|High Energy Physics Libraries Webzine|
HEP Libraries Webzine
Issue 5 / November 2001
Jean-Blaise Claivaz (*), Jean-Yves Le Meur (*), Nicholas Robinson (*)
15 Oct 2001
For many years, CERN has been collecting all High Energy Physics documents to make them easy to access to physics researchers. A repository of up to 170,000 electronic documents is available via the CERN Document Server , the main gateway to CERN's digital library. On top of the creation of this digital archive, the computing support has been looking towards possible improvements in order to ease the task of searching through and reading the articles kept. In addition to the acquisition, cataloguing and indexing of standard metadata, specific treatments have been applied to the data itself.
In this paper, after a brief description of process applied to fulltext data,
we shall focus on the specific work done within a collaboration between CERN,
Geneva University and the University of Sunderland in order to successfully
achieve the automated acquisition of structured citations from fulltext
When fulltext documents are received on the CERN Document Server (CDS), either via a direct submission  or via a batch upload procedure , the standard process is triggered which runs a set of procedures according to the rules shown below. This is completely separate from treatments done directly on metadata.
Automatically enriching the metadata of a document with the set of all citations made within it is interesting for many reasons. Various similar enrichments have been made in the same way in the past. For example, a program developed at the Deutsches Elektronen-Synchroton DESY (called Giva) helps library cataloguers to obtain all authors of a paper by extracting the complete list from the fulltext (which may even be up to several thousand people for large collaborations). The fulltext is also used in a more recent project, which is concerned with acquiring keywords from documents in an automated way .
The CERN Citation project is therefore not our first experience of working with fulltexts to enrich the metadata kept about a document. The specific issues to be addressed by the project are that:
1- It is not possible to ask of those submitting documents, the heavy task of
submitting references separately (as is done for the abstract). This would
be far too time consuming for them.
2- The position and format of references within a document is almost unpredictable, due to the huge variations in the ways in which different authors cite documents.
3- The exploitation of a good database is considerable.
If the first point above is obvious (we propose as an option, the inclusion
of citations when submitting documents to CERN Document Server and ... we never
get them that way!), the second point is being improved thanks to a SLAC
initiative, pushing HEP authors to use a standard format for citations . This is also difficult to achieve, in particular at CERN where
authors from very large world-wide institutes may not be easily convinced to
follow a "citation-writing" policy. The third point will be detailed
later, with its three main consequences: improving the calculation of eprints'
impact, enabling one to search terms appearing in references, and allowing the
navigation to cited documents.
Of course, CERN is not the only organisation to be interested in the exploitation of document citations.
Beyond the task of isolating "well"-defined references, the automated linking of related information within collections of documents is being investigated in many places. Let us quote here a few initiatives, such as the LIGHT project, started at CERN and taken over by industry , the SFX technology, proposing a solution for reference linking , the Crossref repository  using the DOI identifiers  and the S_Link_S XML based linking system . These are various approaches to help semi-automated linking between documents, based on meta-data and link management technologies. They do not focus particularly on the references written by authors in their documents.
Still, in the HEP domain where there is a long tradition of eprints (research papers freely available), various initiatives exist with the same and precise goal to reach a comprehensive coverage of eprints citations. Mentioned below are some of the projects in this field, but some may have been omitted due to its extremely active nature.
1- Science Citation Index, by Institute for Scientific
The Science Citation Index  keeps citations of papers in virtually all scientific journals (not just physics) since 1982. It is accessible only to subscribing institutions, either electronically or in paper form. Academic libraries often subscribe to this professional tool. It does not cover papers which are not published (conference articles, etc.) and it is not free of charge.
On top of the (Los Alamos, then Cornell) ArXiv.org eprint archives, SLAC has built a database of references and a search system enabling the counting and ordering of the most cited papers .
The following is the warning that prevents abusive interpretation of the results.
"The citation search should be used and interpreted with great care. At present, the source for the citation index in the HEP database is only the preprints/eprints received by the SLAC Library, and not the (unpreprinted) journal articles. Citations to a paper during the months it was circulated as a (non-eprint) preprint may also be lost, because only references to journal articles and e-print papers are indexed. Still, the citation index in HEP (SPIRES-SLAC) is formed from an impressive number of sources. For example, in 1998, the citation lists were collected from almost 14,000 preprints."
This project, "Reference Linking and Citation Analysis for Open Archives" , is a collaboration between Southampton University, Cornell University and the Los Alamos National Laboratory. Among other goals, one of them was to enrich fulltext documents from the ArXiv.org mirror site of Southampton with all references linked inside the PDF files and also to derive rules related to the impact of eprints .
4- Research Index, from NEC Research Institute:
ResearchIndex is a digital library that aims to improve the dissemination, retrieval, and accessibility of scientific literature. Specific areas of focus include the effective use of the Web, and the use of machine learning. Autonomous Citation Indexing (ACI)  automates the construction of citation indexes (similar to the Science Citation Index (R)). It has a more general scope than High Energy Physics and it is not based on a special collection, as documents can be directly retrieved from the Web.
It may look as though the work carried out at CERN is redundant with the projects mentioned above! Actually, the first analysis and development started at CERN was in 1994, with the first construction of a "CIT" database, containing only raw references of electronic documents for which the automatic parsing was successful. The interesting feature of this database  is that it is possible to look for any term (author names, report number or title) and obtain details of the papers that use this term within their references. It is complementary to the SLAC citation system where not all of the text is kept and indexed, but only the pointer to the corresponding article. Another difference between CERN and SLAC citation treatment is that while SLAC is pointing the references to its own database (where the preprint and its publication information are available), CERN decided to link journal article references directly to the e-journal site, whenever available to CERN members.
The scope of the project covers all of the CERN Document Server, which not only contains documents from ArXiv but also many CERN preprints, internal and scientific notes.
At CERN, no human resource for manual editing was allocated to the project for the long term, making mandatory the building of a more and more complete and complex acquisition algorithm. A new step was reached this summer 2001 thanks to a successful collaboration with a librarian from the University of Geneva (CH) and a student in computing from the University of Sunderland (UK). The reference analysis, the technological choices and the software have been deeply studied and completely renewed.
This is described in detail in the next part of this article.
The Reference Extraction Process
As already explained, upon receipt of a new document on the CERN Document Server, a conversion is made from its original format to the PDF format. This is mainly done because these files are of a platform-independent nature and because they are relatively small in file size allowing space on servers to be used more efficiently. They are therefore ideal for viewing by users of the service, but not when it comes to extracting references. At that point, it becomes necessary to have a more simple form of file to search through: the plain text format. Even if PDF is of a complicated nature, not organised in a linear fashion (as read by humans), but rather as a series of reference tables that point to different byte locations at which the various objects that make up the file are stored, it was deliberately chosen as a format from which to create the plain text document due to its stability over other file formats such as PostScript. Many conversion tools were tested to find the most suitable one, and the « pdftotext » tool from Foolabs  was eventually chosen. A paper about the comparison study of these different tools with their advantages and drawbacks has also been released . Having converted the document into a plain text format, the extraction process can be started.
2- Extracting the references section
The way the extraction script works is easy to understand: starting from the
very last line of the document, the program scans it upwards, searching for
the beginning of the references section, usually indicated by words such as
'References', 'Bibliography', etc. Having found this references section
title, the script reads down the text and extracts all the lines until it
encounters the next section title (e.g. 'Figures') or until it reaches the end
of the document. If no references section title can be found, a second
scan is done, this time looking for the first two reference lines, but only when
those lines are numbered with 'square brace' style:
| ...................... |
The reason is that it is a fairly safe assumption that when a line beginning with '' is followed (within a few lines) by a line beginning with '' the references section has been found. The same cannot be said however for other styles of line numerator such as '1.', as they are far more commonly found within the document body.
 Preparing the LaTeX List of Publications from the SPIRES BibTeX output. http://www.slac.stanford.edu/spires/hep/bibliotex.html
 LIGHT project,
It can clearly be seen from the above that for each page break line there should be a page number (Page X) on the line above it, and a document title ( Le Meur, JY et al. From Fulltext Documents to Structured Citations: the CERN Treatment) on the line below it. The program would be able to detect this pattern, and thus be able to remove this unwanted information. The technique is not perfect however. Often in documents, authors place titles for each section of a document in the page headers. This would mean that the page headers would differ throughout the document, and the program would not therefore be able to identify the headers - they must effectively remain fixed throughout the document. The point here, however, is that nothing is lost from this process when it fails, but when it succeeds, much is gained from it.
The process is also often able to prevent other recurring information traps, such as the repeated presence of the word "References" in the document (for example in the current title of the chapter). However, limitations are clear. A well formatted preprint with clearly indicated chapters and numbered references leads to a good result. On the other hand, if the document has bad pagination with figures inserted in the middle of the references section, which often occurs, then the result becomes unpredictable.
3- Rebuilding individual reference lines that may have been broken due to line wrapping
Having located and extracted the references section from the document body, there still remains one necessary task before the process of recognition of the cited items can begin. As is apparent to anyone who is familiar with academic papers, there are often many reference lines present in a document. Often, a given reference line can be rather long, as it can cite several documents, or simply details a large title or many authors of a paper. When a large line such as this is viewed on screen, it is broken across one or more lines of text, depending upon how long it is and how large the 'canvas' size of the document is set to be. When a human reads such a line, it is apparent to them that this reference line that stretches across several actual lines of text, is actually only one true line - it has simply been wrapped for convenience sake, as it would stretch off the printable boundaries of the page were it all shown on one line. Take for example, the following reference line:
| H. Van de Sompel, P. Hochstenbach. "Reference Linking
in a Hybrid Library Environment, Part 2: SFX, a Generic Linking
Solution'' D-Lib Magazine |
(April 1999), http://www.dlib.org/dlib/april99/van_de_sompel/04van_de_sompel-pt2.html
It is clear that it is really only one line, but has been broken for convenience sake. However, a computer program is not aware that this is really supposed to be one line, because during the conversion-to-text process, carriage return characters are inserted at the point at which lines are wrapped. Carriage return characters are also inserted at legitimate line break points as well, however, which means that it is not possible to determine between what is really a line wrap, and what is a genuine line break. This means that long reference lines are broken during the conversion-to-text process. At first sight, this breaking of long lines may not seem to be a problem. However, the process for the recognition of citations within reference lines operates on a line by line basis. This means that an accidentally broken reference line will be considered as more than one reference line by the citation recognition process. Consider the broken reference line shown above. It can be seen that the date of the article has been separated from the title. This would mean that the citation recognition would fail due to the absence of the date in the citation. In short, it is necessary to rebuild broken reference lines so that citation information is not destroyed and thus lost.
In order to rebuild reference lines, two main cases are considered. The first case is when the reference lines have no form of markers to identify the start of a new reference line. This is the most difficult case in which to rebuild the reference line, and basically involves attempting to use blank lines between reference lines (if present) to rebuild reference lines, of co-ordinate information taken from a PostScript version of the document (if available) to rebuild the individual reference lines. In this situation, if the reference lines cannot be rebuilt correctly, they are all simply joined to form one very large reference line. This solution is perfect from the view of citation instance recognition, as the recognition process simply receives one large line, and searches through it for citation information. However, from the point of view of human reading of this line, the solution is messy, as it is difficult to search through a huge block of text for one single citation item. It is not a terrible situation, however, as large lines can be split up into smaller, more manageable lines at display time.
The second situation is when the reference lines start with markers of some description (such as '*', '', '(1)', etc.). In this case, the program can simply join lines together until encountering a line beginning with the identified marker type, at which point it can make a split, having identified the start of a new reference line. At this stage, having rebuilt all reference lines, the process of identifying and tagging cited items within the lines can begin.
The reference extraction
script was run during summer 2001, extracting the backlog files of
fulltext documents starting in 1994. From 102,530 PDF files stored in the
CERN database, 94,221 references sections were extracted, reaching a total of
91.90% success. Non-English articles or documents without a references section
were also counted as failures.
The Recognition of Citations Within Reference Lines
There are essentially three types of citation item recognised by the process at this stage. These citation types are the internet address of the item (the URL), the report number of an item, and the title of the journal/periodical in which the item appears. The details of each of these items are recognised and recorded so that the information can be used.
1- Recognising Internet Addresses
Since the World Wide Web has become a very common research medium for academics, due to the wide range of information available from it both easily and immediately, it has become more and more common for authors to include the internet addresses of items that they have cited. This is often done in one of two ways:
Internet addresses are easily recognisable, using a form of pattern matching. It is possible to search for the HTML anchor tags surrounding something which resembles an internet address, or simply to search for the plain internet address format (e.g. http://........, or ftp://......., etc.). When a citation to an internet address is found with the HTML tagging information, both the address and the description ("The CERN Web site" in the above example) can be recorded. When it is simply a plain internet address however, obviously only the internet address is recorded, as there is no description present.
2- Recognising Report Numbers
3- Recognising Journal/Periodical Titles
The biggest step during the citation recognition process is that of recognising the Journal/Periodical title details. The reason for this is that more elements must be taken into account. In order to make a link to an item cited in a Journal/Periodical, it is necessary to know the Journal/Periodical title, the series name, the volume, the year of print, and the pagination information. The title of the Journal/Periodical is identified using a knowledge base of Journal/Periodical titles that CERN either knows about, or subscribes to. This knowledge base is of a rather large nature, as for each Journal/Periodical title, there is a standardised way of writing the title, along with many non-standard ways of writing the title. Unfortunately, authors seem quite unaware or unconcerned with the standard ways of writing the titles, so it is necessary to record the non-standard forms of titles that have been encountered, along with the standard. This way, when a non-standard title form is encountered by the program, it can be recognised and replaced by a standard form, which is then recorded for later use. The standard form of writing a title is set out by the ISO 4 standard. Currently, about 1800 entries are kept in the titles knowledge base for about 800 actual titles. In addition to the recognition of the title information, it is of course necessary to recognise the series name, volume, year of print and pagination information (collectively known as the numeration information). This numeration information is identified and recorded by the use of several standard models. Unfortunately, with this numeration information, there are many non-standard ways of writing it (writing the components of it in different orders), and it is necessary to use the different models to recognise these non-standard forms, and change them into a standard format - the sequence "volume (year) page". Notice that there is no punctuation in this standard form.
The matching against this list allows the script to recognise and standardise the titles with a high reliability. Take for example, the following:
Among the 94 211 eprints for which the reference section extraction was
successful during summer 2001, almost 3 million (2 896 541) individual citations
were collected and the recognition process led to the following:
These results give a good picture of the trends followed in terms of citing habits in HEP over the past seven years or so. It can clearly be seen that journal articles are massively quoted (~70%), URLs are very rarely provided (0.44%), and report numbers are often used (~13%). A closer look at report number recognition shows an obvious increase in the use of report numbers in citations throughout the years: during the years 1994 to 1999, 11.3% of citations contain report numbers, whereas during the years 2000-2001, this number has reached 18.6%. It is probable that this results from the increasing availability of free literature (eprints to which authors refer with report numbers).
Probably, the most important factors of this result are that we have now the
- create 379 440 links from the references to the eprints
- create 12 684 links with the provided URLS
- create up to 2 329 286 links to the journal articles, if available on-line, or to their corresponding eprint, if available on the CERN Document Server.
The Linking of Citation Items to Their Respective Fulltexts
Older CERN Citation Recognition System:
| I. Lauer et al., Phys. Rev. Letters 81, (1998), p. 2052
 M.B.H.Breese, Nucl. Instrum. Methods Phys. Res., B : 132 (1997) 540
W.H. Zureck, Phys. Rev. D 24, 1516 (1981); W.G. Unruh and W.H. Zureck, ibid. 40, 1071(1989)
 J.L. Feng, Int. J. Mod. Phys. A13, 2319 (1998) [hep-ph/9803319]
 D. E. Kaplan and A. E. Nelson, J. High Energy Phys. : 0001 (2000) 033 [arXiv:hep-ph/9901254]
| I. Lauer et al., Phys. Rev. Lett. : 81 (1998) 2052
 M.B.H.Breese Nucl. Instrum. Methods Phys. Res., B: 132 (1997) 540
 W.H. Zureck, Phys. Rev., D : 24 (1981) 1516; W.G. Unruh and W.H. Zureck, Phys. Rev., D : 40 (1989) 1071
 J.L. Feng, Int. J. Mod. Phys., A :13 (1998) 2319 - hep-ph/9803319
 D. E. Kaplan and A. E. Nelson J. High Energy Phys.: 0001 (2000) 033 [arXiv [hep-ph/9901254]
The end result of this project can be seen on the Web as it has now become
part of the CERN Document Server. From the CDS Search interface (http://weblib.cern.ch), when retrieving
eprints, the "Show References" links point to these automatically linked
Considerable time and effort has been invested in the area of reference extraction and recognition at CERN. The result has been the development of a system that is considerably more robust, and able to extract references from many more documents than was possible under the older system. Under this new system, the number of cited items within given reference lines has also been greatly increased, resulting in a better situation for both CERN and its users. The citation information that has been extracted is organised in a very structured and efficient manner. Not only does this make for better system performance and less wasted space in the CERN bibliographic information databases, but it also allows the exploitation of the data in order to provide many more services to the user.
These services are now being investigated. Possible examples of them are:
 CERN Document Server, http://cds.cern.ch/
 T. Baron, J-Y Le Meur. "From Desk to Web: the Electronic Document Submission", presentation at CERN, Aug 1999
 N. Pignard, I. Geretschläger, J. Jerdelet. "Automation of electronic ressources in the Scientific Information Service at CERN", High Energy Phys. Libr. Webzine : 3 (2001) 003
 D. McGlashan, J-Y. Le Meur. "The CERN Conversion Service", presentation at CERN, Aug 1999
 D. McGlashan, J-Y Le Meur. "SetLink: the CERN Document Server Link Manager", CERN-ETT-2000-001, High Energy Phys. Libr. Webzine : 1 (2000) 1
 Fulltext searching of HEP eprints, http://weblib.cern.ch/Fulltext
 D. Dallman, J-Y Le Meur. "Automatic keywording of High Energy Physics", CERN-AS-99-005, 4th International Conference on Grey Literature : New Frontiers in Grey Literature, Washington, DC, USA, 4 - 5 Oct1999 / Ed. by Farace, D J and Frantzen, J - GreyNet, Amsterdam, 2000. - pp.230-237
 Preparing the LaTeX List of Publications from the SPIRES BibTeX output. http://www.slac.stanford.edu/spires/hep/bibliotex.html
 LightObject, sharing knowledge, http://www.lightobjects.com/
 H. Van de Sompel, P. Hochstenbach. "Reference Linking in a Hybrid Library Environment, Part 2: SFX, a Generic Linking Solution''
D-Lib Magazine (April 1999)
 Crossref, The central source for reference linking, http://www.crossref.org
 The Digital Object Identifier System, DOI, http://www.doi.org
 E. Hellman. ScholarlyLink Specification Framework (S-Link-S), http://www.openly.com/SLinkS/
 Science Citation Index, http://www.isinet.com/isi/
 arXiv.org e-Print archive, http://arxiv.org
 SLAC Spires Collection of References, http://www.slac.stanford.edu/spires/hep/references.html
 Reference Linking and Citation Analysis for Open Archives, http://opcit.eprints.org/
 P. Ginsparg, J. Halpern, C. Lagoze, S. Harnad, W. Hall, L. Carr. "Integrating and navigating eprint archives through citation-linking: the open citation (OpCit) linking project"
 S. Lawrence, C. Lee Giles, K. Bollacker. "Digital Libraries and Autonomous Citation Indexing", IEEE Computer, Volume 32, Number 6, pp. 67-71, 1999
 The CERN CIT database, References in E-prints, http://weblib.cern.ch/Home/References_in_E-prints/
 A PDF Viewer for X. http://www.foolabs.com/xpdf/
 N. Robinson. "A Comparison of Utilities for Converting from PostScript or Portable Document Format to Text", CERN-OPEN-2001-065 ; Geneva : CERN , 31 Aug 2001
 MARC 21 - MARC formats for the 21st century
 E. Lodi, M. Vesely, J. Vigen. "Link managers for grey literature", CERN-AS-99-006, 4th International Conference on Grey Literature : New Frontiers in Grey Literature, Washington, DC, USA, 4 - 5 Oct 1999 /Ed. by Farace, D J and Frantzen, J - GreyNet, Amsterdam, 2000. - pp.116-134
 Direct Impact: Request on the frequency of download for documents on CERN Document Server, http://doc.cern.ch/impact/
Jean-Yves Le Meur
CERN, European Organization for Nuclear Research, Switzerland
Jean-Yves Le Meur works at CERN in the Education and Technology Transfer division to develop and promote the CERN Document Server: a set of databases and Internet & Intranet software for storing and delivering scientific documentation to the HEP community. He is the Deputy Group Leader of the Document Handling Service leading the Photo, Video, Desktop Publishing, Printshop and CDS sections.
University of Sunderland, UK
Nicholas Robinson is currently studying in the final year of a BSc Computing course at the University of Sunderland in the UK. As part of
the third year of this course, he spent fourteen months working as a technical student with the CDS team at CERN. His final year of studies at
university includes work in areas such as Artificial Intelligence, Object-Oriented Development and Software Engineering Methods.
For citation purposes:
Jean-Blaise Claivaz, Jean-Yves Le Meur, Nicholas Robinson, "From Fulltext Documents to Structured Citations: CERN's automated Solution", High Energy Physics Library Webzine, issue 5, November 2001