Search Result Refinement Based On Semantic Relation Recognition

Since current commercial information retrieval systems are based on keyword matching, there are lots of results that do not meet the user's requirements, while the keywords of queries and documents are matched, which leads to the unsatisfactory performance. A new model was proposed for search result refining based on semantic relation recognition. This approach could effectively filter out the irrelative search results by recognizing the semantic relation among the keywords of documents retrieved from the original query. It correctly made up for the defects of keyword-based retrieval. Base on this model we built a retrieval system; experimental evaluation was done on the semantic relation of object by using this system.This paper mainly researched on the model and algorithm of the search result refinement base on semantic relation recognition. Our work was as following:First, we discussed the classical retrieval model and conceptual graph based retrieval model, and the method of search result refinement, then proposed a search result refinement model base on the semantic relation recognition. In our approach the initial search results were refined by filtering out the irrelative documents which the semantic relation among the keywords in the documents does not match with the one in the query. The model was based on the current commercial search engine, and using the semantic relation among the concepts, it is really a retrieval method on semantic level.Then, we build a system to implement the retrieval model on the computer. We use the SVM method to learn the semantic relation among the keywords and then use it to classify the original retrieved documents. In case the semantic relation among the keywords in the document is not same as the one in the query, this document will be discarded. We improved the precision of retrieval by selecting the features and parameters of the SVM by the method of trial and error and making use of the hash table to speed up the matching.Finally, we made an experiment to compare our retrieval model with the original search engine. Original search results of the dozen pair of keywords having the specified"(motion) object"relationship from Google after manually analyzed and classified were used as experimental data. Experimental results show that our approach can effectively filter out the irrelative search results on the most of the user queries, and for every semantic relation, we can learn it by only using a handful of training examples.The innovations of this paper are as below:In our retrieval model the keyword matching and semantic relation matching were divided into two stages. First we match the keyword; it can be achieved by any classical retrieval model. Then we match the semantic relation among the keywords. Thus, the query and document were matched on semantic level gradually.