|High Energy Physics Libraries Webzine|
Issue 10 / December 2004
Most information retrieval systems nowadays use full-text searching because algorithms are fast and very good results can be achieved. However, the use of full text indexes has its limitations, especially in the multilingual context, and it is not a solution for further information access requirements such as the need for efficient document navigation, categorisation, abstracting and other means by which the document contents can be quickly understood and compared with others. We show that automatic indexing with controlled vocabulary keywords (descriptors) complements full-text indexing because it allows cross-lingual information access. Furthermore, controlled vocabulary indexing produces document representations that are useful for human users, for the Semantic Web, and for other application areas that require the linking and comparison of documents with each other. Due to its obvious usefulness, controlled vocabulary indexing has received increasing attention over the last few years. Aiming at a better understanding of the state-of-the-art in this field, we discuss the various approaches to automatic keywording and propose a taxonomy for their classification.
We launch our web browser and, after clicking on a bookmark, a one-field form appears embedded in the page. Once a few words are typed inside the text field, we click on the `submit' button expecting the answer to our question. A few seconds later the browser shows a page containing a list of items: those the system considers most suitable to answer our needs. The discrimination of results becomes a non-trivial operation due to the large number of entries returned. Sometimes we can get rid of some of them at a glance: the title or the text provided along with the item is enough to know we are not interested, but sometimes we have to click and check the real document to see whether it is the information we want, or not.
Many of us will recognize the sequence of steps performed above. We were looking for information using a full-text search engine. This operational mode in information searching and retrieval has populated almost every digital system which stores information. We can find forms like the one described when:
Though the usefulness of full text search engines has been widely proven and, therefore, accepted, they are still not good enough in some cases and totally inappropriate in others. The first kind of less-successful cases are those where the collection contains a huge range of subjects and documents: for example, the World Wide Web. Old approaches using purely full-text-based engines were abandoned, since the quality of results provided was declining with the growth of the collection. Therefore, new techniques arose with the aim of filtering and re-qualifying the rank (the Page Rank algorithm is one of the most successful examples ). They index every word in a page so they can perform full-text searches later. The problem with this approach is that language is complex, ambiguous and rich in variation, so the quality of the results is still not as good as we would like. But this technique of indexing is solving the big problem of searching for information on the web. It is an implementable solution in very general contexts.
The second field where full text-search techniques do not do so well is when textual information is not available. There are still some kinds of collections which are not suitable (yet) for this genre of engines. We refer here to pieces of information like images, sounds, etc. The current solution is to provide, beforehand, textual information related to every item (that is, enrich the data with text) so that later we can search using this related text as an access point. Many techniques have been developed in order to automate such a process by pattern recognition, clustering and so on.
Imagine you had to organize your personal library, what sort of ideas do you think you would try in order to achieve well organized shelves? Maybe one of your first ideas would be to group books by theme, then to label them and put their details in a kind of index. Later on you might find you have so many books, it would be better to arrange them by size (large repositories do so). Whatever method you used, in the end you would have to index them in one way or another. Now the question could be: which indexes should I use? It is not an easy task to define them because several considerations must be taken into account. Vickery already emphasizes this reality :
The problem of subject representation is therefore far less straightforward than other aspects of document description.
In the beginning, the use of keywords for information storage and retrieval was due to two major needs: the need for classification and the need for retrieval. The former need had a double benefit: first, it let librarians organize physical volumes into logical clusters; second, the possibility to search within a defined cluster was regarded as a way to speed up the searching for information (as pointed out by the so-called 'cluster hypothesis' of Rijsbergen ).
Hence, two major goals of indexing are to:
Alphabetical terminologies and classification structures (known as 'thesauri') were thought of as tools to improve the two main measures in information retrieval: precision and recall. These refer to the quality of retrieved documents when compared to the search query. 'Precision' is the number of relevant documents retrieved over the total number of documents retrieved. 'Recall' is the number of relevant documents retrieved over the total number of relevant documents in the collection. These two measures show the problem of an antagonistic relationship: if we try to improve one of them, the other will decay. For example, if we retrieve the whole collection in answer to a given query, our recall will be 100%, but our precision will be so low that the result will be unusable. The challenge resides, then, in finding a method which shows a good performance for both measures.
In earlier times, techniques were used to improve these two values for a defined retrieval system; i.e. the implementation of these techniques was oriented to the purpose and content of the retrieval system. The techniques traditionally used rely on setting relationships between words in a controlled vocabulary. Using those relations on a given query we can improve recall (by expanding to related terms) or precision (by narrowing with less generic terms). These are the reasons for the use of thesauri.
There are several definitions for the word 'thesaurus'. In an old work of Vickery  we find a definition for thesaurus which summarizes in a few words the rationale associated with it:
''The thesaurus is a display of the terms in a retrieval language showing semantic relations between them.''
Here, Vickery shows, on the one hand, the main purpose of a thesaurus: it defines a retrieval language, whatever the retrieval method might be. On the other hand, he does not define the kind of relationships between entries (synonyms, broader terms...), specifying only that a set of semantic relations is defined. We will see that this brief definition fits perfectly with any type of existing thesaurus.
One of the earliest thesauri (and maybe the most famous one) is Roget's Thesaurus . Dr. Peter Mark Roget's main idea behind this compilation was to create a system which would offer words to express a given meaning, while conversely traditional dictionaries offer meanings for a given word. This would help writers to express their concepts in the most suitable form. Such users had the thesaurus as a reference book for writing of texts. Thus, it was mostly intended to be useful in the document creation phase.
The power of reducing a language to its basic concepts has become more and more useful, especially since the "semantic network" has arisen in electronic form. WordNet  is an on-line reference system" (their authors state). English nouns, verbs, adverbs and adjectives are organized into synonym sets (also called synsets), each representing one underlying lexical concept. Nowadays we can assume that almost every thesaurus (specialized or not) is available in electronic form.
Thesaurus descriptors are normally unambiguous because they are clearly defined, whereas full text indexing does not provide any differentiation for words such as 'plant' in 'power plant' versus 'green plant'.
There is even a multilingual thesaurus based on WordNet called EuroWordNet , which, using English as central node, maps synsets between different European languages. This work represents a milestone in multilingual information retrieval.
Both WordNet and Roget's Thesaurus are general reference sources, i.e. they don't focus on specialized terminologies. But the areas where thesauri become useful tools are in specialized domains (Law, Medicine, Material Science, Physics, Astronomy...). One example is the INSPEC thesaurus , focused on technical literature; or the ASIS thesaurus, specialized in Information Science . NASA, the European Union, and other organizations produce their own specialized thesauri (like the multilingual EUROVOC thesaurus ).
Each thesaurus has its own organization, according to the purpose it needs to accomplish. But we can summarize any of them by the following components:
There are two relations which are commonly used among existing thesauri:
Of course, depending on the purpose of the thesaurus, some of these relations may be ignored. Also new relations could occur. WordNet, for example, includes all of the given relations. INSPEC and Eurovoc thesauri condense meronym relations into the ''related'' relationship (see , RT means ''related terms''). Synonymy is implemented by the application of the ''USE'' statement.
[...] penalty NT1 alternative sentence NT1 carrying out of sentence NT2 barring of penalties by limitation NT2 reduction of sentence RT repentance RT terrorism (0431) NT2 release on licence NT2 suspension of sentence NT1 conditional discharge NT1 confiscation of property RT seizure of goods (1221) NT1 criminal record NT1 death penalty NT1 deprivation of rights RT civil rights (1236) [...]
Usually in specialized thesauri either the synonymy is neglected or a preferred word representing the meaning is given, since the purpose is to provide a list of controlled terms (and that ''control'' refers to the use of just one word for a given meaning). Nevertheless, most of them include synonymy in one way or another.
There are, however, some special cases of thesauri where there is more than just terms and relations. In some cases the thesaurus is a complex reference of specific relations, with specially defined rules to build a document's keywords. This is the case of the DESY thesaurus , specializing in high energy physics literature. With the entries given we can construct particle combinations, reaction equations and energy declarations among other examples.
These facts bring us to the conclusion of Vickery that there is a tight relationship between the thesaurus and its domain of retrieval.
The construction of hand-crafted thesauri for use in computer applications dates back to the early 1950s with the work of H. P. Luhn, who developed a thesaurus for the indexing of scientific literature at IBM. The number of thesauri and systems is now growing steadily because manually or automatically keyworded documents have many advantages and offer additional applications over simple documents not linked to thesauri. Depending on whether people or machines make use of keywords assigned to documents, we distinguish the following uses:
For high energy documents this can speed up the search process in specialized collections which grow by hundreds of documents every week .
If the thesaurus is multilingual, this summarising function also works across languages, i.e. a user will see a list of keywords (a summary) in their own language of a document written in another language .
However, the use of keywords for classification goes beyond the pure retrieval domain. The freedesktop.org project  promotes the use of keywords for the arrangement of icons (representing application launchers) in the main menu of the desktop where applications are internally attached to a list of categories. It means that there is no predefined taxonomy to which program launchers are classified. Instead, programs are labelled with keywords from which menus are created in the graphical interface of an operating system (like the Gnome  desktop available for Linux and other operating systems).
Therefore, we could imagine a scenario where we want to look for a service, e.g. a database of iron manufacturers. We get the keywords of the service which may have been generated from the content of top web pages in the portal of this service (the pages which let us access the database via web forms or any other web based interaction). These keywords show us that there is another database which offers iron toys, since the thesaurus splits the keyword iron manufacturer into the subtopics iron-made toys manufacturer, naval manufacturer, etc. Thanks to the semantic network created from keyword relationships we are able to find the provider we need.
F. W. Lancaster  gives us the following definition for indexing:
''The main purpose of indexing and abstracting is to construct representations of published items in a form suitable for inclusion in some type of database.''.
This is a very general description, but it still summarizes the goal of indexing: provide an abstraction of document contents for better storage and retrieval (which is the goal of any database). We find several types of indexing. For example, web search engines (like Google, Altavista and others) generally full-text-index web pages automatically, but for some specialized and popular subject areas, they ask teams of professional indexers to carry out careful manual indexing.
We distinguish two main types of indexing :
The taxonomy we propose focuses on those systems that do assignment of keywords instead of extraction.
Figure 2 provides a graphical view of this process. The indexer reads the document and selects keywords from a thesaurus or a controlled vocabulary.
Although there has been extensive work done on automatic thesauri generation, less work has been done on automatic descriptor assignment. Although research is advancing slowly in this area, it benefits from development in other IR (information retrieval) areas. We mention here some of the systems developed for automatic assignment, along with a taxonomy proposal for this kind of tool.
We can classify automatic keyword assignment systems (AKWAs) into two main categories, depending on their degree of automation:
The following is a taxonomy that summarizes the various AKWA approaches developed so far:
Advantages of this relatively simple approach are its relatively easy implementation and the fact that, depending on the kind of merge function used, training may not be necessary. A disadvantage is that the merging function might be difficult to implement.
An example for this category is the system developed by Ezhela et al. , which is fully based on the references linked to a publication. Even with a quite simple approach, the results were acceptable, due to the high specialization of the subject domain (High Energy Physics). The merging strategy used was a pure intersection operation.
Most current systems fall under this category. One of them is the first version of the HEPindexer system , which uses a vector space model fully dedicated to the automatic assignment of descriptors.
Finally, the last criterion for classifying AKWAs is based on training needs, i.e. on the amount of effort required to develop the system:
An example of this kind of approach is the European Commission's EUROVOC indexing tool , which represents each class by a combination of the most frequent lemmatized keywords of all of its documents, uses a large number of parameters and calculates various types of vector space similarities between the new document and all of its classes in order to determine the most appropriate descriptor. A more standard approach to Machine Learning was taken by Hulth , but her work is limited to keyword extraction rather than assignment.
A general, positive feature of these systems is that they can rather easily be adapted to new thesauri and languages as they automatically optimise the selection of features from the feature set provided as input. Language-independent systems typically work with little linguistic input (e.g. only with a text's frequency list of word tokens). Better performance can be achieved using more linguistic processing such as lemmatization and parsing or using linguistic resources such as synonym lists and other dictionaries.
We have shown that manual or automatic indexing of document collections with controlled vocabulary thesaurus descriptors is complementary to full-text indexing and that it provides both human users and machines with the means to analyse, navigate and access the contents of document collections in a way full-text indexing would not permit. Indeed, instead of being replaced by full-text searches in electronic libraries, a growing number of automatic keyword assignment systems are being developed that use a range of very different approaches.
In this paper, we have given an introduction to automatic keyword assignment, distinguishing it from keyword extraction, and proposing a classification of approaches, referring to sample implementations for each approach. This presentation will hopefully help researchers in the area to better understand and classify emerging approaches. We have also summarized some of the powerful applications that this kind of tool is offering in the field of information retrieval, and we have structured them into comprehensive categories and have shown real examples and working solutions for some of them.
As the number of different systems for automatic keyword assignment has been increasing over recent years, it was our aim to give some order to the state of the art in this interesting and promising field of research.
 L. Page, S. Brin, R. Motwani, and T. Winograd. "The pagerank citation ranking: Bringing order to the web." Technical report, Computer Science Department, Stanford University, 1998.
 B. C. Vickery. Information Systems. London: Butterworth, 1973.
C. J. van Rijsbergen. Information Retrieval. London: Butterworths,
 S. M. Lloyd, editor. Roget's Thesaurus. Longman, 1982.
George A. Miller, Richard Beckwith,
Christian Fellbaum, Derek Gross and Katherine Miller. Introduction to
WordNet: An On-line Lexical Database. Cognitive Science Laboratory,
Princeton University, August 1993.
 P. Vossen. "Eurowordnet: a multilingual database for information retrieval" in: Proceedings of the DELOS Workshop on Cross-language information retrieval, Zurich, Mar 5-7, 1997.
ASIS thesaurus of information science.
The high energy physics index keywords, 1996.
A. Montejo-Ráez and D. Dallman.
"Experiences in automatic keywording of particle physics literature." High Energy Physics Libraries Webzine, (issue 5), November
 R. Steinberger. "Cross-lingual keyword assignment." In: L. A. U. na López, editor, Proceedings of the XVII Conference of the Spanish Society for Natural Language Processing (SEPLN'2001), pages 273-280, Jan (Spain), Sept. 2001.
The grace engine.
The freedesktop project. Homepage
The gnome website.
 L. A. Vassilevskaya. An approach to automatic indexing of scientific publications in high energy physics for database spires-hep. Master Thesis, September 2002.
 L. W. Wright, H. K. Grossetta Nardini, A. R Aronson and T. C. Rindflesch. "Hierarchical concept indexing of full-text documents in the unified medical language system information sources map." Journal of Education for Library and Information Science 50(6) :514-523, 1999.
 Ralf Steinberger, Bruno Pouliquen, and Johan Hagman. "Cross-lingual Document Similarity Calculation Using the Multilingual Thesaurus EUROVOC", in: A. Gelloukh, editor, Computational Linguistics and Intelligent Text Processing: Third International Conference CIC Ling 2002. Springer, LNCS 2276, 2002.
 Ralf Steinberger, Johan Hagman, and Stefan Scheer. "Using thesauri for automatic indexing and for the visualisation of multilingual document collections.", in: Proceedings of the workshop on Ontologies and lexical knowledge bases (Ontolex, 2000), pages 130-141, Sozopol, Bulgaria, sept 2000.
 T. Berners-Lee, J. Hendler and O. Lassila. "The semantic Web." Scientific American, 284(5) (May 2001) :34-43.
 T. H. Fran Berman, Geoffrey Fox, editor.Grid Computing: Making the Global Infrastructure a Reality.Wiley, 2003.
 F. W. Lancaster.Indexing and Abstracting in Theory and Practice.London: Library Association Publishing, 1998.
O. R. Gail Hodge.
"Cendi agency indexing system descriptions: A baseline report."
Technical report, CENDI, 1998.
 N. Vieduts-Stokolo. "Concept recognition in an automatic text-processing system for the life sciences". Journal of the American Society for Information Science, 38 (4): 269-287, 1987.
 A. Anjewierden and S. Kabel. "Automatic indexing of documents with ontologies." In: B.Krose, M. de Rijke, G. Screiber and M. van Someren, editors, Proceedings of BNAIC 2001(13th Belgian/Dutch Conference on Artificial Intelligence). Amsterdam, Neteherlands, 23-30, 2001.
 V.V. Ezhela, V.E. Bunakov, S.B. Lugovsky, V.S. Lugovsky and K.S. Lugovsky. "Discovery of the additional knowledge and their authomatic indexing via citations (in Russian). " In: third All-Russian conference, Digital Libraries; Advanced Methods and Technologies, Digital collections (RCDL'2001), Petrozavodsk, sept 11-13, 2001.
 A. Montejo-Ráez. "Toward conceptual indexing using automatic assignment of descriptors." In: Stephano Mizzaro and Carlo Tasso, editors, Personalization Techniques in Electronic Publishing on the Web: Trends and Perspectives, proceedings of the AM' 2002 Workshop on Personalization Techniques in Electronic Publishing, Malaga, Spain, May 2002.
 Bruno Pouliquen, Ralf Steinberger and Camelia Ignat. "Automatic annotation of multilingual text collections with a conceptual thesaurus." In: A. Todirascu, editor, Proceedings of the workshop 'Ontologies and Information Extraction' at the Summer School 'The Semantic Web and Language Technology' (EUROLAN'2003), Bucharest, jul 28 - Aug 8 2003.
 A. Hulth. "Improved automatic keyword extraction given more linguistic knowledge." In Proceedings of the Conference Empirical Methods in Natural Language Processing (EMNLP'2003), Sapporo, Japan, July 2003.
Marjorie M.K. Hlava and Richard Hainebach.
Multilingual machine indexing.
Tel: +41 (0) 22 767 3833
Joint Research Centre
T.P. 267, 21020 Ispra (VA)
Tel: +39 - 0332 78 6271