Theory, Vol. 17, pp. 155-164, 1993 Printed in the USA. All rights reserved.
Library Acquisitions: Practice &
0364~6408/93$6.00 + .OO Copyright 0 1993Pergamon Press Ltd.
TECHNOLOGY FOR ACQUISITIONS AND ACCESS
AN EXPERT SYSTEM FOR PREDICTING APPROVAL PLAN RECEIPTS
Abstract - This paper explores the use of expert systems technology in one aspect of library acquisitions, the approval plan projile. The methodologies for determining whether a requested title would be received on approval were encoded in an expert system. The system design and test is described. Some of the advantages to using expert system technologies and the difficulties which were encountered are described. The future potential of expert systems for library acquisitions is addressed. A selected bibliography is included.
INTRODUCTION In recent years there have been a number of books and articles that have addressed the potential of expert systems for libraries. Most of that literature, however, focuses on the application of artificial intelligence to reference services -capturing the expertise of reference librarians in systems that assist library users in identifying appropriate reference tools [ 11. A few articles have addressed the potential of expert systems for cataloguing, and Pam Zager [2] applies expert systems technology in the acquisitions department by means of a system which assists with vendor selection. Yet most libraries are not currently using expert systems technology [3]. This is unfortunate, since it seems that the next logical step for library automation is to incorporate expert system technologies into our integrated library systems, particularly for acquisitions and cataloguing processes. The predictive serials check-in system, which monitors publication cycles of periodicals and identifies titles for claiming when issues are not received as expected is a start in this direction. The next step for these systems is for them to “learn” about publication patterns of individual titles and adjust claiming cycles, based on actual receipt dates, rather than a mathematically derived length of time between issues. This paper explores the use of expert systems technology in one aspect of library acquisitions, the approval plan profile. The methodologies for determining whether a requested title 155
156
L. C. BRANCHE
BROWN
would be received on approval were encoded in an expert system. The system design and test is described. Some of the advantages to using expert system technologies, and the difficulties which were encountered are described. The future potential of expert systems for library acquisitions is addressed [4]. Expert systems encode the knowledge of experts into “knowledge bases” that allow them to draw logical conclusions. The banking industry utilizes expert systems to decide whether a person qualifies for a mortgage or car loan. Insurance companies use expert-system technology to determine the insurability of customers. Using the information about a particular person, in conjunction with company policies, expert systems evaluate the criteria and make recommendations. In all aspects of library service, we also evaluate known criteria in terms of policy or procedure. Based on that evaluation, we choose a vendor, estimate a price, assign a shelving location, choose a reference source, value a gift, and so on. Rule-based expert systems technology incorporated into our automated library systems can apply library policy and procedures to known data elements and generate new data, especially in the acquisitions department. This has the potential to free staff from the repetitive keying of standard data into database fields and reduce training of new employees. No longer would they have to memorize publishers from whom we order direct or check lists of standing orders. At the Pennsylvania State University Libraries, we investigated the use of expert systems technology within the acquisitions department for predicting approval plan receipts. The receipt of books on an approval plan is determined by a set of rules, commonly known as the approval plan profile. These rules are applied by resident experts, the approval plan staff, to the order requests that are submitted by subject selectors. To apply expert system technology to this evaluation process, the methodologies used for determining whether a requested title would likely be received via a PSUL approval plan were encoded using rule-based expert system technology. A prototype expert system was developed using the VP-Expert expert system shell package by Paperback Software International, Berkeley, CA.
ASSESSMENT
OF THE PROCESS
Approval plans gather materials from around the world, without the work involved in ordering books title by title. They often require coordination with the firm order process at an institution. Pennsylvania State University Libraries has 16 active approval plans, acquiring materials worldwide in a variety of subject areas. The approval program is budgeted separately from the firm order and serial accounts. Purchase requests must be reviewed to determine the possibility of receipt via any one of these active approval plans. This review includes examining a list of approval publishers, the subject parameters of appropriate plan profiles, and vendor treatment of titles. Expertise is required by the approval plan staff to analyze information on the purchase request and apply that information to approval plan criteria. This application of criteria, or rules, to known data is a natural application for expert systems technology. Approval plan activities, full of various rules for the kinds of books the library desires, are particularly appropriate for expert system application. Libraries contract with vendors for approval plan services in order to assure that current imprints are received without the intensive labor of firm ordering. In addition, bibliographers are assured that publications from major publishers will be shipped to the library, for their review. The vendor determines which books to send based on a library “profile,” a description of publishers, subjects, and nonsubject “parameters” (such as cost, language, format, and type) required by the library. The
Expert Systems and the Approval Plan
157
vendor monitors current publishing output, examining each title published in terms of the subject and nonsubject parameters of its customers. The vendor ships to the approval customer those titles whose profiles match the profile of the library. Encoding decision criteria for likely receipt has multiple benefits. Rather than requiring the highly specialized knowledge of staff in the approval unit, a larger number of staff would have access to the information needed to make the determination, thereby enhancing their skills and freeing approval unit staff for other tasks. Additionally, using a system that consistently evaluates each purchase request would aid in the training of new staff members and new subject selectors on the intricacies of approval plan profiles. Further, the knowledge gained in writing a stand-alone expert system provides the opportunity to explore the technology and consider possibilities for incorporating rule-based programming into an automated acquisitions system. Using rule-based technology in an acquisitions system could make it possible for the system to prevent firm order of titles that will come on approval, or automatically assign a vendor to a purchase based on the type of publication. Expert system technology, incorporated in an automated acquisitions system, could track agreements with vendors, assigning vendors to assure proper order mix and agreed-upon expenditures. Incorporated into other parts of an integrated library system, rule-based programming could assign shelving locations, automatically determine binding cycles or other preservation requirements, and assist staff with cooperative purchasing or lending agreements. At Penn State, the approval plans unit is responsible for monitoring book shipments and title treatment of its approval plan vendors. Order requests for current imprints are screened at the preorder search stage against a printed list of approval publishers. Those titles published by an approval publisher are sent to the approval plan unit for further research. The approval unit screens the order request to determine vendor treatment of the title and whether it will be shipped as part of the approval plan. During this screening, staff must determine the country of origin of the title, the imprint year, whether the subject is included in the plan that services that country; verify that the cost of the item is under the price limit; and examine the nonsubject parameters assigned to the title by the vendor. This screening process was targeted for expert system application.
SYSTEM DEVELOPMENT Through an analysis of the steps in the process, a system was designed to help make the determination as to whether a title would come on approval. Figure 1 illustrates the decision making criteria which were incorporated into the program. As order requests are received in the ordering unit, a judgment is made as to whether the title is likely to come via one of the active approval plans. Staff use a variety of “knowledge bases” or sources of information when evaluating each title. These include a list of publishers whose titles are received on approval, a spreadsheet that identifies subject inclusions organized by LC class for each of the plans in the approval program, a list of contemporary authors whose works are received on approval, a list of geographic coverage of the approval plan program, and accessing the vendor database to determine vendor treatment of each title (including nonsubject parameters such as academic level and publication type). One of the objectives in the system design was to incorporate these various lists, or “knowledge bases” into the expert system. This would provide broader accessibility to these various sources of information, thus allowing a greater number of people to establish likelihood of receipt via an approval plan. Not only would it provide information to a wider audience of acquisitions personnel, it would also be possible to pro-
158
L. C. BRANCHE
BROWN
Figure 1. Workflow for Determining If an Item Should Come on Approval.
Expert Systems and the Approval Plan
159
vide access to collection development subject specialists. Information on likely receipt would benefit subject specialists as they decide whether to submit a purchase request for a title. It also provides the staff with a single resource for evaluating titles and predicting receipt. Rather than consulting five different resources, they can simply respond to the questions posed by the system, and allow the system to consult the various databases of information. An expert system shell, VP-Expert, was used to encode approval profile information. This program provides a framework for developing the expert system, and allowed us to use existing databases as part of the system’s knowledge base. Initially, an action statement is assigned, which sets the goal for the decision making process. FoIlowing the action statement, a set of rules is described, by which the system draws conclusions and reaches the set goal. In this system, the goal is to determine if the title will come on approval. This is encoded in the action statement, “FIND come_on_approval.” The rules are then described, using the expert knowledge of the approval staff and the knowledge bases contained in the lists of publishers and subject inclusions and nonsubject parameters of the various plans. Rules consist of a statement, “IF,” and a conclusion, “THEN.” For example, we have limited the approval plan to those titles with prices per volume of less than $200.00. The rule that encodes this knowledge reads: IF price per volume > $200.00 THEN come_on_approval = no In addition to evaluating data using the ruIe base, VP-Exprt can access dBase databases and import information from these for use in the evaluation. Using this feature, it is possible to identify the LC class of a title and allow the system to retrieve information about that LC class from the separate database (such as retrieving the data on plans for which that subject is included in the approval program), and then incorporate this additional information into the evaluation process. The criteria identified in Figure I were incorporated into the knowledge base, using a set of rules and information from databases. The success of the expert system design rests in the logic of its rule base. In order to maintain the integrity of the logic that drives the analysis, the system was developed incrementally. Basic criteria, such as price and country of origin were encoded in the rule base first. Following that, publisher parameters, then subject parameters were added as part of the knowledge base. Additional criteria to be added to the knowledge base include a database of contemporary authors and composers whose works will be sent on approval by our vendors, and a mechanism for determining the academic level of the titles requested. Incremental development allows testing during each phase of development. Using the system does not require special training or knowledge on the part of the staff. The system begins by describing the scope and querying the user for information. (See Figures 2 and 3.) The user, working with an order request, answers each question posed for a particular title. Based on user response, the system works through the rule base, accessing the databases and incorporating information from them into the knowledge base as necessary. It reaches a conclusion about the title by using the rule base and the action statement to evaluate known criteria and infer logical conclusions. Unknown information is solicited from the user. (See Figure 4 for a sample session.) System users do not need sophisticated skills to make use of the technology. VP-Expert provides a user-friendly enviro~ent, with natural language questions and the option to provide menu choices for a response. (See the question about place of publication in Figure 3.) The system simply requires that the user have basic information about the title with which to answer the questions posed.
160 This
L. C. BRANCHE expert
systeLndcumMcs
BROWN
whether
aparticularitemislikelytocomeviaa PSU approval plan. Press any key to begin the consultation.
I
IHelp 2G
3whtif
VP-Expert Menu Options.
TESTING
THE SYSTEM
To test the decision-making skill of the system, a small group of titles requested for purchase were run against the system. These titles had already been evaluated by the approval plan staff. Ten were determined to be out of scope and therefore not expected on approval. Ten
%%o 1s the Pubhsher? Which LC class has this title been given? Does the item cost less than $200.00? No YeS What year was the item published? Where was the item originally published? Africa AWlia Germany New Zealand SwiQerland United States Other
Belgium Holland United Kingdom
Will this item come on approval? SRule 6Set 4Vaxiable 1H 1 2G 3whfif IHz$ 2Htw 3whay? 4Slow 5Fast 6Quit
7Edrt 8Qurt
Figure 3. System Queries.
161
Expert Systems and the Approval Plan
t&o IS the Pubhshti OMord University Pr’ess Which LC class has this title been given? t9 Does the item cost less than $200.00? YCS No Wii this item come on approval? NO
1Help 2Go 3Whatif 4Variable 5Rule 6Set 1Help 2How 3Why? 4Slow SFast 6Quit
7Edit
@it
Figure 4. Sample Session. User Response Is Italicized. System Response Follows the Final Question (in bold).
were determined to be within the criteria for approval plan shipment and therefore could be expected to come on approval. Because the outcome was already known, the system could be evaluated for flaws if the results of the expert system test did not match the staff’s decision about the title. A more rigorous test is planned, which will include a larger number of titles tested prior to evaluation by the approval plans staff. Most of the titles were correctly identified by the system. In the cases where the system did not agree with the staff decision, errors in the logic of the rule base were found. When these were corrected, the system tested accurately. For example, an error in logic occurred because the system tests the rules consecutively. If a conclusion is reached within a rule, it does not evaluate subsequent rules. Changing the order of the rules corrected the error. Another problem that was encountered was caused by the use of the “ELSE” clause in a rule. “ELSE” within a rule is extremely powerful in the inference engine because it forces a conclusion within that rule. For example, the rule base contains a rule that evaluates the LC class of a title in terms of its inclusion or exclusion for each approval plans. Originally this rule was expressed as: IF BTA = EX or CAM = EX or COU = EX THEN come_on_approval ELSE come_on_approval
= no = yes
In this rule, the system is seeking a value for “come_on_approval,” based on what it knows about subject inclusions for each vendor. It looks to the database to determine whether the subject is included (“IN”) or excluded (“EX”) for each of the three vendors (BTA, CAM, and COU). If it finds that the subject is excluded (“EX”) for any of these vendors, the value of “come_on_approval” would be “No.” But, if the subject is included for each of the vendors, the “ELSE” clause triggered the conclusion, “come_on_approval = yes.” This prevented the system from looking further in the rule base for parameters that may have excluded
162
L. C. BRANCHE
BROWN
the title from the approval plan profile (such as price or year of publication). When the “ELSE” clause is removed, the system proceeds to the next rule if none of the criteria in this rule are met, as it should. During system testing another dif~culty was encountered. There are frequently instances in evaluating titles that are not easily encoded as either yes or no, such as subject areas where our profile indicates “limited coverage. ” “Limited coverage” is used to indicate an LC class where we receive books in some areas, but not in others. The knowledge base we used to assign LC class inclusion in profiles contained only the first two alpha characters of an LC class number, which meant that a rule could not be written to include RM 214-258 but exclude RM 283-298. This limitation resulted in erroneous results and would have to be addressed before the system could be put into regular use. Adding a more detailed list of LC classes to the knowledge base, with their inclusion or exclusion from the profile, would overcome this difficulty. A third difficulty involved the way the system accessed information from the database. To develop the publisher database, our existing publisher list was imported into dBase, each publisher as a separate record. This meant that in order to retrieve a record, the publisher’s name had to be entered, during a query session, exactly as it appears in the database record, e.g., John Wiley & Sons. If a staff person entered just “Wiley,” the system would not find the publisher name in the database, and would reach an erroneous conclusion that the publisher entered was not part of the approval program. The program was modified to query the user a second time if the publisher was not found in the database. Users are instructed to look up the publisher on the printed list and reenter the publisher’s name exactly as it appears on the list. This is a time-consuming task. As users become familiar with the system, they will be more familiar with the “correct” way they enter the name. This has implications for new staff and staff training, however, and is indicative of the problem with automating the human thought process. As approval plan staff review order requests, they easily identify publishers regardless of how a requestor has listed the publisher’s name. The expert system, being a literal machine, cannot identify “Wiley” as “John Wiley & Sons.” One method of overcoming this would be to add every possible variation of a publisher’s name to the database. Requiring staff to learn the “correct” way of entering a publisher’s name is a limitation of the system, since one of the objectives of developing the system was to provide a means for disseminating approval information that requires minimal training or expertise. This problem also demonstrated the need for database m~ntenance. As publishers are added or removed from the approval plan, it is necessary to modify the databases of the expert system. Other profile modifications would require system maintenance. While streamlining the workflow for some staff members, the staff person responsible for the system would need to perform regular updates. This maintenance could also utilize expert systems technology. A secondary expert system was created to modify the databases. Similar to the approval plan analysis, this system queries the user for the change, and adds or deletes data in the dBase databases.
CONCLUSION The development of the prototype system look at how we evaluate titles in terms of our nology into the daily approval staff routine include criteria such as the author list and to LC classifications.
afforded us the oppo~u~ty to take a detailed approval plan coverage. Incorporating this techwill require continued system development, to further refine database access for more detailed
Expert Systems and the Approval Plan
163
The system has the potential for providing a wide range of PSU library faculty and staff with detailed information about approval plan receipts. Additionally, there is potential for incorporating similar technology into the local acquisitions system that is under development. Rules could be incorporated into the system, so that data is evaluated as a new record is added. Price, publisher, country of publication, and publication date could all be examined by the system and potential approval plan receipts “flagged,” so that they are not firm ordered. This would eliminate the need for the list checking and prescreening that is currently done. Rulebased programming also holds potential for system-generated vendor selection and determining appropriate price estimates for items in the acquisitions system. Expert systems are capable of both storing and applying knowledge. A system that incorporates expert system technology can take the data from a database and apply it within the context of a set of parameters. Within the context of library acquisitions, this means that the acquisitions components of integrated library systems could apply “rules” of ordering and information from a database of potential vendors, and apply that to each individual order that is processed by the acquisitions unit. Using a database of approval plan parameters, an expert system can apply what it “knows” about the library’s approval plan to predict which requested titles are likely to be received on approval, reducing the work done to an-order request for a book that ultimately is received on approval. Stand-alone expert systems, such as the one described in this paper, simply streamline existing acquisitions workflows. As expert systems technology is incorporated into existing systems, libraries can expect more sophistication in the integrated systems. Such library systems will have the ability to infer outputs based on criteria which are encoded within the system. Further, it is possible that systems could “learn” from these inputs, where the rules that govern inference are written by the system itself, based on examples, rather than by programmers. As researchers of artificial intelligence have found, this potential comes not without difficulties. It is difficult for humans to describe how they make decisions. Encoding-the process by which decisions are made requires a preciseness that is difficult to achieve. Yet, the potential of the technology to streamline decision making and provide broader access to expert solutions is enticing enough for its continued pursuit. NOTES 1. See Ardis, Susan B., “Online Patent Searching: Guided by an Expert System,” Online, 14 (March 1990), 56-62; Morris, Anne, The Application of Expert Systems in Libraries and Information Centres, London: Bowker-Saur, 1992; Zahir, Sajjad and Chew Lik Chang, “Online-Bxpert:An Expert System for Online Database Selection,” Journal of the American Society for Information Science, 43 (1992), 340-357; and Bailey, Charles W., “Intelligent Library Systems: Artificial Intelligence Technology and Library Automation Systems,” Advances in Library Automation and Networking, vol. 4, ed. Joe A. Hewitt. Greenwich, CT: JAI Press, Inc., pp. l-23. 2. Zager, Pam and Omar Smadi, “A Knowledge-based Expert Systems Application in Library Acquisitions: Monographs,” Library Acquisitions: Practice & Theory, 16 (1992), 145. 3. See Expert Systems in ARL Libraries, SPEC Kit #174, Washington, DC: Office of Management Services, Association of Research Libraries, 1991. 4. Library Hi Tech published a special issue, l&l/2 (June 1992), describing expert system technology and its potential for libraries.
SELECTED
BIBLIOGRAPHY
Ardis, Susan B. “Online Patent Searching: Guided by an Expert System.” Online 14(2):56-62; March 1990. Bailey, Charles W. “Intelligent Library Systems: Artificial Intelligence Technology and Library Automation Systems.” In Advances in Library Automation and Networking, vol. 4. Greenwich, CT:JAI Press; 1991.
164
L. C. BRANCHE BROWN
Carrington, Bessie M. “Expert Systems: Power to the Experts.” Do&base 13:47-S& 1990. Expert Systems in ARL Libraries, SPEC Kit #174. Washington, DC: Office of Management Services, Association of Research Libraries; 1991. Holthoff, Tim. “Expert Librarian Applications of Expert Systems to Library Technical Services.” Technical Services Quarter/y 7:1-16; 1989. Morris, Anne. The Application of Expert Systems in Libraries and Information Centres. London: Bowker-Saur; 1992. Riggs, Donald E. “The Library Perspective.” In The Evolution of Library Automation: Management Issues and Future Perspectives, ed. Gary M. Pitkin. Westport, CT: Meckler; 1991: pp. 19-38. Riggs, Donald E. “Artificial Intelligence and Expert Systems.” In Information Technologv: Design und Applications. Ed. Nancy D. Lane and Margaret E. Chisholm. Boston, MA: G.K. Hall & Co; 1991:227-243. Zager, Pam, and Omar Smadi. “A Knowledge-based Expert Systems Application in Library Acquisitions: Monographs.” Library Acquisitions: Practice & Theory 16:145; 1992. Zahir, Sajjad and Chew Lik Chang. “Online-Expert: An Expert System for Online Database Selection.” Journal of the American Society for Information Science 43:340-357; 1992.