Content and Knowledge Management in a Digital Library
and Museum
Jian-Hua Yeh, Jia-Yang Chang, Yen-Jen Oyang
Department of Computer Science and Information Engineering, National Taiwan University, Taipei, Taiwan, Republic of China. E-mail: yjoyang@csie.ntu.edu.tw
This article discusses the design of a digital library that addresses both content and knowledge management. The design of the digital library features two major dis-tinctions: (1) the system incorporates a two-tier reposi-tory system to facilitate content management, and (2) the system incorporates an object-oriented model to facilitate the management of temporal information and exploits information extraction and deductive inference to derive implied knowledge based on the content of the digital library. The two-tier repository system relieves the system manager from manually maintaining the hy-perlinks among the Web pages, when the digital library content is updated. The task of maintaining hyperlinks among Web pages can become cumbersome to the sys-tem manager if there are a large number of Web pages and hyperlinks. With respect to knowledge manage-ment, this design aims at facilitating temporal informa-tion management and deriving implied relainforma-tions among the objects in the digital library. The motivation behind developing these knowledge processing utilities is to create a system that complements the capabilities of human beings. Deriving a comprehensive list of implied relations is an exhausting task if the digital library con-tains a great amount of information and the number of implied relations is great. With such knowledge-pro-cessing utilities, specialists are released from perform-ing tedious work and can, therefore, spend more time with more productive philosophical activities to derive advanced knowledge. Applying knowledge management utilities effectively can extend the applications of digital libraries to new dimensions.
Introduction
Advances in computer network and storage technologies have inspired the design of digital libraries in recent years (Arms, Blanchi, & Overly, 1997; Gladney et al., 1994; Kahn & Wilensky, 1995; Nuernberg, Furuta, Legget, Marshall, & Shipman, 1995; Schatz & Chen, 1996). The emergence of digital libraries has introduced a number of important issues (Daniel & Lagoze, 1997; Glymour, Madigan, Pregibon, & Smyth, 1996; Imielinski & Mannila, 1996). One issue that
has not attracted much attention, but is essential to digital library development, is the management of digital library contents. Because a typical digital library contains a large quantity of documents, the conventional practice used to maintain the contents of web sites will not work for digital libraries. In particular, the hyperlink that is widely incorpo-rated in conventional web sites to structure the web pages poses a major challenge. For a digital library with a large quantity of materials, attempting to manually maintain the hyperlinks among web pages will become a cumbersome task for the manager of a digital library. Due to this obser-vation, the design of the National Taiwan University Digital Library and Museum (NTUDLM) incorporates a two-tier repository architecture that eliminates the need to manually maintain hyperlinks among the object Web pages.
Another interesting issue that emerged with the develop-ment of digital libraries is knowledge managedevelop-ment. In some cases, if the digital library can also provide knowledge management capabilities (Chen, 1990; Chen, Smith, Lars-gaard, Hill, & Ramsey, 1997; Wong et al., 1997), then the application of the digital library is extended to a new di-mension. The design of the NTUDLM aims at facilitating the management of temporal information and exploiting information extraction (Cowie & Lehnert, 1996; Solderland, 1996; Solderland, Fisher, & Lehnert, 1997) and deductive inference (Fagin, Halpern, Moses, & Vardi, 1995; Russel & Norvig, 1995) to derive implied knowledge based on the content of the digital library. One application of the incor-porated knowledge-processing utilities is to derive the im-plied relationships among the objects in the digital library from a large volume of information. The following is an example. Assume that one piece of information in the digital library reveals that a person named A served in a govern-ment agency for some time, and another piece of informa-tion reveals that a person named B also served in the same agency for some time. Should the periods during which A and B served in the agency overlap, then these two pieces of information together imply that A and B were colleagues for a period of time. Deducing a piece of this kind of informa-tion may be quite trivial to most people. However, if the
digital library contains a great amount of information and the implied relationships are extensive, completing this kind of task may exhaust the person who attempts to do it manually. With the incorporation of knowledge-manage-ment utilities, specialists are released from carrying out such tedious work, and can spend more time in the pursuit of philosophical activities to derive more advanced knowl-edge.
To facilitate the management of temporal information and derive implied knowledge, an object-oriented model was developed. In recent years, there have been several articles that address the representation of temporal informa-tion using object-oriented structures (Goralwalla, O¨ zsu, & Szafron, 1997; Goralwalla, Leontiev, O¨ zsu, & Szafron, 1997; Tong, Wu¨thrich, & Sankaran, 1995). However, the developments in those structures aimed at the recording and retrieval of temporal information in databases. Because the main issue addressed in this article is deriving implied knowledge based on temporal information, an object struc-ture was developed with this goal in mind. In the process for deriving implied knowledge, the temporal logic system de-veloped by James F. Allen was employed (Allen, 1991).
In the following discussion, the next section elaborates on the structure of the two-tier repository system and the content management mechanism employed at NTUDLM. Then the following section describes the object-oriented model developed to facilitate management of temporal knowledge. The Information Extraction Section addresses how information can be automatically extracted from the content of a digital library and the process for deriving implied knowledge based on temporal information. Appli-cation of the Knowledge Processing Utilities provides ex-amples of knowledge management utility applications. The System Implementation discusses system implementation
and evaluation. Finally, concluding remarks are presented in the Conclusion Section.
Structure of the Two-Tier Repository System Figure 1 shows the major components of the NTUDLM. There are three main modules: the agent, the objects repos-itory, and the content repository. The agent is the module that directly interacts with the user. Upon receiving a query from the user, the agent expands the query based on the thesaurus (Chen, Yim, Fye, & Schatz, 1994; Miller, 1997; Park & Choi, 1997; Schatz, Johnson, Cochrane, & Chen, 1996; Schutze & Pedersen, 1997) of the NTUDLM content and then sends the expanded query to the repository system. Figure 2 exemplifies the query expansion carried out by the agent. In this example, Zealandia, Tayouan, and Tayovan are ancient names of a place now called An-Ping. Because a user may not know all of An-Ping’s ancient names, by which An-Ping is referred to in ancient documents, carrying out a query expansion with the thesaurus is an effective way to improve the recall rate of the search result. The repository system will retrieve all related materials based on the ex-panded query and return the retrieved materials to the agent. The agent will then integrate the retrieved materials to generate a Web page.
One of the main distinctions in the NTUDLM system design is that a two-tier repository system is incorporated. The two-tier repository system comprises the objects repos-itory and the content reposrepos-itory. The main reason why the two-tier design is employed is to facilitate content manage-ment. The objects repository stores and manages the content the NTUDLM in an object-oriented form, and is responsible for resolving information retrieval requests on the fly. To the agent, the objects repository is an object-oriented data-base.
The content repository stores and manages the content of the NTUDLM in a relational database. The content reposi-tory does not respond to information retrieval requests on the fly, which is the objects repository’s responsibility. The main role that the content repository plays is to facilitate content management. When updates to the NTUDLM con-tent are to be carried out, the updates are first made to the content repository. Then, the content of the content
repos-FIG. 1. System architecture of the NTUDLM.
itory is compiled to generate a new object structure for the objects repository. During the generation process, crossref-erences among the descriptive data of the objects are estab-lished. This practice eliminates the need to manually main-tain hyperlinks or crossreferences among the descriptive Web pages of the objects.
Figure 3 depicts how the objects in the NTUDLM is stored in a relational database in the content repository. Each object in the NTUDLM has an entry in the objects table. The objects are classified based on the types of Web page templates that will be used to synthesize the Web pages of the objects. The associated files fields of each object specify the files that will be invoked in synthesizing the Web page of the object. The table shown in Figure 3 is maintained by the manager of the content repository. When a new object is to be added into the content repository, the manager needs to fill out the fields in the table.
What are missing in the structure shown in Figure 3 are the crossreferences among the Web pages contents of the objects. In the NTUDLM, the crossreferences are estab-lished using an annotation utility as shown in Figure 4. The annotation utility examines the text files whose contents are to be incorporated on Web pages and looks for instances of the primary key terms in the table shown in Figure 3. At each instance of the primary key terms, the annotation utility annotates a CGI (Common Gateway Interface) tag.
When a CGI tag on a Web page is clicked, a query con-taining the primary key of the tag is sent to the objects repository.
The HTML files with CGI annotations, which are gen-erated in the annotation process shown in Figure 4, are stored in the objects repository. When an object is invoked by the agent, the objects repository retrieves the correspond-ing Web page template accordcorrespond-ing to the object type and the files associated with the object to synthesize the Web page of the object. Figure 5 illustrates the process of synthesizing the Web page of an object on the fly. During the synthesis process, the text files used are those that have been anno-tated with CGI tags, not those plain text files stored in the content repository.
The main advantage that the practice described here offers is a very effective method for managing the contents of a digital library. When the manager of the digital library wants to add new objects or modify the descriptive texts for some objects, the manager only needs to make updates to the content repository. Then, the annotation utility will automatically establish all crossreferences among the de-scriptive texts for the objects. This eliminates the cumber-some task of manually maintaining the hyperlinks or cross-references among Web pages.
Organization of Temporal Knowledge in the NTUDLM
The discussion in the last section focuses on how the contents of the NTUDLM are structured to support brows-ing operations conducted by the users. This section will address the object-oriented model employed in the objects repository to facilitate the management of temporal knowl-edge. Because of the nature of the NTUDLM, which con-tains mainly historical materials, temporal knowledge is the most important type of knowledge in this application. One example of temporal knowledge is the relationship between two people. Two people may have been classmates when at a school, and subsequently, colleagues, during their career development. As a result, time qualifiers are needed to precisely describe the classmate and colleague relationships
FIG. 3. The structure of the content repository.
FIG. 4. Crossreferences established by the annotation utility.
between these two persons. In the past, there have been a lot of studies on the representation and reasoning for temporal knowledge (Allen, 1991; Galton, 1987; Goralwalla et al., 1997a; Vila, 1994). What is addressed in this article is an object-oriented model developed to facilitate the manage-ment of temporal knowledge.
In the object-oriented model of the NTUDLM objects repository, there are two categories of objects, namely the informative object, and the cognitive object. Informative objects are the building primitives of the NTUDLM. Typi-cal informative objects include paragraphs of text, pictures, maps, video programs, etc. Each informative object has its own well-defined contents, and the contents normally do not change over time. Informative objects collectively are not structured except that they may be classified according to their nature. However, classification of informative objects is simply for easy management, and has no structural im-plication.
Informative objects define information piece by piece. However, in human cognition, there is a higher layer of abstraction. In the NTUDLM, objects of the higher layer of abstraction are called cognitive objects. Typical cognitive objects include persons, events, places, etc. A cognitive object is a concept described by a collection of related informative objects, and a cognitive object itself does not have well-defined contents. Human’s interpretation of a cognitive object is more dynamic than that of an informative object. As mentioned earlier, each informative object has its own well-defined contents, which normally do not change over time. On the other hand, the meaning and characteris-tics of a cognitive object may change from time to time. When new information related to a cognitive object is found and added to the digital library, how the cognitive object stands in our interpretation may change.
A typical informative object or cognitive object is asso-ciated with one or more attributes and relationships. The attributes and relationships are the structures used to store temporal information in the NTUDLM. The attributes as-sociated with an object describe the object from various aspects. Each attribute of an object is described by a se-quence of state tuples. A state tuple contains a state and a period of time during which the state is valid. The state tuples in a sequence are sorted according their order in time domain. Some attributes are valid indefinitely. Such at-tributes are described by single-state sequences in which the states are valid indefinitely. While an attribute describes an object from one aspect, a relation associates one object with another object. A relation can exist between two informative objects, between two cognitive objects, or between one informative object and one cognitive object. Because two objects in the real world may be related in different aspects, more than one relationship can coexist between two objects. For example, two persons can be both relatives in their family life and colleagues in their career development. In such situations, it is better to have two relations to describe these two persons’ relationship from the two different as-pects. A relationship, just like an attribute, is described by a
sequence of state tuples. Figure 6(a) depicts the attribute and relation structures of a typical object. Figure 6(b) illustrates the structures among the objects in the NTUDLM.
Like the crossreferences among the Web pages, the at-tribute and relation structures of the objects are derived from compiling the content in the content repository. In the content repository, there are one attributes and one relations table that record all the attributes and relations, respectively. Information Extraction and the Process of
Deriving Implied Knowledge
The object structures presented in last section were de-veloped to facilitate the management of temporal edge. The next question is how the desired temporal knowl-edge is acquired. The NTUDLM relies upon two sources to acquire knowledge. The first source is from the metadata (Baldonado, Chang, Gravano, & Paepcke, 1997; Weibel, 1995) associated with the objects in the digital library. The second source is through information extraction. In the NTUDLM convention, each object is associated with a set of metadata. The metadata is created manually by contents specialists and contains knowledge that can not be easily derived by information extraction utilities. Figure 7(a) shows an example of metadata. Some metadata fields
ify the attributes of the object and some specify how this object is related to other objects. Figure 7(b) depicts how the knowledge presented in the metadata in Figure 7(a) is recorded in the object structure described in the last section. Figure 8 shows an example from a history book that is a good case to which to apply the information extraction utility. The example summarizes the backgrounds and ca-reers of the officers that served in the Taiwan county gov-ernment during the Chin Dynasty. It is apparent that the list exhibits certain rules. For example, in the description of each officer, the first time period always specifies when the officer took the job, and the second time period always specifies when the officer left the job. Therefore, it is quite straightforward to automatically extract information about the time intervals during which these officers served in their jobs.
As far as deriving implied knowledge is concerned, the temporal logic developed by James F. Allen (Allen, 1991) is employed. In the temporal logic, a time argument is added to the conventional predicate to form the temporal predicate of form P(a1, a2, . . . , ak; t), where a1, a2, . . . , ak are
nontemporal arguments, and t is the interval during which the predicate is of concern. Figure 9(a) illustrates various relations between two time intervals identified by James F. Allen, and Figure 9(b) shows the primitive operations be-tween two intervals.
The deductive inference utility first employed in the NTUDLM project is for deriving the relations among the objects in the digital library that the temporal information
stored in the attributes and relations of objects implies. The process of deriving implied relations starts with a set of m objects, denoted by {O1, O2, . . . , Om}, and a set of n
FIG. 7. An example of metadata from NTUDLM. FIG. 8. An example from a history book.
temporal inference rules, denoted by {R1, R2, . . . , Rn}.
These m objects are structured in accordance with the object structures presented in the last section, and each of these n temporal inference rules specifies how a new relationship can be derived from the attributes and existing relations recorded in the object structures. Following is an example of the temporal inference rule. The example derives a col-league relationship between two persons based on their career histories. Figure 10 shows the pseudocode of the deductive inference process.
Job共PersonA, CompanyX; T1兲ˆJob共PersonB, CompanyX; T2兲ˆ
overlaps共T1, T2兲 f Colleague
共PersonA, PersonB; intersection共T1, T2兲兲
Applications of the Knowledge-Processing Utilities
The discussion in this section first uses a simple example to illustrate the process of deriving implied relationships
from temporal information stored in a digital library. Then, a number of situations are presented to demonstrate how such knowledge processing utilities can be applied to extend the applications of digital libraries.
The example used is to figure out the bureaucratic rela-tionship among the government officers whose career records are present in Table 1. Figure 11 shows the bureau-cratic hierarchy of the county government of Taiwan during the Chin dynasty. In this example, the criterion is that if one person held a job position that is on the subtree branching from another person’s job position, and the periods during which they held these two job positions overlap, then the first person is said to be a subordinate of the second person. The inference rule is as follows.
Job共PersonA, Position1; T1兲ˆJob共PersonB, Position2; T2兲ˆ
Successor共Position1, Position2兲ˆoverlaps共T1, T2兲 f
Subordinate共PersonA, PersonB; intersection共T1, T2兲兲
Figure 12 shows the results derived from the deductive inference process.
FIG. 10. Pseudocode for deriving implied relations among objects.
TABLE 1. Officers and their positions.
Name Position
From (year/month)
To (year/month) Shen, Chao-pin County governor 1644/01 1644/01 Jiang, Hsiang County governor 1646/01 1648/01 Wang, Chao-sheng County governor 1648/01 1651/01 Wang, Shi-june County governor 1665/01 1665/01 Chao, Shing-ke Secretary-general 1644/01 1648/01 Chen, Jia Secretary-general 1649/01 1651/01 Chang, Yuan-ying Secretary-general 1651/01 1651/01 Wang, Li-chong Secretary-general 1665/01 1665/01 Chang, Yuan-chu Executive secretary 1644/11 1648/11 Kao, Lang Executive secretary 1649/01 1651/01 Li, Ting-quai Executive secretary 1665/01 1665/01 Chie, Weng-da Director of police
division
1644/01 1648/01 Chang, Wen-mo Director of police
division
1649/01 1652/01
The dates are in Lunar calendar.
FIG. 11. Bureaucratic hierarchy of Taiwan provincial government during the Chin dynasty.
There are several occasions in which extensive search of the implied relationships among objects in the digital library can be applied to complement human beings’ capabilities. One occasion is when a new object is to be added to the digital library. In such an occasion, the search can start with known facts about the new object’s relations with the ex-isting objects in the digital library. The implied relations identified by the extensive search provide a clear picture about how the new object is related to the existing objects in the digital library, and are essential for maintaining the integrity of the digital library contents. For example, if the new object is an ancient aboriginal leader, then the search can start with the village where the aboriginal leader re-sided, and the major events that the aboriginal leader were involved. An extensive search of the implied relations could identify the aboriginal leader’s relations with other impor-tant historical figures living in the same era.
Another occasion in which an extensive search of im-plied relations among objects will be invoked is when a content specialist finds a new relationship between two objects and wants to find out whether the new relationship affects our interpretation about existing objects or implies new relationships among existing objects. For example, a newly found piece of information reveals that a person named Y is actually a child of another person named X, and it is already known that a person named Z is a child of X, then a sibling relationship exists between Y and Z.
The extensive search mechanism described in this article can also be applied in informational retrieval. The search mechanism will yield high recall rates, and can rank the search outputs according to the level of association that an output has to the original search target. The power of this search mechanism is that it can find extended relationships among cognitive objects from the existing relationships in the digital library. As mentioned earlier, such extended associations often lead to new interpretations about the subjects of interest.
System Implementation and Evaluation
Figure 13 depicts the system architecture of the NTUDLM. The agent and the objects repository run on a SUN Enterprise 3000 server with four 200-MHz Ultrasparc CPUs, while the content repository is installed on a HP NetServer with a 233-MHz Pentium II CPU. The reason behind incorporating two different platforms is to exploit the advantages of these two different types of platforms. As of today, UNIX-based platforms in general offer higher system performance and higher availability than Windows-NT based platforms. Because the agent and the objects repository are the modules that directly respond to users’ queries, it is more desirable to install these two modules on a UNIX-based platform. On the other hand, software utili-ties for Windows-NT– based platforms are more widely available and at lower costs than those for UNIX-based platforms. As a result, it is more appropriate to install the content repository module on a Windows-NT– based plat-form. In the implementation, a Microsoft SQL server was employed to build the content repository.
As of today, the NTUDLM contains over 27,000 ob-jects, and the total content volume in the content repos-itory exceeds 400 Gigabytes. The uncompressed image files of the objects account for most of the volume. To expedite transmission of images over networks, the im-age files are resampled to reduce their resolution, and are compressed before they are loaded into the objects re-pository. With resampling and compression, the total volume of the image files in the objects repository is about 5 Gigabytes.
With the current volume of the NTUDLM, the com-pilation process that automatically establishes crossref-erences among the Web pages of the objects takes 1,583 seconds. This amount of time does not include the time taken by resampling and compressing image files. Among
FIG. 13. System implementation of the NTUDLM. FIG. 12. Results derived by the knowledge discovery process.
the 1,583 seconds, 1,026 seconds are used to retrieve data from the Microsoft SQL server database, and the remain-ing 557 seconds are used to build crossreferences among the Web pages of the objects. To evaluate the response time of the system, 10,000 query terms were randomly selected from the content repository primary key list. The system, on average, required 0.157 seconds to service a query.
Conclusion
This article discussed the design of the National Tai-wan University Digital Library and Museum (NTUDLM), which addresses content and knowledge management. The digital library design features two ma-jor distinctions: (1) the system incorporates a two-tier repository system to facilitate content management, and releases the system manager from manually maintaining hyperlinks among the web pages. (2) The system incor-porates an object-oriented model to facilitate the man-agement of temporal information, and exploits informa-tion extracinforma-tion and deductive inference to derive implied knowledge based on the content of the digital library. The incorporated knowledge-processing utilities not only re-leases specialists from performing simple but tedious inference work, but also extends the applications of dig-ital libraries into a new dimension.
This article also discussed the implementation and evaluation of the NTUDLM. The entire system was im-plemented on two platforms— one UNIX-based platform, and one Windows-NT– based platform. The main reason behind this practice is to exploit the advantages of two different types of platforms. As far as system perfor-mance is concerned, operational experiences revealed that the system can respond to a user’s query within a reasonable period of time.
What this article presents are just a few first steps taken by the NTUDLM research team toward developing a knowledge-based oriented digital library. This article fo-cused on deriving the implied relationships among the ob-jects in a digital library containing mostly historical articles and records. Similar ideas can be applied to derive other types of knowledge based on the contents of various types of digital libraries. One issue that is currently under inves-tigation by the NTUDLM research team is the representa-tion and reasoning of spatial knowledge. This study is essential to the goal of incorporating cultural atlas materials into the NTUDLM.
References
Allen, J.F. (1991). Temporal reasoning and planning. In Reasoning about plans (pp. 1– 68). San Mateo, CA: Morgan Kaufmann Publishers, Inc. Arms, W.Y., Blanchi, C., & Overly, E.A. (1997). An architecture for
information in digital libraries. D-Lib Magazine. [Online]. Available: http://www.dlib.org/dlib/february97/cnri/02arms1.html.
Baldonado, M., Chang, C.-C., Gravano, L., & Paepcke, A. (1997). Meta-data for digital libraries: Architecture and design rationale. In Proceed-ings of ACM digital libraries ’97. Philadelphia, PA.
Chen, H. (1990). A knowledge-based design for hypertext-based document retrieval system. In Proceedings of the international conference and workshop on database and expert systems applications (DEXA 1990). Vienna, Austria.
Chen, H., Smith, T.R., Larsgaard, M., Hill, L.L., & Ramsey, M. (1997). A geographic knowledge representation system for multimedia geospatial retrieval and analysis. International Journal on Digital Libraries, 1(2), 132–152.
Chen, H., Yim, T., Fye, D., & Schatz, B.R. (1994). Automatic thesaurus generation for an electronic community system. Journal of American Society for Information Science (JASIS), 46(1), 52–59.
Cowie, J., & Lehnert, W. (1996). Information extraction. Communications of the ACM, 39(1), 80 –91.
Daniel, R., Jr., & Lagoze, C. (1997). Distributed active relationships in the Warwick Framework. In Proceedings of the second IEEE metadata conference. Maryland.
Fagin, R., Halpern, J.Y., Moses, Y., & Vardi, M.Y. (1995). Reasoning about knowledge. Cambridge, MA: The MIT Press.
Galton, A. (1987). Temporal logics and their applications. New York: Academic Press.
Gladney, H.M., Fox, E.A., Ahmed, Z., Ashany, R., Belkin, N.J., Lesk, M., Tong, R., & Zemankova, M. (1994). Digital Library: Gross structure and requirements. In Proceedings of digital libraries ’94. College Station, TX. [Online]. Available: http://www.csdl.tamu.edu/DL94/paper/ fox.html.
Glymour, C., Madigan, D., Pregibon, D., & Smyth, P. (1996). Statistical inference and data mining. Special Issue of the Communications of the ACM on Data Mining and Knowledge Discovery, 39(11), 35– 41.
Goralwalla, I., Leontiev, Y., O¨ zsu, M., & Szafron, D. (1997a). Modeling temporal primitives: Back to basics. In Proceedings of 6th international conference on information and knowledge management (CIKM’97) (pp. 24 –31). Las Vegas, NV.
Goralwalla, I., O¨ zsu, M., & Szafron, D. (1997b). A framework for temporal data models: Exploiting object-oriented technology. In Proceedings of 1997 conference on technology of object-oriented languages and sys-tems (TOOLS USA 97). Santa Barbara, CA.
Imielinski, T., & Mannila, H. (1996). A database perspective on KDD. Special Issue of the Communications of the ACM on Data Mining and Knowledge Discovery, 39(11), 58 – 64.
Kahn, R., & Wilensky, R. (1995). A framework for distributed digital object services (Tech. Rep.). CNRI. [Online]. Available: http://WWW. CNRI.Reston.VA.US/home/cstr/arch/k-w.html.
Miller, U. (1997). Thesaurus construction: Problems and their roots. In-formation Processing & Management, 33, 481– 493.
Nuernberg, P.J., Furuta, R., Leggett, J.J., Marshall, C.C., & Shipman, F.M., III. (1995). Digital libraries: Issues and architectures. In Proceedings of digital libraries ’95. Austin, TX, USA. [Online]. Available: http://csdl. tamu.edu/DL95/papers/nuernberg/nuernberg.html.
Park, Y.C., & Choi, K.S. (1997). Automatic thesaurus construction using Baysian networks. Information Processing & Management, 32, 543– 553.
Russel, S.J., & Norvig, P. (1995). Artificial intelligence: A modern ap-proach (pp. 151–184). Englewood Cliffs, NJ: Prentice-Hall Interna-tional, Inc.
Schatz, B.R., & Chen, H. (1996). Building large-scale digital libraries. Computer theme issue on US Digital Library Initiative.
Schatz, B.R., Johnson, E.H., Cocharane, P.A., & Chen, H. (1996). Interactive term suggestion for users of digital libraries: Using sub-ject thesauri and co-occurrence lists for information retrieval. In Proceedings of ACM digital libraries ’96. Hyatt Regency Bethesda, MD.
Schutze, H., & Pedersen, J.O. (1997). A cooccurrence-based thesaurus and two applications to information retrieval. Information Processing & Management, 33, 307–318.
Solderland, S., Fisher, D., & Lehnert, W. (1997). Automatic learned vs. hand-crafted text analysis rules (Tech. Rep.). CIIR. [Online]. Available: http://cobar.cs.umass.edu/pubfiles/te-44.ps.
Solderland, S.G. (1996). CRYSTAL: Learning domain-specific text anal-ysis rules (Tech. Rep.). CIIR. [Online]. Available: http://cobar.cs. umass.edu/pubfiles/te-43.ps.
Tong, W.C., Wu¨thrich, B., & Sankaran, K. (1995). A temporal and probabilistic, deductive and object-oriented query language. In Workshop on Temporal reasoning in deductive and object-oriented database (pp. 61–69). Singapore.
Vila, L. (1994). A survey on temporal reasoning in artificial intelligence. AI Communications, 7(1), 4 –28.
Weibel, S. (1995). Metadata: The foundations of resource description. D-Lib Magazine. [Online]. Available: http://www.dlib.org/dlib/July95/ 07weibel.html.
Wong, S.T.C., Hoo, K.S., Knowlton, R.C., Hawkins, R.A., Laxer, K.D., & Tjandra, D. (1997). Issue and applications of digital library technology in biomedical imaging. International Journal on Digital Libraries, 1(3), 209 –219.
