Parameter Estimation for a Simple Hierarchical Generative Model for XML Retrieval

Abstract. This paper explores the possibility of using a modified Expectation-Maximization algorithm to estimate parameters for a simple hierarchical generative model for XML retrieval. The generative model for an XML element is estimated by linearly interpolating statistical language models estimated from the text of the element, the parent element, the document element, and its children elements. We heuristically modify EM to allow the incorporation of negative examples, then attempt to maximize the likelihood of the relevant components while minimizing the likelihood of non-relevant components found in training data. The technique for incorporation of negative examples provide an effective algorithm to estimate the parameters in the linear combination mentioned. Some experiments are presented on the CO.Thorough task that support these claims.

• Type: Conference
• Year: 2005
• Where: INEX
• Authors: Paul Ogilvie, Jamie Callan
• Keywords: Generative Model, Hierarchical Generative Model, INEX 2005, Information Management, Negative Examples