Extended structural relevance framework: a framework for evaluating structured document retrieval

  • Authors:

  • M. Sadek Ali;Mariano Consens;Mounia Lalmas

  • Affiliations:

  • Bahen Center, University of Toronto, Toronto, Canada;Bahen Center, University of Toronto, Toronto, Canada;Yahoo! Research, Barcelona, Spain 08018

  • Venue:
  • Information Retrieval
  • Year:
  • 2012

Quantified Score

Hi-index 0.00

Visualization

Abstract

A structured document retrieval (SDR) system aims to minimize the effort users spend to locate relevant information by retrieving parts of documents. To evaluate the range of SDR tasks, from element to passage to tree retrieval, numerous task-specific measures have been proposed. This has resulted in SDR evaluation measures that cannot easily be compared with respect to each other and across tasks. In previous work, we defined the SDR task of tree retrieval where passage and element are special cases. In this paper, we look in greater detail into tree retrieval to identify the main components of SDR evaluation: relevance, navigation, and redundancy. Our goal is to evaluate SDR within a single probabilistic framework based on these components. This framework, called Extended Structural Relevance (ESR), calculates user expected gain in relevant information depending on whether it is seen via hits (relevant results retrieved), unseen via misses (relevant results not retrieved), or possibly seen via near-misses (relevant results accessed via navigation). We use these expectations as parameters to formulate evaluation measures for tree retrieval. We then demonstrate how existing task-specific measures, if viewed as tree retrieval, can be formulated, computed and compared using our framework. Finally, we experimentally validate ESR across a range of SDR tasks.