Combining Families of Information Retrieval Algorithms Using Metalearning

Abstract

This chapter describes some experiments that use metalearning to combine families of information retrieval (IR) algorithms obtained by varying the normalizations and similarity functions. By metalearning, we mean the following simple idea: a family of IR algorithms is applied to a corpus of documents in which relevance is known to produce a learning set. A machine learning algorithm is then applied to this data set to produce a classifier that combines the different IR algorithms. In experiments with TREC-3 data, we could significantly improve precision at the same level of recall with this technique. Most prior work in this area has focused on combining different IR algorithms with various averaging schemes or has used a fixed combining function. The combining function in metalearning is a statistical model itself which in general depends on the document, the query, and the various scores produced by the different component IR algorithms.