Research On Query Performance Prediction In Search Result Diversification

For a given query,the search engine first analyzes it,then retrieves the documents on the pre-built index,and finally produces a sorted list of documents according to a ranking algorithm.In order to evaluate the performance of the returned document list,it usually requires manual judgments,which are time-consuming and costly.It has great practical significance to develop the query performance prediction technique without any manual judgment.For search engines,improving the performance of some difficult queries is particularly important.If we can predict the performance of queries and identify those queries with poor performance,then we will take the necessary remedial measures to improve the quality of these query results,and certainly,it will improve the user's satisfaction.Therefore,it is a meaningful work to design an effective query relevance performance prediction method,which is also a research direction in the field of information retrieval at present.A query often contains multiple sub intents,and different users often have different intentions for the same query.In order to allow more users to get better search experience,the search engine should make the highly-ranked search results cover as many sub-intents as possible.This process is called diversification,and the performance of the diversified results is referred to as the diversification performance.Under the background of the search result diversification,in order to avoid returning the query results with poor diversification performance to the user,the search engine needs to predict the diversification performance of the search results in response to the query.Therefore,this paper studies the prediction for the query diversification performance and to the best of our knowledge,there is no research found in the literature on the diversification performance prediction so far.This dissertation mainly focuses on the following aspects:(1)For predicting query relevance performance,from the view of predicting the class of the query difficulty(hard,median,or easy),we propose a method based onthe support vector machine to predict the class of the query difficulty.The experimental results show that the proposed method has good prediction effect,especially in the prediction of difficult queries.(2)In predicting the diversification performance of the search results in response to the query,we have used 5 predictors,and tested these predictors upon the runs from TREC Web Track 2010-2011 Diversity Task.The experimental results show that the presented predictors are effective.(3)We analyze the impact of different ways to obtain the search results of sub-intents in response to a given query on the effectiveness of the diversification predictors.The diversification predictors need to analyze the search results of sub-intents and this can be done from the external resources or from the intermediate results in the search result diversification directly.The effectiveness of proposed predictors is tested on a group of diversified results.The experimental results show that the performance of proposed predictors is still better than that of the traditional ones.