Learning To Rank For Biomedical Information Retrieval

In recent years,with the continuous development of biomedical,the scale of data has also continued to increase,which makes it difficult for researchers to extract information manually.In order to satisfy the researcher's information requirement,information retrieval techniques for the biomedical field have been resolved.Different from traditional information retrieval,information retrieval in the biomedical field faces some specific challenges,most of that are due to the complexity and particularity of medical terms,which may lead to the retrieval of irrelevant articles.Research in the field of biological information retrieval usually has two problems to solve: First,the understanding of the user query is to retrieve the most similar document according to the input query;Second,set up the retrieval model,measure the similarity between the query and the document and give the ranking result.The above two challenges complement each other.A well-understood user query allows the search system to locate information requirements more accurately.This thesis focuses on the above issues from three aspects of research,including:By using learning to rank for technical research on optimizing,a model combining query improvement and query expansion is proposed.Based on the improvement of the query,our method uses LTR to reorder the query expansion words,so that the original query can retrieve the document with higher accuracy.Only by biological field resources,the query expansion can obtain extended words,but it can not accurately describe the degree of relevance between extended words and query words.Introduce LTR methods can fully consider the relevance of extended words and original queries.Improve the inaccurate problem of single-extension.The experimental results show that the query optimization technology based on LTR can effectively improve the efficiency of information retrieval in the biomedical field.For the 2017 TREC evaluation task,an information retrieval system for clinical medicine was established.Precision medicine aims to use genomic information to find more effective treatments for patients.The evaluation task is similar to the problem of determining a clinical diagnosis for a patient,and the evaluation focuses on providing clinical decision support for cancer patients who may influence the genetic variation of treatment options.According to the pertinence of the task,we propose a method of query expansion based on the special treatment of the data set and the theme,and filters the output results to achieve secondary sorting output.The experimental results show that the method of improving the query and filtering the results can effectively search biomedical literature and clinical trials.Using deep learning to perform query expansion techniques,a method of constructing extended word features using deep learning distributed vector representations is proposed.Based on the learning to rank,our method improves the extended word relevance annotation strategy.It uses word vectors to construct features,strengthens the expression of similarity between documents and queries,and is used to construct and optimize extended word ranking models.The experimental results show that our method has a high accuracy in the TREC public data set,which has great significance for deep learning in the IR field.The research work of this thesis is oriented to the information retrieval problem in the biomedical field.Mainly for query optimization,our method expands the query through learning to rank algorithms,deep learning models,and resources in the biomedical field,so that we can understand user intent more accurately and improve retrieval efficiency.