54

User Interfaces for Digital Libraries

Andreas Rauber, A Min TjoaDepartment of Software Technology

Vienna University of TechnologyFavoritenstr. 9 - 11 / 188A - 1040 Vienna, Austria

{rauber,tjoa}@ifs.tuwien.ac.at

Abstract: Digital Libraries have gained tremendous interest with several research projects addressing the wealth of challenges in this field. While computational intelligence systems are being used for specific tasks in this arena, the majority of projects relies on conventional techniques for the basic structure of the library itself. With the SOMLib project we created a digital library system that uses a neural network-based core for the representation of the library. The self-organizing map, a popular unsupervised neural network model, is used to topically structure a document collection similar to the organization of real-world libraries, with extended models such as the growing hierarchical self-organizing map allowing the detection of topic hierarchies. Based on this core, additional modules provide information retrieval features, and automatically label the various topical sections in the document collection. A metaphor graphics based interface further assists the user in intuitively understanding the library, providing an instant overview.

Keywords: User Interface, Digital Libraries, Information Visualization, Metaphor Graphics.

1. INTRODUCTIONDigital libraries experience a much wider propagation than we might realize, and most of us interact with one every other day. They exist in several different forms, ranging from "real libraries" such as publicly accessible library catalogs, via bookstores on the Internet (e.g. Amazon, Amadeus, Books-Online, etc.) to information repositories making text, images, and

PDF Cre

ate! 5 T

rial

www.n

uance.

com

55

multimedia elements available on-line, allowing users to search them and retrieve information. Depending on how broad we draw the scope of a digital library, some might even consider the World Wide Web as one huge digital library, albeit a - in terms of librarianship - very badly organized and managed one.Any of these digital libraries represents a tremendous potential, be it in terms of marketing, making huge collections of books, CDs etc. available that can be ordered in a few mouseclicks, or by providing almost any piece of information that can be possible imagined. Yet, we find this potential not to be exhausted. Even worse, many of the true benefits of digital libraries are not even touched upon. Still, this should not be seen too much as a reproach. Digital libraries are relatively young - at least if compared with their conventional counterparts, which have been around in some form or other for centuries, and which continuously adapted to the needs of their users. We thus might want to take a look at these libraries, or information warehouses, to learn some lessons for digital libraries, how we should organize and present information to users to best serve their needs and to gain the biggest benefit from the way information is made available.The range of issues that may be considered in this respect is vast, covering all fields from being able to store, access, and explore these repositories, locating the required information, and finally obtaining it, including all issues related to buying physical or electronic objects, copyright issues, and so on. For the scope of this paper we will concentrate on two aspects, namely the organization of information and its visualization, i.e. firstly, how we can organize documents in a way that helps users find the correct piece of information without having to specify complicated boolean queries, and secondly, how we can present the pieces of information to the user in a way that allows them to understand, what kind of information it actually is.

The reminder of this paper is structured as follows. Section 2 takes a brief look on conventional libraries to provide some inspirations for ways how the challenges might be addressed in digital collections. Section 3 then presents an approach to automatically organize document collections according to their content using neural networks, with experimental results being presented in Section 4. This is followed by a presentation of the libViewer system, a metaphor-graphics based visualization of document collections in Section 5.

PDF Cre

ate! 5 T

rial

www.n

uance.

com

1 We have to acknowledge the fact, that this topic-oriented organization of libraries is no longer possible for

some very large libraries, as the continuous addition of new documents would cause reorganizations of the entire collection too frequently. These collections usually are merely sorted by order of acquisition, with catalogues being the primary means of access to documents. However, the reason for abandoning the topical structure is merely due to the physical limitations of documents in a collection.

_________________________

56

2. DIGITAL AND CONVENTIONAL LIBRARIESLibraries usually exhibit a clearly detectable structure by organizing books by topic into sections and shelves. This structure allows us to gain insight into the contents of the library as well as to get a rough overview of the amount of information available on specific topics1. When entering a library or large-scale book store, in spite of the overwhelming amount of information present in such locations users usually manage to orient themselves and find the way to their section of interest quite easily. Without being able to read the title of books from the far distance, not knowing actually where to find a book by a specific author or even without knowing the title or the author of a book, most people are able to locate the respective sections when looking for a dictionary, a poem collection, or a story book for children.Searching a library can take several forms: you might start browsing from the entrance via different floors to any specific section and shelve, which is then searched entry by entry. Note, that at most libraries you find a map of the library at the entrance, giving an overview of books on which topic may be found in which section. A second approach may be by searching keyword, author and title catalogues. Third, you might also ask a librarian to help you find the requested pieces of information by giving a rough idea of the desired book. The outcome of such an inquiry is usually not only a list of titles or a pile of books, but also includes some recommendations based on the experience of the librarian. Locating one book in the library usually leaves you, due to the topical structure, with several other relevant ones nearby. Once you find the corresponding shelf, by scanning the books sorted there, it is usually easy for you to tell the age of a book, the number of times it has been used before (at least in a public library rather than in a bookstore), as well as the amount and type of information to be expected in the books simply by looking at them. The cover of the book, the title, type of binding, the shape of the binding (brand new versus well-thumbed and almost torn apart), the size of the book, color and other properties of an item on the shelve contain a wealth of information that most people are accustomed to and able to interpret intuitively.PDF

Create!

5 Trial

www.n

uance.

com

57

Thus, it is easy for us to gain an overview of the contents of a library, the type of information present, how many items of a specific title can be found etc. All these features make orientation rather easy in spite of the wealth of information present.

While we find this kind of organization to be the most prominent one for public access libraries (supported by card catalogues for alphabetical listings or author searches), it is hardly found in any digital library. One of the reasons for this lies with the difficulty of being able to create such an organization automatically for large collections of documents. Wherever such manual indexing and structuring is possible, we find it to be readily included even in the electronic world, such as the topically structured collections of Amazon [W01] or Yahoo [W02]. Adopting these characteristics of conventional libraries for electronic media to combine the benefits of the evolved structures of conventional systems with the benefits of digital systems provides a challenging task, but an important step forward in making those information repositories widely usable.

3. SELF-ORGANIZING MAPSThe self-organizing map (Kohonen 1982, Kohonen 1995) is a general unsupervised tool for the ordering of high-dimensional data in such a way that similar input items are grouped spatially close to one another. The model consists of a number of neural processing elements, i.e. units. Each of the units i is assigned an n-dimensional weight vector mi, mi = [mi1, mi2, ..., mi

n]T, mi n. It is important to note that the weight vectors have the same dimensionality as

the input patterns. The training process of self-organizing maps may be described in terms of input pattern presentation and weight vector adaptation. Each training iteration t starts with the random selection of one input pattern x(t). This input pattern is presented to the self-organizing map and each unit determines its activation. Usually, the Euclidean distance between the weight vector and the input pattern is used to calculate a units activation. In this particular case, the unit with the lowest activation is referred to as the winner, c, of the training iteration, as given in Expression (1).

c : mc(t) = arg mini || x(t) - mi(t) || (1)

Finally, the weight vector of the winner as well as the weight vectors of selected units in the vicinity of the winner are adapted. This adaptation is implemented as a gradual reduction of

PDF Cre

ate! 5 T

rial

www.n

uance.

com

58

the difference between corresponding components of input pattern and weight vector, as shown in Expression (2).

mi(t+1) = mi(t) + (t) hci(t) [x(t) - mi(t)] (2)

Geometrically speaking, the weight vectors of the adapted units are moved a bit towards the input pattern. The amount of weight vector movement is guided by a so-called learning rate,

, decreasing in time. The number of units that are affected by adaptation is determined by a

so-called neighborhood function, hci. This number of units also decreases in time such that towards the end of the training process only the winner is adapted. A Gaussian may be used to model the neighborhood function. It is common practice that at the beginning of training a wide area of the output space is subject to adaptation. The spatial width of units affected by adaptation is reduced gradually during the training process allowing the formation of large clusters at the beginning and fine-grained input discrimination towards the end of the training. The movement of weight vectors has the consequence, that the Euclidean distance between input and weight vectors decreases and thus, the weight vectors become more similar to the input pattern. The respective unit is more likely to win at future presentations of this input pattern. The consequence of adapting not only the winner alone but also a number of units in the neighborhood of the winner leads to a spatial clustering of similar input patters in neighboring parts of the self-organizing map. Thus, similarities between input patterns that are present in the n-dimensional input space are mirrored within the two-dimensional output space of the self-organizing map. The training process of the self-organizing map describes a topology preserving mapping from a high-dimensional input space to a 2-dimensional output space where similar patterns are mapped to geographically close locations in the output space.

The SOM has been used repeatedly for the unsupervised classification of free-form text documents, cf. (Merkl 1998, Rauber et al. 1999, Kohonen et al. 2000). Text documents can be thought of topical clusters in the high-dimensional feature space spanned by the individual words in the documents. A trained SOM thus represents a topical ordering of the documents, meaning that documents on similar topics are located close to each other on the two-dimensional map. This is comparable to what one can expect from a conventional library, where we also find the various books ordered by some content-based criteria. Thus, the SOM offers by its very architecture an ideal way for the organization of document repositories. The items to be included in the SOMLib library system are represented in the form of feature

PDF Cre

ate! 5 T

rial

www.n

uance.

com

2 We use the notion (x/y) to refer to the unit located in row x and column y of the map,

_________________________

59

vectors, which are created by parsing the texts and processing the resulting word histograms to provide a compact representation of the texts. These feature vectors are used as input to train a standard self-organizing map. Extensions to the SOM, such as the growing hierarchical self-organizing map (GHSOM) (Dittenbach et al. 2000) further allow the detection of hierarchical structures in the topic hierarchy. Furthermore, using the LabelSOM technique (Rauber 1999) we are able to automatically extract descriptive keywords from the trained map describing the topics of the clusters, allowing the map to be intuitively read and interpreted.

4. CONTENT-BASED DOCUMENT ORGANIZATION

For the experiments presented hereafter we use the classic TIME Magazine article collection available at [W03]. It consists of a collection of 420 articles from the TIME Magazine dating from the early 1960s. This collection, while being small enough to be presented in sufficient detail, provides the benefits of a real-world article collection covering a wide range of topics from foreign affairs to high-society gossip, thus forming an ideal testbed for the evaluation of our approach. Please note, that the consecutive numbering is not complete, i.e. not all articles are available in the set.To be used for map training, a vector-space representation of the single documents is created.For each document collection a list of all words appearing in the respective collection is extracted while applying some basic word stemming techniques. Words that do not contribute to contents description are removed from these lists. Instead of defining language or content specific stop word lists, we rather discard terms that appear in more than 90% or in less than 3 articles in each collection. Thus, we end up with a vector dimensionality of 5.923 for representation of the 420 documents. The individual documents are then represented by feature vectors using a tf x idf, i.e. term frequency times inverse document frequency, weighting scheme (Salton 1989). This weighting scheme assigns high values to terms that are important as to describe and discriminate between the documents. These feature vectors are further used to train a self-organizing map consisting of 10 x 15 units.We find, that the SOM has succeeded in creating a topology preserving representation of the topical clusters of articles. For example, in the lower left corner we find a group of units representing articles on the conflict in Vietnam. To name just a few, we find articles T320, T369 on unit (14/0)2, T390, T418, T434 on unit (14/1) dealing with the government

PDF Cre

ate! 5 T

rial

www.n

uance.

com

starting with (0/0) in the upper left corner

__________________________________________________

60

crackdown on Buddhist monks, next to a number of articles on units (14/3), (14/4) and neighboring ones, covering the fighting and suffering during the Vietnam War. A cluster of documents covering affairs in the Middle-East is located in the lower right corner of the map around unit (14/9), next to a cluster on the so-called Profumo-Keeler affair, a political scandal in Great Britain in the 1960s, on and around units (10/9) and (11/9). Above this area, on units (6/10) and neighboring ones we find articles on elections in Italy and possible coalitions, next to two units (2/9) and (3/9) covering elections in India. Similarly, all other units on the map can be identified to represent a topical cluster of news articles. This co-location of similar, yet not identical topics, is one of the most important characteristics of SOMs making them particularly suitable for the organization of document collections for interactive browsing. The clusters are described by labels extracted by the LabelSOM method, reading for example, saigon, vietname, viet cong, province, buddha, monk, combat, helicopter, pilot, rocket, etc. for the Vietnam cluster, or christian, democrats, socialist, italy, pietro, nenni, fanfani etc. for the cluster on elections in Italy. For a more detailed discussion of the articles and topic clusters found on this map, we refer to (Rauber et al. 1999) and the on-line version of this map with the respective articles, available at [W03].

5. METAPHOR GRAPHICS FOR DOCUMENT VISUALIZATIONWhile the spatial organization of documents on the 2-dimensional map in combination with the automatically extracted concept labels supports orientation in and understanding of an unknown document repository, much information on the documents cannot be told from the resulting representation. Information like the size of the underlying document, its type, the date it was created, when it was accessed for the last time and how often it has been accessed at all, its language etc. is not provided in an intuitively interpretable way. Rather, users are required to read and abstract from textual descriptions, inferring the amount or recent-ness of information provided by a given document by comparing size and date information.We thus developed the libViewer, a metaphor-graphics based interface to a digital library (Rauber et al. 2000). Documents are no longer represented as textual listings, but as graphical objects of different representation types such as binders, papers, hardcover books, paperbacks etc, with further metadata information being conveyed by additional metaphors such as spine width, logos, well-thumbed spines, different degrees of dustiness, highlighting glares, position

PDF Cre

ate! 5 T

rial

www.n

uance.

com

61

in the shelf and others. Based on these metaphors we can define a set of mappings of metadata attributes to be visualized, allowing the easy understanding of documents.The libViewer is a Java-Applet interfacing with a number of servers providing the data to be visualized. Its main task is to provide a library representation that is intuitively understandable by the untrained user by relying on concepts and metaphors taken from conventional, real world libraries. It relies on a server to provide the appropriate mapping from the metadata attributes available in a specific document repository onto (a subset of) the supported metaphors. Thus it can be adopted to a variety of application domains, serving as an interface to library catalogs, on-line bookstores, and internal information repositories, and even manual pages or project documentations.

Figure 1: libViewer Interface

Figure 1 provides an example of the visualization of documents in a digital library using the libViewer interface. A number of different document types such as hardcover books, paperbacks, technical reports and papers can be easily identified as their corresponding physical representations, such as the libViewer and somViewer technical reports in green binders, the 4 different Langenscheidt dictionaries as yellow hardcover books or various paperback books published by e.g. Springer. Further attributes are depicted in a similar fashion, e.g. having the logo identify the publisher of a book if a corresponding logo is available (e.g. Springer, Langenscheidt, IEEE), or having the thickness of the binding represent the size of the underlying resource as e.g. for the different Langenscheidt Dictionaries. Another straight-forward representation is provided by the degree to which dust has accumulated on the back of the books, ranging from a few dust particles to a spider-web

PDF Cre

ate! 5 T

rial

www.n

uance.

com

62

covering half of a book that has not been referenced for a long time, as it is the case for the fifth book in the lower shelve. On the other hand, the third book in the lower shelve is clearly identified as being frequently referenced due to its rather distorted, well-thumbed binding indicating its frequent use. Albeit hardly noticeable in the printed representation, we find a highlighting glare in the first book in the upper shelve, indicating -- similar to shiny new books in libraries -- the fact that it was added to the library only recently. Furthermore, some books like the first ones in the upper shelve as well as most binders are not aligned with the backs of all the other books, making them stand out and thus promoting easier picking. Contrary to that, some books like the third in the upper shelve or the second in the lower shelve have been pushed far into the back of the shelves. The alignment can thus be used to indicate some kind of relevance recommendation or, with respect to electronic book stores, indicate promotions.

Figure 2: libViewer visualization of the TIME collection in (a) distant and (b) close-up view

An example for a libViewer library visualization of the TIME Magazine article collection is depicted in Figure 2, showing part of the Vietnam cluster of the flat SOM both in the distant as well as the close-up view. The documents are colored according to their cluster membership as determined by the shared labels, with for example, the Vietname documents colored yellow, and those of the neighboring African cluster colored green. Please note, that the relative age of a document, to pick just one example, become intuitively visible from the graphical representation, as does the relative size of a document, relieving the user from having to read and to interpret the metadata as textual descriptions.

PDF Cre

ate! 5 T

rial

www.n

uance.

com

63

6. REFERENCESDittenbach M., Merkl, D. and Rauber, A. (2000) The Growing Hierarchical Self-Organizing Map. Proceedings of the International Joint Conference on Neural Networks (IJCNN 2000), IEEE, 2000

Kohonen, T. (1982) Self-organized formation of topologically correct feature maps. Biological Cybernetics, 43, 1982

Kohonen, T. (1995) Self-organizing maps. Springer Verlag, Berlin, 1995.

Kohonen, T., Kaski, S., Lagus, K., Salojrvi, J., Honkela, J., Paatero, V., and Saarela, A. (2000) Self-Organization of a Massive Document Collection, IEEE Transactions on Neural Networks, vol. 11(3), IEEE, 2000.

Merkl, D. (1998) Text classification with self-organizing maps: Some lessons learned. Neurocomputing, vol. 21(1-3), Elsevir, 1998.

Rauber, A. and Merkl, D. (1999) Using Self-Organizing Maps to Organize Document Collections and to Characterize Subject Matters: How to Make a Map Tell the News of the World. Proceedings of the 10. International Conference on Database and Expert Systems Applications (DEXA99), LNCS 1677, Springer, 1999.

Rauber, A. (1999) LabelSOM: On the Labeling of Self-Organizing Maps. Proceedings of the International Joint Conference on Neural Networks (IJCNN99), IEEE, 1999.

Rauber, A., and Bina, H. (2000) Visualizing Electronic Document Repositories: Drawing Books and Papers in a Digital Library. Advances in Visual Database Systems: Proceedings of the IFIP TC2 WG2.6 5. Working Conference on Visual Database Systems, Kluwer Academic Publishers, 2000.

Salton, G. (1989) Automatic Text Processing: The Transformation, Analysis, and Retrieval of Information by Computer. Addison-Wesley, 1989.

7. LINKS[W01] http://www.amazon.com[W02] http://yahoo.com[W03] http://www.ifs.tuwien.ac.at/~andi/somlib/

PDF Cre

ate! 5 T

rial

www.n

uance.

com