Research On Information Retrieval Ranking Optimization Methods

Recently,as the dramatic development of Internet,the data and information resources on the Web show an explosive increase trend,it's difficult for users to quickly search the relevant results from enormous quantity of Web data.As the advent and development of technologies about Information Retrieval,search engine has become the main way for users to look for information resources.However,the queries submitted by users may be very long or short,which will cause queries inaccurately to express users' search intention and vocabulary mismatch between queries and words in Web documents,which makes it more difficult for users to retrieve their desired information.Furthermore,in most cases,users are usually interested in the top results in the result list,while only parts of results can meet users' information needs.While the returned top results will contain much irrelevant information that is not related to users' search intention,thus causing the search results unsatisfactory.Therefore,how to improve the accuracy of top results in information retrieval system and users' satisfactory is the hot topic.To tackle these problems,there are two main methods to solve it,which includes: search result re-ranking and query expansion.Query expansion method is to dig terms relevant to users' information needs,then we use these terms to expand the original queries and use the expansion queries to perform the second retrieval.Search result re-ranking method is to utilize ranking model to re-rank the documents in initial result list.However,search result re-ranking is the most popular way to deal with these problems from practical perspective.The existing methods are taking consideration of relationships between documents,using similarities of documents to construct a graph for documents,combining with a center-node of graph or based on semi-supervised learning for search result re-ranking.However,most of them only consider relationships between documents,which may have some limitations.On the other hand,there exist a few researches combining search result re-ranking technique with query expansion technique.Based on the existing approaches,this paper makes two contributions in this thesis.1.A search result re-ranking method utilizing multiple-relationships is proposed.Based on relationship between documents,this approach also considers relationships between key words in the top documents in the result lists,as well as relationships between documents and key words in the top documents.Then we try to incorporate the three relationships into a normalized framework to re-rank documents based on the initial results.Experiments are conducted on information retrieval standard evaluation dataset.The experimental results show that this method are better than other methods based on relationship between documents,which can significantly improve accuracy of top results in the result lists.2.Based on search result re-ranking and query expansion,this paper proposes an optimization approach.At first,this method incorporates relationship between documents into a normalized framework and re-ranks documents based on the initial results.Next,we utilize the top documents in initial result list to construct expansion terms set and select candidate terms from this set to expand the original queries.Experiments results conducted on real dataset show that the proposed method can improve the precision for top results and users' satisfactory.