SAPHIRE: A CONCEPT-BASED APPROACH TO INFO RETRIEVAL

The major goal of this project is to develop and evaluate innovative
approaches to information retrieval (IR) in the biomedical domain.
Building on the initial work done by the Principal Investigator (PI)
with the SAPHIRE Project and taking advantage of the efforts of the
Unified Medical Language Systems (UMLS) Project, we aim to design and
test new methods for automated indexing and retrieval. The underlying
thesis of the SAPHIRE approach to IR is movement of information
representation from the level of terms to that of concepts. Terms, such
as MeSH entries, are surface string representations of underlying
concepts. A problem with their use in representing concepts is that
they cannot account for the different ways a concept may be expressed in
medical texts or information system queries. SAPHIRE is a first step in
the direction of concept-based IR, and we plan to investigate several
enhancements to this approach. The major goal will be achieved with six
separate but interrelated tasks: 1. Develop methodology for evaluation of IR systems in laboratory and
clinical settings. 2. Assess the utility of computational linguistic approaches to concept
discovery in text using constrained natural language processing and
knowledge base construction. 3. Refine strategies for automated indexing of a wide variety of textual
material, including abstracts, full text of articles, textbooks, and
hypertext. 4. Explore different user interfaces, aiming to allow optimal retrieval
for both novice and expert users. 5. Assess the use of semantic relationships between concepts in indexing
and retrieval. 6. Integrate the SAPHIRE approach with other programs, such as the CODEX
system and Explorer-2, and scale up to large text collections. In the course of the project we will create an IR system that will help
meet the information needs of busy health care providers. Such a system
should have a diverse variety of content available as well as quality
indexing to represent content accurately. It should also feature
retrieval capability that is fast and easy to use. In this grant, we
propose to iteratively build an IR system that utilizes concept-based
probabilistic indexing and retrieval, and evaluate it each step along
the way in laboratory as well as real world settings.