- Editorial
- In This Issue
Bert R. Boyce
- Research
What Is Information Discovery About?
H. A. Proper and P. D. Bruza
Proper and Bruza's believe that information discovery is an attempt
to
broadly model information retrieval outside the context of any operational
retrieval model. They define user need in terms of supply and demand for
something undefined called infions, or information particles. Relevance is
meeting a set of requirements stated in the same terms. Aboutness is
apparently a relation on a set of key words, but satisfaction is a
relationship on information carriers and their logical descriptions. A
distinction is made between system and user satisfaction, which will be
close in factual descriptions but not necessarily so in cases of aboutness.
Text Segmentation for Chinese Spell Checking
Kin Hong Lee, Qin Lu, and Mau Kit Michael Ng
Since Chinese text has no natural delimiters text must be segmented
into
valid words before error correction can take place. Many words are
represented by single characters but others require multiple character
strings. Lee, Lu, and Ng. test the Block of Combinations (BOC) segmentation
method which uses a 60,000-word dictionary with the 2,000 most frequently
used words grammatically tagged. A user dictionary for adding words not
predefined is available, and otherwise unidentified words are stored in a
temporary file. The 200 most frequent single character words are accepted
but others are suspected to be errors and presented to the user for
clarification with similar words as suggestions. Since the number of
possible segmentations increases rapidly as the number of characters grows,
a sliding five-word window is used rather than a complete sentence. The
procedure is more accurate than another method which takes about the same
computational time.
A Fuzzy Genetic Algorithm Approach to an Adaptive Information Retrieval
Agent
Maria J. Martin-Bautista, Maria-Amparo Vila, and Henrik Legind Larsen
The Genetic Information Retrieval Agent Filter (GIRAF) is a software
agent, tested here by Martin-Bautista, Vila, and Larsen, that can work
offline to filter and rank retrieved documents from an Internet search
engine. Query terms are extracted from evaluated initial search documents
or from ideal documents provided by a user, and each of these terms is
given a weight that is the average of its occurrence frequency in all
documents analyzed. One of four gene types is assigned randomly to each
term to form a triple with its weight, and chromosomes (strings of these
gene triples) are randomly formed. Type one genes use as their weight a
number of occurrences of the term in a document that will give complete
satisfaction (a higher number reducing satisfaction). Type two genes are
satisfied completely by documents that have no occurrences of the term.
Type three genes use the weight as a traditional threshold with complete
satisfaction achieved if the number is reached or exceeded. Type four
divides each document into three parts and will be satisfied by any of the
parts. Chromosomes are ranked by their similarity to relevant documents and
modified by choosing parents where the first is random and the second
higher than the first and preforming a gene crossover. Mutation also
randomly occurs. New chromosomes cause those at the bottom of the list to
be removed. Tests with virtual users indicate that type three genes work
best, except that when user profiles are permitted to change during the
process, type one gains some advantage
A Distance and Angle Similarity Measure Method
Jin Zhang and Robert R. Korfhage
The typical angularity measure (cosine) identifies documents whose
index
term distributions are similar. Despite this similarity, they may be far
apart in the document space if the level of detail in the discussion of the
topics is different. Document similarity with a distance measure depends
upon the length of the hypersphere radius from the reference point to the
document. Zhang and Korfhage present a similarity measure which combines
distance and angularity based measures. This measure ``s'' is the product
of two parameters, a and c, where ``a'' has the negative radius as an
exponent (a typical distance measure) and ``c'' (between 0 and 1) has the
value of the angle divided by the maximum value of the angle as an
exponent. The maximum value of ``s'' is equal to the distance-based measure
and the minimum is smaller but in the same position as the cosine measure.
Varying ``a'' and ``c'' will reflect a user's emphasis on distance or
angularity.
DARE: Distance and Angle Retrieval Environment: A Tale of the Two
Measures
Jin Zhang and Robert R. Korfhage
In a second paper, Zhang and Korfhage present a visualization model
which can display both distance and angle measures simultaneously and
handle both conjunction and intersection. Changing the slope and position
of a straight line in the visualization space results in a modification of
the threshold-defined contour in the document vector space and thus expands
or contracts the scope and emphasis of the retrieved set.
- Perspectives Issue on...Visual Information Retrieval Interfaces
Introduction and Overview: Visualization, Retrieval, and Knowledge
Mark Rorvig and Lois F. Lunin
This Perspectives issue is assembled to provide an historical
background
to visualization in information retrieval. It is a review of the
assumptions and technology configurations by which the current literature
may be interpreted. The techniques of the authors of this issue differ, but
all treat their techniques as manuals of description flowing from a history
of common mathematical and technical influences. All technologies have
histories of development. The historical forces of visualization frame the
current efforts and comprise the field in which new problem dimensions are
addressed. No field of scientific inquiry emerges without a background.
This issue adds to the depth necessary for the study of visualization by
new students and new scholars.
The NASA Image Collection Visual Thesaurus
M. E. Rorvig, C. H. Turner, and J. Moncada
The first visual interface to a collection was designed and
implemented
at the Johnson Space Center of NASA in the years 1988-1992. In this
interface, described in the article entitled, "The NASA Visual Thesaurus,"
the Rorvig et al assumed that the task of inferring images from terms and
terms from images would introduce invariance in image indexing. The system
remained in use for two years, but eventually failed because no automatic
method to assign terms to images could be discovered, and the manual cost
of such term assignment was too great to be supported.
Rorvig et al attempted to use image descriptions clustered by cosine
vector methods to identify a unique image for every thesaurus term. The
candidate images suggested by this method were often heartbreakingly close
to the mark. But close was not good enough. These developments were
described in detail by Seloff (1990). Although the Seloff article has been
widely cited, the initial article which specified the design parameters for
the system of his report has remained unpublished. It appears in this
article in the form originally presented at the ASIS mid-year conference of
1988. The article is significant because it represents the first
identification of the components of a visual interface.
Visualizing Science by Citation Mapping
Henry Small
The article by Henry Small of the Institute for Scientific
Information
addresses the two decade long historical use of visualization techniques in
calculating the relationships among scientific fields by their patterns of
co-citation. Small begins with the simplest of algorithms as conceived
within the computational limitations of the 1970's and ends with the most
ambitious ones presently available through Sandia National Laboratories
(SNL). In this article, students and scholars will find algorithms
applicable to many different aspects of the co-citation problem, as Small
frankly describes the research paths that were successful and led to
further enhancements as well as the ones which were eventually discarded
either because of their inefficiency in computation, or their failure to
yield truthful insights validated by earlier techniques. Many of these
algorithms may be transplanted to address similar problems with data that
may be encountered by researchers who require some intermediate processing
alternatives.
The Ecological Approach to Text Visualization
James A. Wise
The article by Jim Wise of Integral Visuals Corporation details the
technical advances of researchers at the Pacific Northwest National
Laboratories (PNNL) over an intense five year history of development.
Wise's "The Ecological Approach to Text Visualization" offers a rich
archive of techniques. This descriptive tour de force begins with the most
brutal techniques (e.g., vectors of length 200K analyzed through
Multidimensional Scaling (MDS)) and, in completely clear and intelligible
detail describes the short cut methods developed for computational
efficiency. These efforts have resulted in the presently available
commercial product "ThemeMedia" offered through a subsidiary of the Smaby
Group of investors (which purchased rights to further develop the
technology). Among the highlights of this article is a description of the
discovery that single and multiple link cluster centroids can be used to
approximate the full text collections originally required for visual
display. Additionally, the transformations of the 2d dot displays to the
present terrain models simultaneously developed at PNNL and SNL are
described in sufficient detail for engineers to reproduce the same
progression of results.
Interactive Graphical Queries for Bibliographic Search
Martin Brooks and Jennifer Campbell
Brooks and Campbell describe the translation of interactive boolean
interfaces with data to a visual display. The "islands" interface which
they illustrate harnesses the power of visualization to the process of
commercial text retrieval. It is a fact that students are still mystified
by these processes. One need only while away a few minutes on any college
campus to realize that most persons still have not got a clue about the
meaning of term conjunction and its impact on search results. Anyone who
has ever performed a boolean search will be able to examine the effect of
visualization on this process, and this article is offered to permit a
broad view of practice changes which can be expected in future systems.
"islands" may not be the ideal interface, but something like it, to
paraphrase an advertising slogan, will be "...coming soon to a library near
you."
A Collection of Visual Thesauri for Browsing Large Collections of
Geographic Images
Marshall C. Ramsey, Hsinchun Chen, Bin Zhu, and Bruce R. Schatz
In Ramsay et al, earth observing images are parsed as texts are
parsed.
These authors use Gabor filters to combine like terrains. No clearer
description of this process is available in the present literature. A Gabor
filter yields textures. By segmenting images into component texture
boundaries, search classes may be derived without resorting to textual
description. This technology thus succeeds where the NASA effort by Rorvig
et al failed. The results reported in this article are concrete and
verifiable; indeed, anyone who has ever traveled over Arizona highways can
authenticate these data. The authors acknowledge the contributions of the
Alexandria Digital Libraries Project, particularly the work of Manjunath
and Ma (1996), but claim their own extensions to this work as well.
Conference Notes--1996: Foundations of Advanced Information
Visualization for Visual Information (Retrieval) Systems
Mark Rorvig and Matthias Hemmje
One of the landmark developments in visual retrieval occurred at a
workshop held in Zurich in the summer of 1996 in conjunction with the
Association for Computing Machinery's Special Interest Group on Information
Retrieval Annual Meeting. For the first time, both European and North
American interests were represented in the development of criteria for
evaluation of visual information retrieval. Among the Europeans, the newly
formed FADIVA (Foundations of Advanced Information Visualization) group
played the dominant role. The workshop report reproduced in this issue has
been widely circulated, but never before published. This conference led to
the first visualization of native TREC/Tipster data as a prelude to formal
visual information retrieval evaluation strategies (Rorvig and Fitzpatrick,
1998; Rorvig, 1998).
[Robert Korfhage, one of the intellectual fathers of the visual information
retrieval effort in both Europe and North America, has contributed his
bibliography on this issue. The bibliography is comprehensive for all work
in this field c. 1997. Such documents are of interest in determining the
scope of future advances. In 1997, this was the known world view of this
area of scholarly effort. For practical use in permitting users to copy
this bibliography, it is available through the ASIS SIGVIZ website
where future editions may
be conveniently updated.]
- Book Reviews
Foundations of Library and Information Science, by Richard E. Rubin
Boyd P. Holmes
Into the Future: The Foundation of Library and Information Services in
the Post-Industrial Era, by Michael Harris, Stan A. Hannah, and Pamela C.
Harris
Ebrahim Afshar
Newspapers of Record in a Digital Age: From Hot Type to Hot Link, by
Shannon E. Martin and Kathleen A. Hansen
Amy E. Sanidas
