Re-ranking Methods Research Based On Cloud Model Theory

In recent years, computer and Internet technology in the information construction of our country has made unprecedented popularization and development, which leads to the continual growth of information content. It’s a big challenge for information retrieval (IR) to improve the retrieval efficiency and the user experience in respect of the continuous expansion of massive information.This thesis first introduces the concept of document re-ranking and its research progress, and thoroughly analyzes the two main methods of document re-ranking, the statistics-based method and the semantic-based method. It has been found out that the two methods both neglect the uncertainty in native language. So this thesis researched the method of document re-ranking in information retrieval based on cloud model theory from the perspective of uncertain knowledge discovery.This thesis has proposed a re-ranking method based on cloud model by means of the uncertain knowledge discovery on the query terms level. The re-ranking method based on cloud model acquired the distribution of the key terms in the documents, used cloud model to convert the distribution into the uncertainty of the document representing the query on the query terms level, and then re-ranked the documents. And then this thesis proposed a re-ranking method based on concept hierarchy using cloud model by means of the uncertain knowledge discovery on the query level. The re-ranking method based on concept hierarchy using cloud model first acquired the uncertainty degree of using the document to represent the query on the query terms level, and then elevated the query terms level to the query based on the concept hierarchy theory using the cloud model synthesized algorithm, therefore acquired the uncertainty degree of using the document to represent the query on the query level, finally used the uncertainty of the two level’s to re-rank the documents.This thesis makes use of the methods proposed in this thesis in document re-ranking, and have designed and implemented the IR system successfully. The system firstly employs the three numerical characteristics of the cloud model to obtain the uncertainty of using the document in the first time research results to represent the query at two levels:the query terms level and the query level which is obtained from the cloud concept hierarchy promotion of the query terms level. And then re-rank the documents based on that uncertainty, returned the re-ranked documents to the user finally.This thesis performed experiments on the information retrieval test collections of NTCIR-5, and evaluated the results under TREC assessments. Experiments showed that the methods make improvements in both relax and rigid assessments and perform more excellent in the rigid assessment.