Bagging to find better expansion words

Abstract

The supervised learning has been applied into the query expansion techniques, which trains a model to predict the “goodness” or “utility” of the expanded term to the retrieval system. There are many features to measure the relatedness between the expanded word and the query, which can be incorporated in the supervised learning to select the expanded terms. The training data set is generated automatically by a tricky method. However, this method can be affected by many aspects. A severe problem is that the distribution of the features is query-dependent, which has not been discussed in previous work. With a different distribution on the features, it is questionable to merge these training instances together and use the whole data set to train one single model. In this paper, we first investigate the statistical distribution of the auto-generated training data and show the problems in the training data set. Based on our analysis, we proposed to use the bagging method to ensemble several regression models in order to get a better supervised model to make prediction on the expanded terms. We conducted the experiments on the TREC benchmark test collections. Our analysis on the training data reveals some interesting phenomena about the query expansion techniques. The experiment results also show that the bagging approach can achieve the state-of-art retrieval performance on the standard TREC data set.