Role of Context in Matching and Evaluation of
XML Information Retrieval
Acta Universitatis Tamperensis No. 1624



By Paavo Arvola
September 2011
Tampere University Press
Distributed by Coronet Books
ISBN: 9789514484742
158 pages
$85.00 Paper original

This dissertation addresses focused retrieval, especially its sub-concept XML (eXtensible Mark-up Language) information retrieval (XML IR). In XML IR, the retrievable units are either individual elements, or sets of elements grouped together typically by a document. These units are ranked according to their estimated relevance by an XML IR system. In traditional information retrieval, the retrievable unit is an atomic document. Due to this atomicity, many core characteristics of such document retrieval paradigm are not appropriate for XML IR. Of these characteristics, this dissertation explores element indexing, scoring and evaluation methods which form two main themes:
1. Element indexing, scoring, and contextualization
2. Focused retrieval evaluation

To investigate the first theme, an XML IR system based on structural indices is constructed. The structural indices offer analyzing power for studying element hierarchies. The main finding in the system development is the utilization of surrounding elements as supplementary evidence in element scoring. This method is called contextualization, for which we distinguish three models: vertical, horizontal and ad hoc contextualizations.

The models are tested with the tools provided by (or derived from) the Initiative for the Evaluation of XML retrieval (INEX). The results indicate that the evidence from element surroundings improves the scoring effectiveness of XML retrieval.

The second theme entails a task where the retrievable elements are grouped by a document. The aim of this theme is to create methods measuring XML IR effectiveness in a credible fashion in a laboratory environment. The credibility is pursued by assuming the chronological reading order of a user together with a point where the user becomes frustrated after reading a certain amount of non-relevant material. Novel metrics are created based on these assumptions.

The relative rankings of systems measured with the metrics differ from those delivered by contemporary metrics. In addition, the focused retrieval strategies benefit from the novel metrics over traditional full document retrieval.

 

Return to Coronet Books main page