College and Research Libraries Improving OPAC Subject Access: The ADFA Experiment Alex Byrne and Mary Micco The Australian Defence Force Academy (AD FA) library has embarked on an ambitious project to upgrade subject access in its online public access catalog (OPAC) by adding an average of twenty-one multiword terms from the table of contents and/or index to the 653 field in the MARC record for each book. After reviewing six months of work involving some 6, 000 books, this study attempts a preliminary evaluation of the impact on the workload and resources, as well as on subject access. he Australian Defence Force Academy (ADFA), a campus of the University of New South Wales, has embarked on an am- bitious project to improve subject access to their book collection, which supports teaching and research in a full range of ac- ademic subjects. A relatively small aca- demic library of 160,000 volumes, it re- cently completed a total retrospective conversion and came online with the inte- grated Urica Library System. Urica is sup- plied by A WAin Australia and McDonnell Douglas in Europe and North America as a turnkey package of hardware and soft- ware. It is able to store and access all the fields available in the MARC format and is designed to permit subject access by con- trolled headings and keywords. OPAC PROBLEMS User surveys of existing OPACs have in- dicated that the major problems in data- base access were 1. finding the right subject heading; 2. ambiguous codes and abbreviations; 3. currency and coverage; 4. indexing and database accuracy; 5. updating. 1 The first problem also appears to be the major weakness of the Urica system. A keyword capability attempts to overcome the difficulty by increasing searching flexi- bility, but it has not proved as successful as had been hoped, for retrieval is handi- capped by both data and software limita- tions. In common with most other library catalog systems used in Australia and North America, Urica restricts subject searching to the MARC 650 fields and hence to Library of Congress Subject Headings (LCSH) recorded in those fields. Therefore the keyword capability can use only the words taken from those headings. That is, subject searching is possible only by II direct hit'' on the exact LC heading including subdivisions or by combinations of terms taken from that ar- tificial vocabulary. As programmed, the system first Alex Byrne is Deputy Librarian, Australian Defence Force Academy Library, Canberra, ACT, Australia. Mary Micco is Associate Professor of Computer Science, Indiana University of Pennsylvania, Indiana, Pennsylvania 15705. 432 ful information, to integrate it into the text, and to cite sources clearly. For their section of the course, the bibliographic in- struction librarians reinforce their expla- nations with carefully labeled handouts, and short hands-on exercises that relate to the major research project from the ESL class. To facilitate interaction further, the bibliographic instruction librarians have, with student help, compiled glossaries of library terms in the dominant languages of our international student population. In conclusion, awareness of how ESL students differ from other students would ESL Students 431 help librarians modify their approaches·, leading to improved communication with this growing segment of the university population. A more comprehensive intro- duction to the library than is presently em- ployed is a necessity for such students. An integrated program involving biblio- graphic instruction personnel with inten- sive English, English department, or "Writing across the Curriculum" pro- grams can be an ideal way to help ESL stu- dents acquire the knowledge and skills es- sential for college success. REFERENCES AND NOTES 1. Office of Educational Research and Improvement, Center for Statistics, Digest of Education Statis- tics, 1985-86 (Washington, D.C.: U.S. Dept. of Education, 1986), p.218-19. 2. Kwasi Sarkodie-Mensah, "In the Words of a Foreigner," Research Strategies 4:30-31 (Winter 1986). 3. Mary Alice Ball and Molly Mahony, "Foreign Students, Libraries and Culture," College & Research Libraries 48:160-66 (Mar. 1987); Manuel D. Lopez, "Chinese Spoken Here: Foreign Language Li- brary Orientation Tours," College & Research Libraries News 44:265 (Sept. 1983); Terry Ann Mood, "Foreign Students and the Academic Library," RQ 22:176-80 (Winter 1982); and Sally G. Way- man, "The International Student in the Academic Library," Journal of Academic Librarianship 9:336-41 Oan. 1984). 4. One Lebanese student, for example, was accused of cheating on his history exam because he had duplicated from memory, word for word, comma for comma, five pages of his textbook. However, he was vindicated when the teacher challenged him to repeat the performance and he did so with ease. Student and teacher were obviously working under different sets of assumptions. The stu- dent declared memorizing at length was easier for him than trying to rephrase new ideas and in- formation into language that might not be as clear. 5. Robert L. Kohls, ed., "Resource 8," Developing Intercultural Awareness (Washington, D.C.: Society for Intercultural Education, Training, and Research, 1981), p.18-22. 6. Sarkodie-Mensah, "In the Words," p.31. 7. Wendy Moorhead, "Ignorance was Our Excuse," College & Research Libraries News 47:585-87 (Oct. 1986). 8. This practice can lead to disastrous misunderstandings, as when French Quarter customers, ask- ing a Cambodian waitress if they could use the rest rooms in a fast food restaurant were told, ''yes, yes," but none existed. 9. Louise Greenfield, Susan Johnston, and Karen Williams, ''Educating the World: Training Library Staff to Communicate Effectively with International Students," Journal of Academic Librarianship 12:227-31 (Sept. 1986). 10. One might occasionally find students capable of the most amazing esoterica; for example, the Jap- anese western enthusiast who cries out, "Come back, Shane!" or the Arab lawyer who can recite the issues involved in the Goetz case. searches for a direct exact subject heading match using a simple hashing algorithm. A keyword search is then initiated auto- matically if no direct match is found or at the user's request. The keywords are com- bined by an implied AND. If too many hits are retrieved, additional terms or trun- cated terms may be AND' ed to narrow re- trieval. However, successful retrieval re- quires all terms to be in the same 650 field; searching across headings is not possible at this juncture. If many hits are found, an AND strategy can be used to combine ad- ditional terms (as character strings) from other 650 fields or most other data fields. Free-text searching is not available. .~~Several methods were considered before the subject access approach method advocated by Pauline Ather- ton was chosen.'' PRELIMINARY ANALYSIS Frustration with these limitations led the ADFA librarian, Lynn Hard, and his staff to consider means of improving the subject data and the software capabilities. As it was decided to remain compatible with the standard Urica system for the present, software changes were mini- mized and efforts were concentrated on augmenting the subject data. Several methods were considered before the sub- ject access approach method advocated by Pauline Atherton was chosen. 2 This method consists of selecting supplemen- tary terms from the contents pages of books and, where needed, the index. ADFA changes were accomplished by se- lecting an average of twenty to twenty- five keywords from the chapter-level headings to be keyed into the 653 field for each book. There was some concern that not all books included the needed infor- mation but this proved to be a minor diffi- culty. The background study performed on a sample of 179 items showed that 104 items had both contents page and index, 52 had contents only, 3 had index only, and 20 had neither. Of the latter, twelve were fiction, four bibliographies, three in- Improving OPAC Subject Access 433 dexes, and one a dictionary. Before embarking on the actual process- ing, staff faced three questions: Would the augmented headings be used separately or in combination with LC headings? A small trial with twenty-two queries confirmed that the augmented terms would be most useful if available for searching as an alternative to the LC head- ·ings. The LC terms would be regarded as the main subject of the book and the addi- tional terms as "contents" terms. Care was taken to reduce redundancy by enter- ing each term only once per document. How would the augmented headings be structured? The best solution appeared to be the . entry of all headings in a single field, the 653 field for local subjects. The semantic relationships between the terms were preserved. Separate headings were delimited with ''@, '' with subheadings separated by means of a semicolon, punc- tuation that would not distort the retrieval process. Would any system of weighting be em- ployed? It was considered but rejected as possibly increasing the complexity of in- dexing by requiring value judgments. It was also felt that no reliable method was possible within the constraints of the cur- rent software. 3 The inquiry software, particular ENQ, which is used by the public, needed two changes to the display options. The first was labeled "enhanced full" and in- cluded both LCSH and augmented head- ings. The other was an "enhanced MARC" option. After six months, more than 6,000 books have been processed with enhanced sub- ject access terms. This paper measures and describes the impact. ENHANCED SUBJECT PROJECT Although exact titles and title words could be searched in standard Urica, it was apparent that loading that data into a subject field would enhance subject ac- cess.4 Consequently, the first step in the project was to load every record's title into the 653 field with a $t tag to indicate it was a title. This was accomplished in a batch run using software developed in house. Drawing on the experiment in providing 434 College & Research Libraries enhanced subject access at Syracuse Uni~ versity by Pauline Cochrane and Barbara Settel, the next setup, begun in December 1986, was to add 653 fields containing ad- ditional terms chosen from the book's contents page or index, to the MARC rec- ord. 5 The project was entitled the Enhance Subject Project (ESP}, and the staff began referring to the additional terms selected as ESP terms. At the time of writing, after seven months of adding enhanced subject terms to all new acquisitions, the library has a collection of 6,139 books, more than 4 percent of the database, that have re- ceived this treatment. Library patrons have full keyword access to these terms, which are stored in the 653 fields with spe- cial tags to indicate up to four levels of sub- division. Standard Urica searching soft- ware is used at present but the library's systems staff developed input and main- tenance software for the 653 field. There were several reasons for selecting this as the best and most cost-effective method of enhancing subject access. It can be accomplished easily in a few minutes. It adds many access points and increases the likelihood of retrieving an item. Users have commented that it helps them to judge the relevance of the retrieved items. It uses the terminology of the author, en- suring currency and relevance, and also provides access to the chapter-level terms at acceptable costs. Minimal skills and in- tellectual effort are required, and the time for cataloging is approximately fifteen minutes per item, yielding an average number of 20.7 headings or 53.6 terms per book. 6 IMPLEMENTATION DETAILS A manual outlining the selection proce- dures to be used was drafted, based on Barbara Sette!' s work with a few minor ad- justments. The first step in the process of adding terms is to determine whether the table of contents is usable and, if not, to se- lect suitable terms from the index. Guide- lines for difficult cases as well as many ex- amples are provided. The catalogers work from the photocopies of the title page, ta- ble of contents, and index. They highlight the terms that they select and add the ap- propriate level codes. These sheets are September 1988 then sent on for keying into the system. Terms are selected that describe the con- tents of each substantial chapter, except where they are significantly repetitious or introductory. The table of contents must contain at least one heading for every thirty-five pages or the index is used to provide additional terms. Page number ranges and names of authors of contrib- uted articles (except for poems and short stories) are omitted. IMPACT OF PROJECT Numerous studies on subject access have shown that LCSH performs poorly. Since it provides only 1.7 access points per book/ this is hardly surprising. It was as- sumed that providing table of contents in- formation would significantly increase re- call and improve precision of searches without adding unduly to the cost of pro- cessing and storing the information. The entries would provide naturally occurring hierarchies of terms that could be dis- played, preserving the syntactical struc- ture of the contents or index pages, to as- sist users in refining their search strategies and evaluating the relevance of retrieved citations. The major question is whether the terms added are useful or simply meaningless clutter. In this study, we have attempted to view the impact of the Enhanced Subject Project (ESP) in both its quantitative and qualitative aspects. IMPACT ON WORKLOAD AND RESOURCES The first consideration was the impact of this additional load on personnel and the system. The cataloging librarian reports that it takes the equivalent of one full-time entry-level clerical person to photocopy the contents pages and key in the selected data. Most selecting and coding of terms for the 1,000 new items a month are per- formed by one paraprofessional. The cata- loging section has absorbed the additional work with little difficulty. Relieved of former duties by the replacement of a to- tally manual system with a fully auto- mated one, the section has taken on this activity as well as continuous upgrading of descriptive cataloging data while losing only one clerk and experiencing a 25 per- cent increase in acquisitions. The systems manager reports that addi- tional disc storage requirements have not been as great as anticipated. Before copy- ing the titles so that they could be searched together with LCSH and ESP terms, the bibliographic database occu- pied 244 megabytes (MB) of storage (of a total of 350MB in use). Copying the titles into the 653 fields caused an increase of 14 percent, to 279MB. By August, after eight months of ESP, during which a total of 8,500 catalog records including 7,000 with ESP were added, there were minimal ef- fects on all files except the file containing the subject/title/ESP keywords. That file increased from 1.8MB to 5.5 MB when ti- tles were added, and then to 7.2 MB over the eight-month period. The increase at- tributable to ESP alone was approximately 2.5 MB. The number of keywords in- creased from 30,153 to 71,338 with the ad- dition of the titles, and then to 101,559 (see Improving OP AC Subject Access 435 table 1). Some 30,000 words in the past eight months are undoubtedly due to ESP. While steep at first, this rate of growth is now declining. By separating the 6,139 titles that had been ESP' ed into a distinct database we were able to evaluate the growth of the vo- cabulary. During this time a total of 51,175 new words were added to the system. It was initially expected that growth would be very rapid until it reached 50,000 and . then would level off, topping out at around 100,000. A sufficiently large data- base is not yet available to verify this opin- ion but it appears unwarranted and will be reevaluated after twelve months. Proper names and jargon are included in the vo- cabulary but there are few non-English- language terms. The distribution of the vocabulary was interesting in that a very large number of words, 89 percent, had fewer than ten occurrences, and 58 per- cent had only one hit (see figure 1). TABLE 1 IMP ACT OF ESP ON VOCABULARY GROWTH NewA~dedords Records Added Ratio January 4,977 448 11.0:1 February 4,247 648 6.6:1 March 4,184 698 5.9:1 ~ril 3,129 707 4.1:1 ay 6,676 802 8.3:1* June 5,717 1,452 3.9:1 *Additional specialist military history materials E.S.P.'ed during these two months as part of a special project (Australasian Military History Database) temporarily interrupted the steady decline. en .c u ~ "'0 c ::J .c .s 50 22 20 18 16 14 12 10 8 6 4 2 70 1-10 11-20 21-30 31-40 41 -50 51-60 61-70 71-80 81-90 91-100 FIGURE 1 Distribution of Vocabulary-word frequency counts 436 College & Research Libraries It is not yet clear whether this pattern is a result of having such a small database and will hold as the database gets larger. It certainly has implications for searching. By putting in the lower frequency terms first, the user may avoid searching for high frequency words at all or may do so only after first culling a small pool of rele- vant documents. It will, however, be very important to provide an "OR" capability to ensure that a sufficiently large pool of documents is retrieved. The vocabulary lists demonstrate again that the number of individual keywords generated from the LC headings is very re- stricted. We need a mechanism to map the natural-language terms put in by users to this very controlled vocabulary. IMPACT ON INFORMATION RETRIEVAL Another interesting question was the impact of the additional access points on information retrieval. In a quantitative analysis we selected 100 significant key- words in computer science identified by the faculty, and an additional 100 key- words in a stratified random sampling of the entire database. (Three or four key- words were randomly chosen from each letter of the alphabet, excluding the words with only one hit and those previously se- lected for the computer science segment.) The focus was on the extent to which the ESP terms are increasing recall in the sys- tem as opposed to adding clutter. Each keyword was submitted for searching and the records retrieved were examined. If the word appeared in the LC subject heading, it was counted as LC and eliminated from the pool. Then the title was checked to see what additional rec- ords were being found by adding key- words from the title. In this study we did not measure the degree of overlap be- tween the terms in the LC headings and ·the title, since our purpose was to evaluate the material being contributed by the chapter-level terms. Finally, if the key- word had not been found in either of these places, it was counted as a record that was being recovered by the chapter-level headings. All of this was made possible by the use of tags to differentiate the ele- September 1988 ments being stored in the 650 and 653 fields and by the programming skills of the systems analysts. See figure 2. The results clearly demonstrate that the ESP terms are retrieving an average of 72 percent of the documents for terms in computer science and, over the whole database, an average of 75 percent. "The terms drawn from the table of contents increased retrieval 300 per- cent." The title contributed little more than 10 percent, while the LC subject headings re- trieved only 15 percent of the available documents, even though this research procedure favors LCSH (see table 2). This means that effectively the terms drawn from the table of contents increased re- trieval300 percent. ANALYZING THE QUALITY OF THE INFORMATION RETRIEVAL Having determined, without a doubt, that the additional access points have in- creased document recall, the next ques- tion is how many irrelevant documents or false drops have also been retrieved. In a first effort to analyze the impact of the ad- ditional headings on precision, computer science terms were searched as precoor- dinated headings or compound headings (e.g., natural language). These headings were selected by computer science faculty to represent significant areas of the field. Admittedly computer science terminol- ogy is very difficult to keep up with since it is rapidly changing. In the future we hope to replicate the study with compound terms selected from other more stable dis- ciplines. Only thirteen of the headings se- lected turned out to correspond to LC sub- ject headings, while eighteen did not. In other words, for 58 percent of the terms picked by the faculty there was no direct match on a heading. Records were key- word searched under these headings and once again we first counted the records re- trieved under the LC heading, then those Improving OPAC Subject Access 437 COMPARISON OF RECALL 6000 5747 5000 4000 3000 2000 1461 1000 231 806 306 0 LCSH TITLE ESP I • COMPLJTER • RANDOM I FIGURE2 Comparison of Recall TABLE2 BOOK NUMBERS RETRIEVED BY NATURAL-LANGUAGE SEARCH STRATEGIES Strategies Artificial intelligence Computer-aidea design Expert systems (computer science) Grammar, comparative and general-syntax Interactive computer systems Language and languages-congresses Linguistics-data processing Linguistics-congresses Machine theory Parsing (Computer grammar) Question-answering systems Semantics *The books were numbered 1- 25 . recovered by the title. The remainder were attributed to ESP headings. Faculty were asked to judge the relevance of the re- trieved articles to the topic as they under- stood it. However, they were told to inter- pret relevance in its broadest sense, that is, articles that could be of interest to someone working on that general topic as it relates to the discipline of computer sci- ence. 1, 2 3 1,14 Book Numbers* 8, 21, 23 11, 13, 16 4,22 5, 8, 9, 10, 11, 15, 16, 17, 21,25 6,19 7 5, 12, 19,21 16, 18, 20 16, 18,24 When an LC heading exists there were only 12.5 percent irrelevant documents, almost half of which are attributable to the LC heading. Surprisingly, the title and ESP terms combined contributed only 16.8 percent to the relevant pool here when the terms have been precoordina- ted. This appears to vindicate the librari- ans who feel that LCSH does provide an invaluable grouping function. This is only 438 College & Research Libraries half the story, however. The LC headings that exist in computer science are of a very general nature, as can be seen from the chart in appendix A. It is important to note that the eighteen headings where no LC match exists recovered 318 documents, of which 220 or 69 percent were deemed rele- vant by the faculty. In all probability these documents would not have been recov- ered otherwise. Title appears to contribute much more where no LC heading exists, that is, only 6 percent of the relevant mate- rial when a heading exists, and up to 49 percent (as well as a higher percentage of irrelevant hits) where no heading exists. The study indicates clearly the value of ad- ditional access points, but failed to clarify the contribution of the title. If title entries were omitted, would the ESP terms have recovered the documents anyway? RETRIEVAL WHERE NO LSCH HEADING EXISTS Another issue is what happens when no LCSH )leading exists, as in the case of nat- ural language, the term chosen for more de- tailed study. Thirty books were retrieved using a keyword search for natural and lan- guage. In this group, five (17 percent) were considered irrelevant. In this small collection of twenty-five relevant books we looked for the possibil- ity of grouping by call number but found twelve different call numbers spread across three schedules, as shown: Science Literature and Q335 QA76.7 QA76.9 (5 books) Language P23 P25 P98 P123 P302 P325 PE1074.5 Technology TA174 A similar scattering occurred under the LCSH headings. There were twenty- seven different subject headings, with a total of fifty-five keywords, giving a book average of 2.2, a slightly higher average yield than has been reported in other stud- ies. Certainly plenty of access points are provided. What is alarming is that no one heading provided any sort of useful grouping function for retrieval purposes. The headings that retrieved three or more books in this group were as follows: September 1988 Grammar, comparative and general-syntax Interactive computer systems Linguistics- data processing Parsing (computer grammar) Semantics 4 3 10 3 3 Linguistics-data processing performed best in retrieving ten of the twenty-five books (40 percent}, but failed to pull two out of three in the highly relevant sub- topics, syntax and semantics. The fact is that twenty-six of the twenty-seven head- ings in this area will locate three or fewer of the twenty-five books on the subject, an effective yield of 12 percent or less. The next question is, given an ambitious searcher, what would be the minimum number of separate subject headings needed to retrieve all twenty-five docu- ments? It took twelve different strategies (see table 2). The most alarming point is that a key- word search restricted to LC headings with the words natural and language will yield no hits even though natural- language processing is a well-recognized topic within artificial intelligence. If we follow the "snowball" strategy discussed in the literature (i.e., drawing on subject headings provided by those re- trieved books deemed relevant}, it is clear that the reader would recall vast amounts of irrelevant literature (see table 3). In all, eight of the twenty-seven headings can be definitely considered relevant and six marginal, while thirteen can be rejected as virtually useless because they are too broad. IMPACT ON USERS When the ESP procedure had been op- erational for six months, a structured in- terview survey of sixty-seven users showed that 38 percent had found the en- hanced terms useful, 19 percent had not, and 43 percent were not sure or were un- aware of the facility. The users, a stratified random sample of those registered with the library, volunteered the following rea- sons: 41 percent had no need to use the Improving OP AC Subject Access 439 ESP CONTRIBUTION TO RECALL High Relevance Ambiguity Anap.hora Discourse analysis Generative grammar Parsing-Computer grammar Question-answering systems Machine translation Semantics 1 0.03°/o II LCSH IIIII TITLE til ESP FIGURE3 Extent to which ESP Increased Document Recovery TABLE 3 RANKING OF LCSH HEADINGS AVAILABLE Relevant Marginal Cleopatra Grammar-Comparative and general -Syntax Artificial Intelligence Linguistics-Data processing Logic-Symbolic and mathematical Machine theory Useless Electronic digital computers-programming Engineering design English language-Data processing Interactive computer systems Language and languages Languages-philosophy Programming languages- Electronic computers-semantics Visual communication 440 College & Research Libraries terms, for seven percent they facilitated access to more material, but for 52 percent they gave more information about indi- vidual books. Besides the clear implication that more publicity and reader education are needed, the responses indicate that the improvement in recall was taken for granted while the increased information available to evaluate the relevance of items was seen as the major benefit. As onere- spondent put it: ''It gives more informa- tion about the book which saves me hav- ing to actually get the book.'' However, some critical comments sug- gested areas for software enhancement, in the areas of synonyms and thesaurus con- trol. Truncation was also suggested as a possible enhancement. Now, that a satis- factory method of increasing the number of access points has been found, and the number of documents recovered has grown dramatically, improved user inter- faces for conducting the search process and increasing the precision of search vo- cabulary are needed. Some of the recom- mendations that emerged from this study were as follows: 1. It is important to provide keyword searching in LC headings, also to display scope notes to indicate where they differ. It is not always clear to the uninformed why a distinction is made between infor- mation retrieval (the activity) and infor- mation storage and retrieval systems (the software that permits the activity). 2. Since the number of records re- trieved even in this small database of 160,000 titles is unmanageable, further study is needed of methods of assisting users to narrow their searches once they have found the topic of interest. In this way they can focus on the books that best address their information needs. CONCLUSION It is important to clarify the issues in- September 1988 volved in providing subject access to books. For years the number of subject headings, or access points, was restricted by the size of the catalog card and a tradi- tion that demanded, for economy of card catalog maintenance only, that subject ac- cess should only be provided for the main topic of the book. Consequently, an aver- age of only two headings per book is pro- vided even today. For periodical articles, experience with . providing keyword access has shown that recall and user satisfaction are increased by providing thirty to fifty access points per document. The cost has not proven to be prohibitive. The Enhanced Subject Project has dem- onstrated that the use of contents terms is a viable and cost-effective technique for dramatically increasing the number of subject access points to the contents of books without a serious increase in false drops. It has also shown once again the failure to provide adequate subject access of the accepted system of providing one or two controlled headings to each book. The controlled headings do help to provide a grouping function · and should not be abandoned but rather enhanced, as sug- gested here. Much further work is needed on the vocabulary distribution patterns we identified and the implications for searching. The ADFA library continues to input subject access terms, thereby providing an increasing pool of research data. It is hoped that additional studies will be im- plemented as more data becomes avail- able. The research reported here is the first . step of a larger project to develop an ex- pert system that can handle authority con- trol by mapping the terms selected by the users to the very restricted vocabulary of LCSH. REFERENCES 1. Books Are for Use, Final Report to the Council on Library Resources, Syracuse University, School of Informa- tion Studies, Subject Access Project, (Syracuse, N.Y.: Syracuse Univ., 1978), p.177. (ED 156131). 2. Ibid. 3. Sue Beatty, "An Experiment in Enhanced Subject Access," Cataloging Australia 11, no.4:86-95 Improving OPAC Subject Access · 441 (Dec. 1985); Alex Byrne, "Life Wasn't Mean to Be Whimsical: Painless Subject Augmentation," Australasian College Libraries 4, no.2:83-90 Oune 1986). 4. AUSMARC, the Australian MARC specification, which is used in the academy library, has no 653 field. The 650 field is used for topical subject headings, 600 and 610 for name headings, while 651 contains geographic subject headings with data, according to AACR2 and LCSH. However, the National Library of Australia advised that the 653 field would probably be specified for data similar to that generated by ESP and that its use would be consistent with LCMARC. American practice would dictate the use of the 690 field for local headings. 5. Barbara Settel, ed., Subject Description of Books: A Manual of Procedures for Augmenting Subject Descrip- tions in Library Catalogs (Syracuse, N.Y.: Syracuse Univ., School of Information Studies, Research Study #3, 1977). 6. Ibid. 7. H. Mary Micco, An Exploratory Study of Three Subject Access Systems in Medicine: LCSH, MeSH, PRECIS (Ph.D. diss., Univ. of Pittsburgh, 1980). APPENDIX A: ANALYSIS OF RETRIEVAL Comparison of Precision of LCSH, Title, ESP Relevant Irrelevant Search Terms LCSH Title ESP LCSH Title ESP LCSHExists Comprehension 4 1 2 7 3 Computer interfaces 17 3 5 2 1 CritiCal path 35 0 2 Expert systems 58 2 20 3 Generative grammar 20 2 1 Information retrieval 222 9 8 3 Machine learnin6 8 2 4 3 Machine translation 4 3 1 Parsing (computer grammar) 8 1 7 1 Pattern recognition 56 15 12 1 2 Question answering systems 4 2 1 2 1 Semantics 55 9 12 30 2 9 Speech recognition 9 4 6 2 3 Total 500 46 85 45 17 31 NoLCSH Automatic classification 1 4 2 Computational linguistics 1 Computer communication 9 3 10 22 Computer memory 3 12 Computer speech 4 8 3 Computer translation 1 2 E:~n~~f~Z systems* 11 13 4 9 ~~~~~~ tutoring 9 2 Natural language 24 1 2 3 Optical storage 1 3 2 Query+ 1 1 9 Risk analysist 3 8 22 Software engmeering 43 7 " 2 7 Text analysis 3 3 21 3 Text recognition 3 5 Video systems 6 3 3 Total 110 105 3 37 58 *The LC match was Decision making-Data processing. +The LC term was Query languages . :j:The LC terms were Risk, risk management.