Machine Learning and Friends Lunch
Discriminative models for Information Retrieval
Ramesh Nallapati
UMass
Abstract
Discriminative models have been preferred over generative models in many pattern classification problems in the recent past owing to some of their attractive theoretical properties. In this paper, we view information retrieval as a binary classification problem as proposed by the binary independence retrieval model and explore the applicability of discriminative classifiers for IR in this framework. We have compared the performance of two popular discriminative models, namely the maximum entropy model and support vector machines with that of language modeling, the state-of-the-art generative model for IR. Our experiments on ad-hoc retrieval indicate that although maximum entropy is significantly worse than language models, support vector machines perform as well as or significantly better than language models in more than 60% of our runs.
We also argue that the main reason to prefer SVMs over language models is their ability to learn arbitrary features automatically as demonstrated by our experiments on the home-page finding task of TREC-10. Our experiments indicate that SVMs improve on the performance of the baseline language model by about 50% in mean reciprocal rank by learning from a variety of features.