusion through the experiment,and prove the validity of the proposed method.Through the above exploration of citation recommendation task,this dissertation systematically compares the utility of different neural network models for text representation in citation recommendation task and the ranking utility of different ranking indexes for recommendation results.For citation recommendation considering time factor,this dissertation provides a new idea,and the proposed citing time preference can be applied to other recommendation models based on user query,so as to conduct dynamic recommendation.Aiming at the deep semantic mining of user query,this dissertation proposes a citation recommendation algorithm considering argumentative structures.Theoretical basis,calculation model and annotated data for user query understanding in citation recommendation are all provided.