Knowledge Base Citation Recommendation Based On Learning To Rank

Knowledge base Citation recommendation refers to filtrate and recommend documents automatically according to the entities included in the knowledge base.With the wide application of Knowledge Base in a variety of fields,knowledge base citation recommendation is becoming more and more popular,such as KBA(Knowledge Base Acceleration)in Text Retrieval Conference(TREC),which has a peculiar mission CCR(Cumulative Citation Recommendation),this paper is based on this mission.At present,the study of knowledge base citation ranking and recommendation focus on modeling it as a retrieval model in learning to rank field.This study came up with a method for citation ranking and recommendation based on learning to rank by contrasting several common used model about citation recommendation.Furthermore,the study summary the three key questions in knowledge base citation ranking and recommendation,which is:Query expansion based on an entity name in knowledge base;feature extraction of documents and entities;predicting model selection of relevance judgment between documents and entities.The main research contents and achievements of this study as follows:1.Putting forward an entity query expansion algorithm on the basis of associated semantic dictionary with word embedding.For the first step,achieving semantic dictionary based entity query expansion algorithm by using property feature in DBpedia.For the second step,accomplishing word embedding based entity query expansion algorithm by using WAF and word2vec algorithm respectively.For the third and last step,combining first-step algorithm with second-step algorithm to get the final result of entity expansion.2.Extracting rarely used features in the field such as semantic feature,syntactic feature and time feature in order to do relevance judgment in subsequent task.Building semantic feature by applying LDA and ESA algorithm to solve the question of polysemy.In addition,the study found that syntactic feature and time feature also have positive effect on knowledge base citation ranking and recommendation.3.On the basis of 1 and 2,achieving relevance judgment between documents and entity by applying point-wise,pair-wise and list-wise methods in learning to rank field.The experiment result of the study turned out the method(that is knowledge base citation ranking and recommendation based on learning to rank)is more effective than common used knowledge base citation ranking and recommendation algorithm.4.Proposing a linear model that combine logistic regression with random,forest classifier to accomplish learning to rank algorithm,which,helps to realize the relevance judgment finally.Also,the experiment result proved that the method which the study used is more efficient in solving issues of knowledge base citation recommendation than other common used learning to rank method.5.Designing and implementing integrated knowledge base citation ranking and recommendation system.The F1 value of this system operated on TREC KBA2014 promoted 19.8%compared to the baseline,which indicated that the method raised by the study is good at dealing with question of citation ranking and recommendation and is feasible.