Applications for Bibliometric Research
in the Emerging Digital Libraries
Sally Jo Cunningham
Department of Computer Science
University of Waikato
Hamilton, New Zealand
email: [email protected]
Abstract: Large numbers of research documents have recently become available on
the Internet through “digital libraries”, and these collections are seeing high levels of
use by their related research communities. A secondary use for these document
repositories and indexes is as a platform for bibliometric research. We examine the
extent to which the new digital libraries support conventional bibliometric analysis, and
discuss shortcomings in their current forms. Interestingly, these electronic text
archives also provide opportunities for new types of studies: generally the full text of
documents are available for analysis, giving a finer grain of insight than abstract-only
online databases; these repositories often contain technical reports or pre-prints, the
“grey literature” that has been previously unavailable for analysis; and document
“usage” can be measured directly by recording user accesses, rather than studied
indirectly through document references.
1. Introduction
In recent years a number of "digital libraries" have become available through the
Internet. While the technology promises in the future to support large, heterogenous
collections, at present the most widely used of the academically-focussed digital
libraries are generally repositories of one or two types of document (typically technical
reports, journal articles, pre-prints, or conference proceedings), grouped by discipline.
A distinguishing characteristic of these digital libraries is that the full text of documents
are often available for retrieval, as well as bibliographic records.The sciences are
represented much more heavily in the present crop of digital libraries than the social
sciences, arts, or humanities. They are maintained by professional societies,
universities, research laboratories, and even private individuals. Access is generally
free, both to search and to download documents.
The emergence of these subject-specific digital libraries is particularly important
given the pattern of access to materials presently employed by research scientists.
Informal exchanges of preprints, reprints, and photocopies of papers passed on by
colleagues currently are major venues for the transmission of scientific information
between researchers in the sciences. In one study, the dependence on these sources
ranges from 12% (for chemistry) to 39% (for mathematics) of all papers cited in
researchers' own publications [11]. A qualitative study of study of how computer
scientists locate and retrieve documents (computing is one of the domains considered
later in this paper) indicates that for that field, technical reports and research documents
found in various locations on the Internet are a preferred source of information [6].
Many of the digital library systems discussed in this paper are repositories for just this
type of literature. The documents tend to be of high quality: primarily technical
reports or working papers from research institutions (both academic and commercial),
as well as advance copies of work accepted for publication in conventional paper
journals. Moreover, these digital libraries are also coming to include refereed work
published digitally (in electronic journals). Anecdotal evidence suggests that in their
fields, these digital libraries are coming to be the resource of choice for locating cutting
edge work.
For specialized subjects such as high energy physics, this dependence on
informal or extra-library dissemination can be much higher. Ginsparg ([9], [10])
reports that fields in physics have traditionally relied heavily on preprint exchanges, and
the digital repositories of physics preprints begun in 1991 (the PHYSICS E-PRINT
ARCHIVES) have to a large extent supplanted conventional publishing and physical
paper mailing of technical reports. By providing ready access to information sources
that are already preferentially utilized by scientists, the digital libraries show potential to
increase access to information that until recently was expensive or difficult to acquire in
paper form. Indeed, in some fields (most notably physics) this process has already
begun, as researchers in less developed countries report access to ongoing research
through the Internet repositories that their local libraries could not afford to acquire
through conventional journal subscriptions ([9], [10]).
The primary use for new bibliographic resources is, of course, for the contents
of the documents involved. A secondary use for emerging resources is as a basis for
bibliometric analysis of the subject field. With the conventionally published scientific
literature, the sheer difficulty of accumulating statistics discouraged bibliometric
research until the advent of large bibliographic databases in the 1960's. Computerized
bibliographic databases sparked a significant increase in the number of large-scale
bibliographic studies, as significant portions of the collection and analysis of data could
be automated ([12], [13]). The availability of CD-ROM versions of bibliographic
databases has been of particular importance, since they provide a cheaper alternative to
the online commercial databases [3].
These computerized bibliographic resources have drawbacks, however. The
greatest is that the full text of documents are rarely available, and even abstracts are not
always present. This obviously limits the types of bibliometric research that can be
conducted solely through these databases. In addition, these databases are generally
limited to formally published documents (those appearing in selected books, journals,
and conference proceedings). The "grey literature" of technical reports, pre-prints, and
other works not formally published are largely ignored, and it is this absence of easy
access to these documents that has hampered the analysis of these important forms of
scientific communication.
The digital libraries currently in existence complement the online and CD-ROM
bibliographic databases. They are best suited for examinations of the "physical"
characteristics of documents (for example, document length), analysis based on
bibliographic information that can be automatically extracted from the document text or
the sometimes unevenly formatted bibliographic records (such as obsolescence
studies), and usage studies (geographic or institutional origin of users, date/time of
access, individual patterns of document retrieval, etc.). Because references are present
in the document file but not identified by field, co-citation and bibliographic coupling
research is not well-supported, and conducting these studies requires considerable
effort on the part of the researcher.
The variety of bibliographic repositories in the available digital libraries in itself
has great potential in conducting bibliometric research. Sigogneau et al [15] present a
case study illustrating the ways in which the strengths of different databases can be
played off each other; they conduct a fine-grained analysis of the emergence of research
fronts in molecular and cellular biology, and demonstrate that the observations gleaned
from two complementary bibliographic databases provide greater insight into their
problem. Similarly, it appears that the types of bibliographic data that can be gleaned
from the relatively unstructured digital libraries can be profitably combined with data
from online databases, CD-ROMS, and other more conventional bibliographic
resources.
This paper is organized as follows: Section 2 discusses the types of indexing
and searching available with current digital libraries; Section 3 gives examples of
conventional bibliometric techniques applied to Internet-accessible archives; Section 4
discusses opportunities to directly measure usage of documents and to detect
information-seeking patterns in researchers; and Section 5 presents our conclusions.
2. Indexing and searching in current digital libraries
At present, the types of indexing fields for most academically-oriented digital
library systems are limited. Many schemes index on user-supplied document
descriptions, abstracts, or similar document surrogates (for example, the PHYSICS E-
PRINT ARCHIVE [10], a collection of physics pre-prints and technical reports). As will
be discussed below, the quality of this user-provided data can be highly variable, and
may unfavorably impact the usefulness of the index for searching. Alternatively, a
designated site librarian may maintain a catalog (eg, the WATERS [14] system, now
subsumed by NCSTRL (http://www.ncstrl.org/), both primarily collections of
computer science technical reports); in this case the quality of the bibliographic
information may be expedited to be higher, but fewer sites will be likely to support
such a librarian and therefore fewer documents are likely to be included in the digital
library. In a “harvesting” system such as the computer science technical report
collections supported by HARVEST [2] or the NEW ZEALAND DIGITAL LIBRARY
computer science technical report collection ([16], [17]), documents are indexed from
passive repositories (that may not even be aware that their documents are being
included in the digital library). Harvesting systems therefore cannot rely on the
presence of bibliographic data of any sort.
Because of the relative paucity of high-quality bibliographic data available to
many of the current academically- or research-focussed digital library collections, their
search interfaces tend to be more primitive than those ordinarily found in online
bibliographic databases or library catalogs. Systems such as NCSTRL can support
author, title, and subject searching, but this more sophisticated search functionality
comes at the expense of requiring participating repositories to use specific software. As
a consequence, these latter systems may provide access to a small number of sites than
harvesting systems. Harvesters may access a broader range of providers, but at the
penalty of being limited to unfielded, keyword searches over the raw text of the
document or document surrogate.
Specifically, the indexing in existing digital libraries has a variety of shortcomings for
bibliometric applications:
•
lack of fielded indexing: As noted above, some large and widely used digital
libraries (such as the computer science technical report collection of the NEW
ZEALAND DIGITAL LIBRARY) may lack formal cataloging entirely, and rely on
keyword searching over the raw document text. Obviously this makes field-
dependent analysis more difficult (for example, locating documents produced by
specific authors), and in the worst case my require a manual examination of all
files in the collection in order to reliably identify a desired document subset.
However, keyword search techniques that approximate fielded searching results
may suffice: for example in the NEW ZEALAND DIGITAL LIBRARY computer
science technical report collection, limiting the keyword search for “Johnson”
to a search of first pages only is likely to retrieve documents written by Johnson
(since for the majority of computer science technical reports, the first page
contains little more than author, title, date, and institution details).
A more principled approach to extracting bibliographic information is embodied
in the CiteSeer tool [1]. This software parses raw, unfielded academic
documents and attempts to identify such indexing information as author, title,
reference list, etc. Obviously such a tool cannot attain 100% accuracy over a
heterogenous document collection, but in practice it appears useful in that it can
make a good first pass in processing a set of documents, providing an initial set
of parsed documents for analysis. The remaining (presumably much smaller) set
of unparsable documents can then be dealt with manually.
•
lack of consistency in field formatting: Current digital libraries usually acquire
bibliographic information from either the authors of submitted articles or
automatic extraction routines (retrieving bibliographic details from catalog files
that may or may not be in a given document site, and that may or may not be in
an easily parsable form). Neither of these methods produce records with
standard formatting, which causes problems with automated bibliometric
analysis. Consider the following examples selected from entries in the hep-th
(high energy physics) collection of the PHYSICS E-PRINT ARCHIVES:
(i)
Authors: A. Yu. Alekseev, V. Schomerus
(ii)
Authors: Adel Bilal and Ian. I. Kogan
(iii)
Authors: Paul S. Aspinwall and David R. Morrison (with an appendix
by Mark Gross)
(iv)
Authors: A. H. Chamseddine and Herbi Dreiner (ETH-Zurich)
In this case, typical for existing digital libraries, there is no standardized format
for authors' names (here, appearing with full names, initials plus last name, and
a mixture of the two); no standard convention for separating author names
(here, either a comma or "and" are used); and parenthetical information can
include a variety of information such as the name of an associate author or the
institutional affiliations of an author. Manual processing or specially crafted
software would be required to reformat these fields for analysis.
•
duplicate entries: Digital libraries that draw documents from a variety of sources
may inadvertently contain duplicate items. Unfortunately, the irregular
formatting of the bibliographic information makes it difficult to automatically
detect these duplicates.
•
implicit field tagging: In some repositories, items are not explicitly tagged with
certain types of information – most commonly the document's date of
publication or production. Instead, the date is implicit in the document's title
(eg, its numeration in a technical report series) or in the location of the document
in the file structure of the repository (eg, separate directories exist for each
year). A second common piece of implicit data is the authors’ institutional
affiliations. This may be contained in the document itself (typically on a cover
page), or may be implicit in the document’s location (for example, a
corporation’s technical reports are stored in its ftp repository). Again, in these
cases special processing is required to append this field information to a
document record for bibliometric analysis.
•
extraction of document text: Few of the documents stored in the research-
oriented digital libraries discussed in this paper are straight ascii text; instead,
documents may appear in a variety of file formats, such as LaTeX, PostScript,
PDF, etc. If the contents of the documents are to be automatically processed
(for example, to count the words in a document, or to extract reference
publication dates for an obsolescence study), then the text must be extracted.
Utilities are available to convert most common document formats to ascii.
It is likely that many of these problems will be addressed as the Internet-based
document indexing systems mature. Even minor changes can greatly increase the
useability of a bibliographic database for bibliometric research. For example, the
addition of an explicit date tag to many online databases in 1975 sparked new
applications in time series research [3].
3. Opportunities for applications of bibliometric techniques
One type of bibliometric research concentrates on quantifying fundamental,
structural details about a subject literature: how many items are published, how many
authors are publishing, over what time period documents are likely to be used, etc.
More complex studies analyze the relationships between documents, such as how
documents cluster into subjects. The following examples give a flavour of the
bibliometric research that is possible using the emerging digital libraries:
examining the “physical” characteristics of archived documents
One relatively straightforward type of bibliometric study characterizes the
formats of different literatures. For example, Figure 1 presents a the range of the size
of computer science technical reports as measured by their length in pages. Of the
45,720 documents in the CSTR collection as of April 1998, nearly 1600 did not contain
page divisions in their files (and hence are excluded from analysis). Note that the
number of pages in the shorter documents (<50 pages) falls into an approximately
normal distribution (slightly skewed to the left), while presumably the longer
documents represent Masters’ and Doctoral theses. A surprising number of documents
are very short (between one and 5 pages); these may represent the type of condensed
results frequently found in the “technical notes”, “short papers”, and “poster sessions”
of computing conferences and journals. The average number of pages per document,
27.5, appears to be slightly longer than the common upper bound for a computing
journal article, although this observation must be confirmed by a similar study of the
lengths of formally published computing articles.
This type of analysis is of particular interest for technical reports, since they
have not been studied in the same detail as formally published papers. A comparison of
the physical characteristics of the formal and informal literature could provide
supporting evidence for common beliefs about the relationship between the two types
of documents. For example, do publishing constraints force journal and proceedings
articles to be shorter than technical reports, and therefore presumably omit technical
details of findings? Do technical reports contain more/less extensive reference sections?
If reference sections of technical reports are longer than those of published articles, then
citation links are being ommitted in published works; if technical reports contain fewer
references, then this may confirm earlier indications that computer scientists tend to
“research first” and do literature surveys later [6].
Figure 1. Range of sizes of CS technical reports, measured by number of pages
obsolescence studies.
A document is considered obsolete when it is no longer referenced by the
current literature. Typically, documents receive their greatest number and frequency of
citations immediately after publication, and the frequency of citation falls rapidly as time
passes. One technique for estimating the obsolescence rate of a body of literature– the
synchronous method – is to find the median date in the references of the documents.
This median date is subtracted from the year of publication for the documents, yielding
the median citation age. As would be expected, this median varies between the
disciplines. Typically the social sciences and arts have a higher median citation age
than the “hard” sciences and engineering, indicating that documents obsolesce more
quickly for the latter fields.
As noted in Section 2, references are not generally explicitly tagged in existing
digital repositories. However, reference dates can usually be extracted from the
document text by first locating the reference section (usually delimited by a "references"
or "bibliography" section heading), and then extracting all numbers in the appropriate
ranges for dates for the field under study.
To illustrate this process, 188 technical reports were sampled from Internet-
accessible repositories1 and used as source documents for a synchronous obsolescence
study. Conveniently, the repositories chosen organize technical reports into sub-
directories by their date of publication. The reference dates for each technical report
were automatically extracted by software that scanned the document’s file for numbers
of the form 19XX, since previous studies indicate that few if any computing reports
reference documents published in previous centuries [5]. Table 1 presents the median
citation age calculated for these documents, broken down by repository and the year of
publication for the source documents from which the reference dates were extracted:
Table 1. Median citation ages for technical report repositories
The median citation age ranges between 2 and 4 years, which is consistent with
previous examinations of computing and information systems literature ([5], [4]).
When graphed, the distribution of reference dates show the exponential curve typically
found in obsolescence studies, including the final droop due to an “immediacy effect”
as fewer very new documents are available for citation [7]. These types of results
provide confirmation that references used in computer science technical reports (the pre-
eminent “grey literature” of the computing field) conforms to the same patterns as
references found in the formally published literature.
co-citation and bibliographic coupling studies
The rate at which documents cite each other (co-citation) or cite the same
documents (bibliographic coupling) can be used to produce "maps" of a subject
literature. These techniques rely on analysis of the references of documents, and these
references must be in a common format. While digital libraries contain full text of
documents, their references are not standardized, and indeed are not even tagged as
such. To perform these studies the references must be manually extracted and
processed–a tedious process that is only worthwhile for documents (such as technical
reports) that are not included in existing citation databases such as the Science Citation
Index and Social Science Citation Index.
detecting cycles or regularities in the rate of production of research
Analysis of trends in the production of technical reports can give indications
about working conditions that affect research; for example, is more research produced
over the summer, when the teaching load is lighter? or is research steadily produced
throughout the year?
Figure 2. Distribution of the number of documents submitted to hep-th, 1992-1994
Figures 2 and 3 present statistics on document accumulation in the hep-th (high
energy physics) e-print server, a part of the PHYSICS E-PRINT ARCHIVE. This system
is one of the oldest formal pre-print archives, and has become the primary means for
information dissemination in its field. Examination of these figures reveals several
trends. Clearly the absolute number of documents deposited in the repository has
tended to increase over the time period. For all three years, research production has its
lowest point in January and February, increases through May and June, then decreases
until August and September. At that point the rate of production steps up, reaching a
yearly peak in November and December. This pattern is less clear for 1992, which
might be expected as the archive was established in mid-1991.
Figure 3. Distribution of the percentage of documents submitted to hep-th, 1992-1994
4. Analysis of usage data
The emerging Internet-based digital libraries will permit research on scientific
information collection and use at a much finer grain than is possible with current paper
libraries or online bibliographic databases. Current bibliometric or scientometric
research of this type must measure information use indirectly – for example, through
examination of the list of references appended to published articles. However, it is well
known that authors do not necessarily include in the reference list all documents that
could have been cited, and conversely that not all references listed may have been
actually “used” in performing the research; citation behavior can be affected by a
number of motivating factors (Garfield lists 15 possible reasons in [8]).
Digital library transaction logs provide a powerful tool for direct analysis of
document “usage”: since digital libraries contain the actual document (rather than only a
document surrogate), the relative amount of “use” that a digital library’s clients make of
a given document sees can be estimated from the number of times the document file is
downloaded (and, presumably, the document is read). Note that file downloading is a
much stronger statement on the part of the user than, for example, having a
bibliographic record appear in the query result set for a conventional bibliographic
system; the user downloads only after the document has been found potentially relevant
through examination of its document surrogate. Additionally, downloading is
frequently time-consuming and sometimes costly (depending on local pricing for
Internet access). Downloaded documents are therefore highly likely at least to be
scanned, if not read closely. The transaction logs for a digital library can provide a
global picture of the use of documents in the collection, since all user interactions with
the library can be automatically logged for analysis. By contrast, it is of course
impossible to track usage of print bibliographies, and very difficult to monitor usage of
bibliographic data available on CD-ROM across more than one or two sites.
Furthermore, analysis of search requests by geographic location, institution,
and sometimes even individual user are also possible. As an example, Table 2 presents
a portion of the summary of usage statistics (broken down by domain code) for queries
to the computer science technical collection of the NEW ZEALAND DIGITAL LIBRARY.
Examination of the data indicates that the heaviest use of the collection comes from
North America, Europe (particularly Germany and Finland), as well as the local New
Zealand community and nearby Australia. As expected for such a collection, a large
proportion of users are from educational (.edu) institutions; surprisingly, however, a
similar number of queries come from commercial (.com) organizations, indicating
perhaps that the documents are seeing use in commercial research and development
units.
Table 2. Accesses to the NEW ZEALAND DIGITAL LIBRARY CS collection by Domain
Code
Of course, usage levels can also be further broken down by IP number
(indicating institutions), and systems requiring users to register may also be able to
analyze usage on an individual basis. Since the query strings themselves are also
recorded in the transaction logs, this domain/institution/individual activity could also be
linked to specific subjects through the query terms. Summaries of this type could be
invaluable for studies of geographic diffusion and distribution of research topics.
Transaction log analysis can also indicate time-related patterns in the
information seeking behavior of digital library users. As a sample of this type of
analysis, Paul Ginsparg notes a seven day periodicity in the number of search requests
made to the PHYSICS E-PRINT archives (Figure 4, reproduced from [9]). From this he
adduces that many physicists do not yet have weekend access to the Internet (an
alternative, slightly more cynical hypothesis is that even high energy theoretical
physicists take the weekend off).
Figure 4. Summary of search requests to the physics pre-print archives
5. Conclusion
This study suggests opportunities for conducting bibliometric research on the
evolving digital libraries. These repositories are suitable platforms for conventional
bibliometric techniques (such as obsolescence studies, quantification of physical
characteristics of documents comprising a subject literature, time analysis, etc.). The
ability to directly monitor access to documents in digital libraries also enables
researchers to explicitly quantify document usage, as well as to implicitly measure
usage through citations. Additional facilities could aid in the performance of
bibliographic experiments, such as: improved tagging of document fields; provision of
utilities to strip out titles, authors, etc. from common document formats; and the ability
to easily eliminate duplicate entries from downloaded library subsets. Unfortunately,
the most useful of these additional facilities – those associated with a higher degree of
cataloging – run counter to the underlying philosophy of many digital libraries: to
avoid, if possible, manual processing and formal cataloging of documents. While
adherence to this principle can limit the accuracy of fielded searching (or indeed,
preclude it altogether), it can also avoid the cataloging bottleneck and permit digital
libraries to provide access to larger numbers of documents.
The digital libraries complement the information currently available through
paper, online, and CD-ROM bibliographic resources. While these latter databases
generally have the advantage of standardized formatting of bibliographic fields, the
digital libraries are freely accessible, often contain "grey literature" that is otherwise
unavailable for analysis, and generally make the full text of documents available. The
insights gained from analysis of digital libraries will add to the store of "information
about information" that we have gained from older types of bibliographic repositories.
References
[1] Bollacker, K.D., S. Lawrence, and C.L.Giles, CiteSeer: An Autonomous Web
Agent for Automatic Retrieval and Identification of Interesting Publications,
Proceedings of the Second International Conference on Autonomous Agents
(Minneapolis/St. Paul, May 9-13), 1998.
[2] Bowman, C.M., P.B. Danzig, U. Manber, and M.F. Schwartz, Scalable Internet
resource discovery: Research problems and approaches, Communications of
the ACM 37(8) (1994) 98-107.
[3] Burton, Hilary D. , Use of a virtual information system for bibliometric analysis,
Informaton Processing & Management 24(1) (1988) 39-44.
[4] Cunningham, S.J., An empirical investigation of the obsolescence rate for
information systems literature, Library and Information Science
Research., 1996, http://library.fgcu.edu/iclc/lisrissu.htm
[5] Cunningham, S.J., and D. Bocock, Obsolescence of computing literature.
Scientometrics 34(2) (1995), pp. 255-262.
[6] Cunningham, S.J. and Lynn Silipigni Connaway, Information searching
preferences and practices of computer science researchers, Proceedings of
OZCHI '96 (1996) 294-299.
[7] de Solla Price, D.J., Citation measures of hard science, soft science, technology,
and nonscience. In: C.E. Nelson and D.K. Pollock (eds), Communication
among scientists and engineers (Heath Lexington, 1970).
[8] Garfield, E., Citation Indexing: Its theory and application in Science, Technology
and Humanities (Wiley, 1979).
[9] Ginsparg, P. After dinner remarks: 14 Oct ‘94 APS meeting at LANL, 1994
(<URL: http://xxx.lanl.gov/blurb> ).
[10] Ginsparg, P., First steps towards electronic research communication, Computers
in Physics 8(4) (1994) 390-401.
[11] Hallmark, J., Scientists' access and retrieval of references cited in their recent
journal articles, College and Research Libraries 55(3) (1994) 199-210.
[12] Hawkins, D.T. , Unconventional uses of on-line information retrieval systems:
on-line bibliometric studies, Journal of the American Society for Information
Science 28 (1977) 13-18.
[13] McGhee, P.E. , P.R. Skinner, K. Roberto, N.J. Ridenour, and S.M. Larson,
Using online databases to study current research trends: an online bibliometric
study, Library and Information Science Research 9 (1987) 285-291.
[14] Maly, K., E.A. Fox, J.C. French, and A.L. Selman, Wide area technical report
server (Technical Report , Dept. of Computer Science, Old Dominion
University,
1994.
Also
available
at
<URL:
http://www.cs.odu.edu/WATERS/WATERS-paper.ps> ).
[15] Sigogneau, M.J. , S. Bain, J.P. Courtial, and H. Feillet, Scientific innovation in
bibliographical databases: a comparative study of the Science Citation Index
and the Pascal database, Scientometrics 22(1) (1991) 65-82.
[16] Witten, I.H., S.J. Cunningham, M. Vallabh, and T.C. Bell, A New Zealand
digital library for computer science research, Proceedings of Digital Libraries
'95 (1995) 25-30.
[17] Witten, I.H., C. Nevill-Manning, and S.J. Cunningham, A public library based
on full-text retrieval, Communications of the ACM 41(4), 1998, p. 71
1Documents were randomly sampled from the DEC
(ftp://crl.dec.com/pub/DEC/CRL/tech-reports/), Sony
(ftp://ftp.csl.sony.co.jp/CSL/CSL-Papers), and Ohio (ftp://archive.cis.ohio-
state.edu/pub/tech-report/) technical report repositories
in the Emerging Digital Libraries
Sally Jo Cunningham
Department of Computer Science
University of Waikato
Hamilton, New Zealand
email: [email protected]
Abstract: Large numbers of research documents have recently become available on
the Internet through “digital libraries”, and these collections are seeing high levels of
use by their related research communities. A secondary use for these document
repositories and indexes is as a platform for bibliometric research. We examine the
extent to which the new digital libraries support conventional bibliometric analysis, and
discuss shortcomings in their current forms. Interestingly, these electronic text
archives also provide opportunities for new types of studies: generally the full text of
documents are available for analysis, giving a finer grain of insight than abstract-only
online databases; these repositories often contain technical reports or pre-prints, the
“grey literature” that has been previously unavailable for analysis; and document
“usage” can be measured directly by recording user accesses, rather than studied
indirectly through document references.
1. Introduction
In recent years a number of "digital libraries" have become available through the
Internet. While the technology promises in the future to support large, heterogenous
collections, at present the most widely used of the academically-focussed digital
libraries are generally repositories of one or two types of document (typically technical
reports, journal articles, pre-prints, or conference proceedings), grouped by discipline.
A distinguishing characteristic of these digital libraries is that the full text of documents
are often available for retrieval, as well as bibliographic records.The sciences are
represented much more heavily in the present crop of digital libraries than the social
sciences, arts, or humanities. They are maintained by professional societies,
universities, research laboratories, and even private individuals. Access is generally
free, both to search and to download documents.
The emergence of these subject-specific digital libraries is particularly important
given the pattern of access to materials presently employed by research scientists.
Informal exchanges of preprints, reprints, and photocopies of papers passed on by
colleagues currently are major venues for the transmission of scientific information
between researchers in the sciences. In one study, the dependence on these sources
ranges from 12% (for chemistry) to 39% (for mathematics) of all papers cited in
researchers' own publications [11]. A qualitative study of study of how computer
scientists locate and retrieve documents (computing is one of the domains considered
later in this paper) indicates that for that field, technical reports and research documents
found in various locations on the Internet are a preferred source of information [6].
Many of the digital library systems discussed in this paper are repositories for just this
type of literature. The documents tend to be of high quality: primarily technical
reports or working papers from research institutions (both academic and commercial),
as well as advance copies of work accepted for publication in conventional paper
journals. Moreover, these digital libraries are also coming to include refereed work
published digitally (in electronic journals). Anecdotal evidence suggests that in their
fields, these digital libraries are coming to be the resource of choice for locating cutting
edge work.
For specialized subjects such as high energy physics, this dependence on
informal or extra-library dissemination can be much higher. Ginsparg ([9], [10])
reports that fields in physics have traditionally relied heavily on preprint exchanges, and
the digital repositories of physics preprints begun in 1991 (the PHYSICS E-PRINT
ARCHIVES) have to a large extent supplanted conventional publishing and physical
paper mailing of technical reports. By providing ready access to information sources
that are already preferentially utilized by scientists, the digital libraries show potential to
increase access to information that until recently was expensive or difficult to acquire in
paper form. Indeed, in some fields (most notably physics) this process has already
begun, as researchers in less developed countries report access to ongoing research
through the Internet repositories that their local libraries could not afford to acquire
through conventional journal subscriptions ([9], [10]).
The primary use for new bibliographic resources is, of course, for the contents
of the documents involved. A secondary use for emerging resources is as a basis for
bibliometric analysis of the subject field. With the conventionally published scientific
literature, the sheer difficulty of accumulating statistics discouraged bibliometric
research until the advent of large bibliographic databases in the 1960's. Computerized
bibliographic databases sparked a significant increase in the number of large-scale
bibliographic studies, as significant portions of the collection and analysis of data could
be automated ([12], [13]). The availability of CD-ROM versions of bibliographic
databases has been of particular importance, since they provide a cheaper alternative to
the online commercial databases [3].
These computerized bibliographic resources have drawbacks, however. The
greatest is that the full text of documents are rarely available, and even abstracts are not
always present. This obviously limits the types of bibliometric research that can be
conducted solely through these databases. In addition, these databases are generally
limited to formally published documents (those appearing in selected books, journals,
and conference proceedings). The "grey literature" of technical reports, pre-prints, and
other works not formally published are largely ignored, and it is this absence of easy
access to these documents that has hampered the analysis of these important forms of
scientific communication.
The digital libraries currently in existence complement the online and CD-ROM
bibliographic databases. They are best suited for examinations of the "physical"
characteristics of documents (for example, document length), analysis based on
bibliographic information that can be automatically extracted from the document text or
the sometimes unevenly formatted bibliographic records (such as obsolescence
studies), and usage studies (geographic or institutional origin of users, date/time of
access, individual patterns of document retrieval, etc.). Because references are present
in the document file but not identified by field, co-citation and bibliographic coupling
research is not well-supported, and conducting these studies requires considerable
effort on the part of the researcher.
The variety of bibliographic repositories in the available digital libraries in itself
has great potential in conducting bibliometric research. Sigogneau et al [15] present a
case study illustrating the ways in which the strengths of different databases can be
played off each other; they conduct a fine-grained analysis of the emergence of research
fronts in molecular and cellular biology, and demonstrate that the observations gleaned
from two complementary bibliographic databases provide greater insight into their
problem. Similarly, it appears that the types of bibliographic data that can be gleaned
from the relatively unstructured digital libraries can be profitably combined with data
from online databases, CD-ROMS, and other more conventional bibliographic
resources.
This paper is organized as follows: Section 2 discusses the types of indexing
and searching available with current digital libraries; Section 3 gives examples of
conventional bibliometric techniques applied to Internet-accessible archives; Section 4
discusses opportunities to directly measure usage of documents and to detect
information-seeking patterns in researchers; and Section 5 presents our conclusions.
2. Indexing and searching in current digital libraries
At present, the types of indexing fields for most academically-oriented digital
library systems are limited. Many schemes index on user-supplied document
descriptions, abstracts, or similar document surrogates (for example, the PHYSICS E-
PRINT ARCHIVE [10], a collection of physics pre-prints and technical reports). As will
be discussed below, the quality of this user-provided data can be highly variable, and
may unfavorably impact the usefulness of the index for searching. Alternatively, a
designated site librarian may maintain a catalog (eg, the WATERS [14] system, now
subsumed by NCSTRL (http://www.ncstrl.org/), both primarily collections of
computer science technical reports); in this case the quality of the bibliographic
information may be expedited to be higher, but fewer sites will be likely to support
such a librarian and therefore fewer documents are likely to be included in the digital
library. In a “harvesting” system such as the computer science technical report
collections supported by HARVEST [2] or the NEW ZEALAND DIGITAL LIBRARY
computer science technical report collection ([16], [17]), documents are indexed from
passive repositories (that may not even be aware that their documents are being
included in the digital library). Harvesting systems therefore cannot rely on the
presence of bibliographic data of any sort.
Because of the relative paucity of high-quality bibliographic data available to
many of the current academically- or research-focussed digital library collections, their
search interfaces tend to be more primitive than those ordinarily found in online
bibliographic databases or library catalogs. Systems such as NCSTRL can support
author, title, and subject searching, but this more sophisticated search functionality
comes at the expense of requiring participating repositories to use specific software. As
a consequence, these latter systems may provide access to a small number of sites than
harvesting systems. Harvesters may access a broader range of providers, but at the
penalty of being limited to unfielded, keyword searches over the raw text of the
document or document surrogate.
Specifically, the indexing in existing digital libraries has a variety of shortcomings for
bibliometric applications:
•
lack of fielded indexing: As noted above, some large and widely used digital
libraries (such as the computer science technical report collection of the NEW
ZEALAND DIGITAL LIBRARY) may lack formal cataloging entirely, and rely on
keyword searching over the raw document text. Obviously this makes field-
dependent analysis more difficult (for example, locating documents produced by
specific authors), and in the worst case my require a manual examination of all
files in the collection in order to reliably identify a desired document subset.
However, keyword search techniques that approximate fielded searching results
may suffice: for example in the NEW ZEALAND DIGITAL LIBRARY computer
science technical report collection, limiting the keyword search for “Johnson”
to a search of first pages only is likely to retrieve documents written by Johnson
(since for the majority of computer science technical reports, the first page
contains little more than author, title, date, and institution details).
A more principled approach to extracting bibliographic information is embodied
in the CiteSeer tool [1]. This software parses raw, unfielded academic
documents and attempts to identify such indexing information as author, title,
reference list, etc. Obviously such a tool cannot attain 100% accuracy over a
heterogenous document collection, but in practice it appears useful in that it can
make a good first pass in processing a set of documents, providing an initial set
of parsed documents for analysis. The remaining (presumably much smaller) set
of unparsable documents can then be dealt with manually.
•
lack of consistency in field formatting: Current digital libraries usually acquire
bibliographic information from either the authors of submitted articles or
automatic extraction routines (retrieving bibliographic details from catalog files
that may or may not be in a given document site, and that may or may not be in
an easily parsable form). Neither of these methods produce records with
standard formatting, which causes problems with automated bibliometric
analysis. Consider the following examples selected from entries in the hep-th
(high energy physics) collection of the PHYSICS E-PRINT ARCHIVES:
(i)
Authors: A. Yu. Alekseev, V. Schomerus
(ii)
Authors: Adel Bilal and Ian. I. Kogan
(iii)
Authors: Paul S. Aspinwall and David R. Morrison (with an appendix
by Mark Gross)
(iv)
Authors: A. H. Chamseddine and Herbi Dreiner (ETH-Zurich)
In this case, typical for existing digital libraries, there is no standardized format
for authors' names (here, appearing with full names, initials plus last name, and
a mixture of the two); no standard convention for separating author names
(here, either a comma or "and" are used); and parenthetical information can
include a variety of information such as the name of an associate author or the
institutional affiliations of an author. Manual processing or specially crafted
software would be required to reformat these fields for analysis.
•
duplicate entries: Digital libraries that draw documents from a variety of sources
may inadvertently contain duplicate items. Unfortunately, the irregular
formatting of the bibliographic information makes it difficult to automatically
detect these duplicates.
•
implicit field tagging: In some repositories, items are not explicitly tagged with
certain types of information – most commonly the document's date of
publication or production. Instead, the date is implicit in the document's title
(eg, its numeration in a technical report series) or in the location of the document
in the file structure of the repository (eg, separate directories exist for each
year). A second common piece of implicit data is the authors’ institutional
affiliations. This may be contained in the document itself (typically on a cover
page), or may be implicit in the document’s location (for example, a
corporation’s technical reports are stored in its ftp repository). Again, in these
cases special processing is required to append this field information to a
document record for bibliometric analysis.
•
extraction of document text: Few of the documents stored in the research-
oriented digital libraries discussed in this paper are straight ascii text; instead,
documents may appear in a variety of file formats, such as LaTeX, PostScript,
PDF, etc. If the contents of the documents are to be automatically processed
(for example, to count the words in a document, or to extract reference
publication dates for an obsolescence study), then the text must be extracted.
Utilities are available to convert most common document formats to ascii.
It is likely that many of these problems will be addressed as the Internet-based
document indexing systems mature. Even minor changes can greatly increase the
useability of a bibliographic database for bibliometric research. For example, the
addition of an explicit date tag to many online databases in 1975 sparked new
applications in time series research [3].
3. Opportunities for applications of bibliometric techniques
One type of bibliometric research concentrates on quantifying fundamental,
structural details about a subject literature: how many items are published, how many
authors are publishing, over what time period documents are likely to be used, etc.
More complex studies analyze the relationships between documents, such as how
documents cluster into subjects. The following examples give a flavour of the
bibliometric research that is possible using the emerging digital libraries:
examining the “physical” characteristics of archived documents
One relatively straightforward type of bibliometric study characterizes the
formats of different literatures. For example, Figure 1 presents a the range of the size
of computer science technical reports as measured by their length in pages. Of the
45,720 documents in the CSTR collection as of April 1998, nearly 1600 did not contain
page divisions in their files (and hence are excluded from analysis). Note that the
number of pages in the shorter documents (<50 pages) falls into an approximately
normal distribution (slightly skewed to the left), while presumably the longer
documents represent Masters’ and Doctoral theses. A surprising number of documents
are very short (between one and 5 pages); these may represent the type of condensed
results frequently found in the “technical notes”, “short papers”, and “poster sessions”
of computing conferences and journals. The average number of pages per document,
27.5, appears to be slightly longer than the common upper bound for a computing
journal article, although this observation must be confirmed by a similar study of the
lengths of formally published computing articles.
This type of analysis is of particular interest for technical reports, since they
have not been studied in the same detail as formally published papers. A comparison of
the physical characteristics of the formal and informal literature could provide
supporting evidence for common beliefs about the relationship between the two types
of documents. For example, do publishing constraints force journal and proceedings
articles to be shorter than technical reports, and therefore presumably omit technical
details of findings? Do technical reports contain more/less extensive reference sections?
If reference sections of technical reports are longer than those of published articles, then
citation links are being ommitted in published works; if technical reports contain fewer
references, then this may confirm earlier indications that computer scientists tend to
“research first” and do literature surveys later [6].
Figure 1. Range of sizes of CS technical reports, measured by number of pages
obsolescence studies.
A document is considered obsolete when it is no longer referenced by the
current literature. Typically, documents receive their greatest number and frequency of
citations immediately after publication, and the frequency of citation falls rapidly as time
passes. One technique for estimating the obsolescence rate of a body of literature– the
synchronous method – is to find the median date in the references of the documents.
This median date is subtracted from the year of publication for the documents, yielding
the median citation age. As would be expected, this median varies between the
disciplines. Typically the social sciences and arts have a higher median citation age
than the “hard” sciences and engineering, indicating that documents obsolesce more
quickly for the latter fields.
As noted in Section 2, references are not generally explicitly tagged in existing
digital repositories. However, reference dates can usually be extracted from the
document text by first locating the reference section (usually delimited by a "references"
or "bibliography" section heading), and then extracting all numbers in the appropriate
ranges for dates for the field under study.
To illustrate this process, 188 technical reports were sampled from Internet-
accessible repositories1 and used as source documents for a synchronous obsolescence
study. Conveniently, the repositories chosen organize technical reports into sub-
directories by their date of publication. The reference dates for each technical report
were automatically extracted by software that scanned the document’s file for numbers
of the form 19XX, since previous studies indicate that few if any computing reports
reference documents published in previous centuries [5]. Table 1 presents the median
citation age calculated for these documents, broken down by repository and the year of
publication for the source documents from which the reference dates were extracted:
Table 1. Median citation ages for technical report repositories
The median citation age ranges between 2 and 4 years, which is consistent with
previous examinations of computing and information systems literature ([5], [4]).
When graphed, the distribution of reference dates show the exponential curve typically
found in obsolescence studies, including the final droop due to an “immediacy effect”
as fewer very new documents are available for citation [7]. These types of results
provide confirmation that references used in computer science technical reports (the pre-
eminent “grey literature” of the computing field) conforms to the same patterns as
references found in the formally published literature.
co-citation and bibliographic coupling studies
The rate at which documents cite each other (co-citation) or cite the same
documents (bibliographic coupling) can be used to produce "maps" of a subject
literature. These techniques rely on analysis of the references of documents, and these
references must be in a common format. While digital libraries contain full text of
documents, their references are not standardized, and indeed are not even tagged as
such. To perform these studies the references must be manually extracted and
processed–a tedious process that is only worthwhile for documents (such as technical
reports) that are not included in existing citation databases such as the Science Citation
Index and Social Science Citation Index.
detecting cycles or regularities in the rate of production of research
Analysis of trends in the production of technical reports can give indications
about working conditions that affect research; for example, is more research produced
over the summer, when the teaching load is lighter? or is research steadily produced
throughout the year?
Figure 2. Distribution of the number of documents submitted to hep-th, 1992-1994
Figures 2 and 3 present statistics on document accumulation in the hep-th (high
energy physics) e-print server, a part of the PHYSICS E-PRINT ARCHIVE. This system
is one of the oldest formal pre-print archives, and has become the primary means for
information dissemination in its field. Examination of these figures reveals several
trends. Clearly the absolute number of documents deposited in the repository has
tended to increase over the time period. For all three years, research production has its
lowest point in January and February, increases through May and June, then decreases
until August and September. At that point the rate of production steps up, reaching a
yearly peak in November and December. This pattern is less clear for 1992, which
might be expected as the archive was established in mid-1991.
Figure 3. Distribution of the percentage of documents submitted to hep-th, 1992-1994
4. Analysis of usage data
The emerging Internet-based digital libraries will permit research on scientific
information collection and use at a much finer grain than is possible with current paper
libraries or online bibliographic databases. Current bibliometric or scientometric
research of this type must measure information use indirectly – for example, through
examination of the list of references appended to published articles. However, it is well
known that authors do not necessarily include in the reference list all documents that
could have been cited, and conversely that not all references listed may have been
actually “used” in performing the research; citation behavior can be affected by a
number of motivating factors (Garfield lists 15 possible reasons in [8]).
Digital library transaction logs provide a powerful tool for direct analysis of
document “usage”: since digital libraries contain the actual document (rather than only a
document surrogate), the relative amount of “use” that a digital library’s clients make of
a given document sees can be estimated from the number of times the document file is
downloaded (and, presumably, the document is read). Note that file downloading is a
much stronger statement on the part of the user than, for example, having a
bibliographic record appear in the query result set for a conventional bibliographic
system; the user downloads only after the document has been found potentially relevant
through examination of its document surrogate. Additionally, downloading is
frequently time-consuming and sometimes costly (depending on local pricing for
Internet access). Downloaded documents are therefore highly likely at least to be
scanned, if not read closely. The transaction logs for a digital library can provide a
global picture of the use of documents in the collection, since all user interactions with
the library can be automatically logged for analysis. By contrast, it is of course
impossible to track usage of print bibliographies, and very difficult to monitor usage of
bibliographic data available on CD-ROM across more than one or two sites.
Furthermore, analysis of search requests by geographic location, institution,
and sometimes even individual user are also possible. As an example, Table 2 presents
a portion of the summary of usage statistics (broken down by domain code) for queries
to the computer science technical collection of the NEW ZEALAND DIGITAL LIBRARY.
Examination of the data indicates that the heaviest use of the collection comes from
North America, Europe (particularly Germany and Finland), as well as the local New
Zealand community and nearby Australia. As expected for such a collection, a large
proportion of users are from educational (.edu) institutions; surprisingly, however, a
similar number of queries come from commercial (.com) organizations, indicating
perhaps that the documents are seeing use in commercial research and development
units.
Table 2. Accesses to the NEW ZEALAND DIGITAL LIBRARY CS collection by Domain
Code
Of course, usage levels can also be further broken down by IP number
(indicating institutions), and systems requiring users to register may also be able to
analyze usage on an individual basis. Since the query strings themselves are also
recorded in the transaction logs, this domain/institution/individual activity could also be
linked to specific subjects through the query terms. Summaries of this type could be
invaluable for studies of geographic diffusion and distribution of research topics.
Transaction log analysis can also indicate time-related patterns in the
information seeking behavior of digital library users. As a sample of this type of
analysis, Paul Ginsparg notes a seven day periodicity in the number of search requests
made to the PHYSICS E-PRINT archives (Figure 4, reproduced from [9]). From this he
adduces that many physicists do not yet have weekend access to the Internet (an
alternative, slightly more cynical hypothesis is that even high energy theoretical
physicists take the weekend off).
Figure 4. Summary of search requests to the physics pre-print archives
5. Conclusion
This study suggests opportunities for conducting bibliometric research on the
evolving digital libraries. These repositories are suitable platforms for conventional
bibliometric techniques (such as obsolescence studies, quantification of physical
characteristics of documents comprising a subject literature, time analysis, etc.). The
ability to directly monitor access to documents in digital libraries also enables
researchers to explicitly quantify document usage, as well as to implicitly measure
usage through citations. Additional facilities could aid in the performance of
bibliographic experiments, such as: improved tagging of document fields; provision of
utilities to strip out titles, authors, etc. from common document formats; and the ability
to easily eliminate duplicate entries from downloaded library subsets. Unfortunately,
the most useful of these additional facilities – those associated with a higher degree of
cataloging – run counter to the underlying philosophy of many digital libraries: to
avoid, if possible, manual processing and formal cataloging of documents. While
adherence to this principle can limit the accuracy of fielded searching (or indeed,
preclude it altogether), it can also avoid the cataloging bottleneck and permit digital
libraries to provide access to larger numbers of documents.
The digital libraries complement the information currently available through
paper, online, and CD-ROM bibliographic resources. While these latter databases
generally have the advantage of standardized formatting of bibliographic fields, the
digital libraries are freely accessible, often contain "grey literature" that is otherwise
unavailable for analysis, and generally make the full text of documents available. The
insights gained from analysis of digital libraries will add to the store of "information
about information" that we have gained from older types of bibliographic repositories.
References
[1] Bollacker, K.D., S. Lawrence, and C.L.Giles, CiteSeer: An Autonomous Web
Agent for Automatic Retrieval and Identification of Interesting Publications,
Proceedings of the Second International Conference on Autonomous Agents
(Minneapolis/St. Paul, May 9-13), 1998.
[2] Bowman, C.M., P.B. Danzig, U. Manber, and M.F. Schwartz, Scalable Internet
resource discovery: Research problems and approaches, Communications of
the ACM 37(8) (1994) 98-107.
[3] Burton, Hilary D. , Use of a virtual information system for bibliometric analysis,
Informaton Processing & Management 24(1) (1988) 39-44.
[4] Cunningham, S.J., An empirical investigation of the obsolescence rate for
information systems literature, Library and Information Science
Research., 1996, http://library.fgcu.edu/iclc/lisrissu.htm
[5] Cunningham, S.J., and D. Bocock, Obsolescence of computing literature.
Scientometrics 34(2) (1995), pp. 255-262.
[6] Cunningham, S.J. and Lynn Silipigni Connaway, Information searching
preferences and practices of computer science researchers, Proceedings of
OZCHI '96 (1996) 294-299.
[7] de Solla Price, D.J., Citation measures of hard science, soft science, technology,
and nonscience. In: C.E. Nelson and D.K. Pollock (eds), Communication
among scientists and engineers (Heath Lexington, 1970).
[8] Garfield, E., Citation Indexing: Its theory and application in Science, Technology
and Humanities (Wiley, 1979).
[9] Ginsparg, P. After dinner remarks: 14 Oct ‘94 APS meeting at LANL, 1994
(<URL: http://xxx.lanl.gov/blurb> ).
[10] Ginsparg, P., First steps towards electronic research communication, Computers
in Physics 8(4) (1994) 390-401.
[11] Hallmark, J., Scientists' access and retrieval of references cited in their recent
journal articles, College and Research Libraries 55(3) (1994) 199-210.
[12] Hawkins, D.T. , Unconventional uses of on-line information retrieval systems:
on-line bibliometric studies, Journal of the American Society for Information
Science 28 (1977) 13-18.
[13] McGhee, P.E. , P.R. Skinner, K. Roberto, N.J. Ridenour, and S.M. Larson,
Using online databases to study current research trends: an online bibliometric
study, Library and Information Science Research 9 (1987) 285-291.
[14] Maly, K., E.A. Fox, J.C. French, and A.L. Selman, Wide area technical report
server (Technical Report , Dept. of Computer Science, Old Dominion
University,
1994.
Also
available
at
<URL:
http://www.cs.odu.edu/WATERS/WATERS-paper.ps> ).
[15] Sigogneau, M.J. , S. Bain, J.P. Courtial, and H. Feillet, Scientific innovation in
bibliographical databases: a comparative study of the Science Citation Index
and the Pascal database, Scientometrics 22(1) (1991) 65-82.
[16] Witten, I.H., S.J. Cunningham, M. Vallabh, and T.C. Bell, A New Zealand
digital library for computer science research, Proceedings of Digital Libraries
'95 (1995) 25-30.
[17] Witten, I.H., C. Nevill-Manning, and S.J. Cunningham, A public library based
on full-text retrieval, Communications of the ACM 41(4), 1998, p. 71
1Documents were randomly sampled from the DEC
(ftp://crl.dec.com/pub/DEC/CRL/tech-reports/), Sony
(ftp://ftp.csl.sony.co.jp/CSL/CSL-Papers), and Ohio (ftp://archive.cis.ohio-
state.edu/pub/tech-report/) technical report repositories