Research On Parallel Learning To Rank Method Based On Topic

In information retrieval,Learning to rank(LTR for short)is a popular ranking technology.Currently,researches on LTR are made on the condition of the independence of document,and they mainly focus on how to precisely predict the relevance score between single document and query.However,they ignore the relationships among candidate documents,which result in redundant information in ranking results.In addition,existing LTR method handle all types of queries by resorting to a unified ranking model,ignoring the differences of queries and failing to specifically dispose the query,which reduces the accuracy of the results.The paper first discusses that it is necessary to consider the differences between queries.On this basis,then this paper further study how to break the independence of the document hypothesis and models relations among candidate documents in the process of training the ranking model.The detailed research content is shown as following:In the off-line phase,this paper jointly consider the relationship among documents and the differences of queries.Firstly,transform queries into the form of query eigenvector,partition the query sets according to the similarities among queries using the cluster approaches,forming different training subsets,and each training subset is constructed as a ranking model.Secondly,taking the relationships among documents into consideration,a relationship ranking learning model is proposed in the training process of sub ranking model,which makes the construction of model more specific and more adjustable.Finally,the score of a document depends on the relevance of its contents and queries,and the relationship between itself and all the previous documents that have been ranked.Based on these observations,the corresponding ranking function and loss function are established.As for the issue of how to tackle with the new arrival queries in the on-line phase,a parallel ranking framework is presented.A candidate ranking model selection method first proposed by choosing k ranking models which are most good at the pending query,and using the k ranking models to score the related documents of pending queries.For different types of queries,we can choose a more suitable sub ranking model and dispose it with this approach help,enabling us partition different ranking models into different points and dispose the queries parallely.Eventually,list the results for each of the candidate ranking model generated,employing a scoring function based on the weight to produce the final list of results.The experiments on standard dataset LETOR show that the approaches we propose,which is the improvement compared to the existing LTR methods,not only can effectivelyincrease the accuracy of the ranked results but solve the problem of diversification of search results to a certain extent.