Automatic Link-Detection in Encoded Archival Descriptions

  1. 1. Junte Zhang

    University of Amsterdam

  2. 2. Khairun Nisa Fachry

    University of Amsterdam

  3. 3. Jaap Kamps

    University of Amsterdam

Work text
This plain text was ingested for the purpose of full-text search, not to preserve original formatting or readability. For the most complete copy, refer to the original conference program.

In this paper we investigate how currently emerging link
detection methods can help enrich encoded archival
descriptions. We discuss link detection methods in general,
and evaluate the identifi cation of names both within, and
across, archival descriptions. Our initial experiments suggest
that we can automatically detect occurrences of person
names with high accuracy, both within (F-score of 0.9620) and
across (F-score of 1) archival descriptions. This allows us to
create (pseudo) encoded archival context descriptions that
provide novel means of navigation, improving access to the
vast amounts of archival data.
Archival fi nding aids are complex multi-level descriptions
of the paper trails of corporations, persons and families.
Currently, fi nding aids are increasingly encoded in XML using
the standard Encoded Archival Descriptions. Archives can
cover hundreds of meters of material, resulting in long and
detailed EAD documents. We use a dataset of 2,886 EAD
documents from the International Institute of Social History
(IISG) and 3,119 documents from the Archives Hub, containing
documents with more than 100,000 words. Navigating in such
archival fi nding aids becomes non-trivial, and it is easy to loose
overview of the hierarchical structure. Hence, this may lead to
the loss of important contextual information for interpreting
the records.
Archival context may be preserved through the use of authority
records capturing information about the record creators
(corporations, persons, or families) and the context of record
creation. By separating the record creator’s descriptions
from the records or resources descriptions themselves, we
can create “links” from all occurrences of the creators to
this context. The resulting descriptions of record creators
can be encoded in XML using the emerging Encoded Archival
Context (EAC) standard. Currently, EAC has only been applied
experimentally. One of the main barriers to adoption is that
it requires substantial effort to adopt EAC. The information
for the creator’s authority record is usually available in some form (for example, EAD descriptions usually have a detailed
fi eld <bioghist> about the archive’s creator). However, linking
such a context description to occurrences of the creator in
the archival descriptions requires more structure than that is
available in legacy data.
Our main aim is to investigate if and how automatic link
detection methods could help improve archival access.
Automatic link detection studies the discovery of relations
between various documents. Such methods have been
employed to detect “missing” links on the Web and recently
in the online encyclopedia Wikipedia. Are link detection
methods suffi ciently effective to be fruitfully applied to archival
descriptions? To answer this question, we will experiment on
the detection of archival creators within and across fi nding aids.
Based on our fi ndings, we will further discuss how detected
links can be used to provide crucial contextual information
for the interpretation of records, and to improve navigation
within and across fi nding aids.
Link Detection Methods
Links generated by humans are abundant on the World Wide
Web, and knowledge repositories like the online encyclopedia
Wikipedia. There are two kinds of links: incoming and outgoing
links. Substrings of text nodes are identifi ed as anchors and
become clickable. Incoming links come from text nodes
of target fi les (destination node) and point to a source fi le
(origin node), while an outgoing link goes from text node in
the source document (origin node) to a target fi le (destination
node). Two assumptions are made: a link from document A to
document B is a recommendation of B by A, and documents
linked to each other are related.
To automatically detect whether two nodes are connected,
it is necessary to search the archives for some string that
both share. Usually it is only one specifi c and extract string. A
general approach to automatic link detection is fi rst to detect
the global similarity between documents. After the relevant
set of documents has been collected, the local similarity can
be detected by comparing text segments with other text
segments in those fi les. In structured documents like archival
fi nding aids, these text segments are often marked up as
logical units, whether it be the title <titleproper>, the wrapper
element <c12> deep down in the fi nding aid, or the element
<persname> that identifi es some personal names. These units
are identifi ed and retrieved in XML Element retrieval. The
identifi cation of relevant anchors is a key problem, as these
are used in the system’s retrieval models to point to (parts of)
related fi nding aids.
Experiment: Name Detection
A specifi c name detection trial with the archive of Joop den
Uyl (1919-1987), former Labor Party prime minister of the
Netherlands, is done as a test to deal with this problem. This
archive consists of 29,184 tokens (with removal of the XML
markup and punctuation), of which 4,979 are unique, and
where a token is a sequence of non-space characters. We
collect a list of the name variants that we expect to encounter:
“J.M. Den Uyl”, “Joop M. Den Uyl”, “Johannes Marten den Uyl”,
“Den Uyl”, etc. We construct a regular expression to fetch
the name variants. The results are depicted in illustration 1,
which shows the local view of the Joop den Uyl archive in our
Retrieving EADs More Effectively (README) system.
Illustration 1: Links detected in EAD
The quality of the name detection trial is evaluated with explicit
feedback, which means manually checking the detected links
for (1) correctness, (2) error, and (3) whether any links were
missing. This was done both within fi nding aids, and across
fi nding aids:
- First, the quality is checked within fi nding aids, by locating
occurrences of creator Joop den Uyl in his archive. For
detecting name occurrences within an archive, our simple
method has a precision of (114/120 =) 0.9500, a recall of
(114/117 =) 0.9744, resulting in an F-score of 0.9620. Some
interesting missing links used name variants where the
prefi x “den” is put behind the last name “Uyl” -- a typical
Dutch practice. Incorrect links mostly are family members
occurring the archive, e.g., “Saskia den Uyl”, “E.J. den Uyl-van
Vessem”, and also “Familie Den Uyl”. Since these names
occur relatively infrequent, few errors are made. The
matching algorithm could easily be refi ned based on these
false positives.
Table 1: Archive “Den Uyl”
Link No link
Name 114 3
No name 6 -
- Second, the same procedure to detect proper names of
Joop den Uyl is applied across fi nding aids with the related
archive of “Partij van de Arbeid Tweede-Kamer Fractie
(1955-1988)” (Dutch MPs from the Labor Party). For
detecting name occurrences across archives, we obtain a
perfect precision, recall, and thus F-score of 1. Table 2: Archive “PvdA”
Link No link
Name 16 0
No name 0 -
Concluding Discussion
In this paper we investigated how currently emerging
link detection methods can help enrich encoded archival
descriptions. We discussed link detection methods in general,
and evaluated the identifi cation of names both within, and
across, archival descriptions. Our initial experiments suggest
that we can automatically detect occurrences of person
names, both within (F-score of 0.9620) and across (F-score
of 1) archival descriptions. This allows us to create (pseudo)
encoded archival context (EAC) descriptions that provide
novel means of navigation and improve access to archival
fi nding aids. The results of our experiments were promising,
and can also be expanded to names of organizations, events,
topics, etc. We expect those to be more diffi cult than personal
name detection.
There are more uses for detecting cross-links in fi nding aids
besides creating extra contextual information. Detecting
missing links is useful for improving the retrieval of separate
fi nding aids, for example, an archival fi nding aid with many
detected incoming links may have a higher relevance. Links can
also offer a search-by-example approach, like given one fi nding
aids, fi nd all related fi nding aids. A step further is to use the
cross-links in the categorization of archival data. Concretely
for historians and other users, who rely on numerous lengthy
archival documents, new insights can be gained by detecting
missing cross-links.
This research is supported by the Netherlands Organization
for Scientifi c Research (NWO) grant # 639.072.601.
Agosti, M., Crestani, F., and Melucci, M. 1997. On the use
of information retrieval techniques for the automatic
construction of hypertext. Information Processing and
Management 33, 2 (1997), 133-144.
Allan, J. 1997. Building hypertext using information retrieval.
Information Processing and Management 33, 2 (1997), 145-159.
EAC, 2004. Encoded Archival Context. http://www.iath.
EAD, 2002. Encoded Archival Description. http://www.loc.
Fissaha Adafre, S. and De Rijke, M. 2005. Discovering missing
links in Wikipedia. In Proceedings of the 3rd international
Workshop on Link Discovery. LinkKDD ‘05. ACM Press, 90-97.
Huang, W. C., Trotman, A., and Geva, S. 2007. Collaborative
Knowledge Management: Evaluation of Automated Link
Discovery in the Wikipedia. In Proceedings of the SIGIR 2007
Workshop on Focused Retrieval, 2007.
INEX LTW, 2007. INEX Link The Wiki Track, 2007. http://inex.
ISAAR (CFP), 2004. International Standard Archival Authority
Record for Corporate bodies, Persons and Families. International
Council on Archives, Ottawa, second edition, 2004.
ISAD(G), 1999. General International Standard Archival
Description. International Council on Archives, Ottawa, second
edition, 1999.
Jenkins, N., 2007. Can We Link It.

If this content appears in violation of your intellectual property rights, or you see errors or omissions, please reach out to Scott B. Weingart to discuss removing or amending the materials.

Conference Info


ADHO - 2008

Hosted at University of Oulu

Oulu, Finland

June 25, 2008 - June 29, 2008

135 works by 231 authors indexed

Conference website:

Series: ADHO (3)

Organizers: ADHO

  • Keywords: None
  • Language: English
  • Topics: None