Massive query expansion for relevance feedback.

摘要

A major task of any information retrieval system is to elicit information from the user about what exactly the information need of the user is.; This thesis explores the process of relevance feedback, in which the user gives the system judgements on whether particular documents are useful, and the system reformulates the user's query based upon the content of those documents. In particular, the thesis focuses on massively expanding the original user query by adding terms occurring in the relevant documents.; The thesis shows that hundreds of terms can be added, producing an effectiveness increase of 20-30% over using just the original query. Various optimization techniques are examined, one of which, Dynamic Feedback Optimization, results in a further effectiveness increase of 10-15%.; The relationship between amount of relevance information used and the gain in effectiveness is also explored. The experimental evidence shows that the effectiveness increases in proportion to the log of the number of terms added, up to a point of diminishing returns, and in proportion to the log of the number of known relevant documents.; Finally, experiments show that massive expansion can improve effectiveness by 10-20% even in the absence of user relevance judgements when terms are added from the top retrieved (unjudged) documents of an initial search.