FILING, FILTERING AND THE FIRST FEW FOUND

INTRODUCTION

It has been customary to arrange catalog records alphabetically in book-form catalogs, card catalogs, and, now, online catalogs. There has been some disagreement over details of filing concerning, for example, the treatment of numerals, modified and non-Latin letters, and the choice between "letter-by-letter" and "word-by-word". "Structured" or "categorical" arrangement, departing from strict alphabetization, has been used in some special cases, notably in the arrangement for prolific authors, of historical subdivisions, and, in the past, for parts of the Bible. A "structured" approach arranges headings based on the categories to which they belong. For example, period subdivisions are arranged chronologically rather than alphabetically. Nevertheless, alphabetization has been the dominant and pervasive principle.

In the U.S.A. there has been a marked preference for providing subject access by alphabetized verbal subject headings rather than by a classified subject catalog. The use of verbal subject headings and the dictionary catalog that then becomes possible increases the prominence (and complexity) of alphabetical ordering.

With some exceptions, the alphabetization of the cards in card catalogs has been carried over to the display of records in online catalogs. At least, alphabetic ordering by main entry has become standard, though dictionary catalog arrangement has not. But the necessity and benefits of alphabetizing catalog records by main entry in online catalogs are less certain than in card files. Feasible and attractive alternatives will be discussed below, but first the issues underlying alphabetic ordering will be reviewed. (For a good, detailed discussion of alphabetical and structured arrangements of subject headings in online catalogs see Headings for Tomorrow (1992). For an older more general discussion see Helfer (1982).

THE FUNCTIONS OF THE CATALOG

The term "retrieval" is ambiguous in that it can subsume three distinct functions:

i. Selecting (or identifying) documents (as in subject or author searching);

ii. Finding a document (or a record for a document) with known individual characteristics ("known item searching"), usually, but not necessarily, to ascertain the location of that document; and

iii. Fetching (delivering) documents.

Since the nineteenth century, library catalogs have been designed to support selecting in addition to the finding (locating) function of a catalog (Buckland 1988). Like bibliographies, modern catalogs are designed to enable the library user to identify or select items in the collections on a particular subject, by a particular author, and, to some extent, with other characteristics, such as form. However, the distinctive function of a library catalog is as a finding list, to enable the library user to ascertain the location (call number) of specific items in the collection once the identity of the item is known. If it did not provide the shelf location of items it would not be considered a catalog.

Alphabetic ordering is, in essence, a very effective device for the locating of words, usually names: names of people, names of places, names of things, and names of subjects. Alphabetic ordering is effective in the card catalog because as a "manual" system it enables the determination, both for filing and for finding, of the proper location of headings, of sub-headings, and of individual records filed under headings and subheadings. It is as a result of this hierarchical process of locating that searching become feasible. Even with a subject catalog on cards arranged in classified order, an alphabetically ordered "relative index" of the names of subjects is needed for effective use. An obvious but significant qualification is that alphabetic ordering becomes useful only as the number of items increases. Where there is one or very few items, alphabetic ordering is of little or no value.

Effective though alphabetic ordering may be for locating names, including the names of subjects, it is a relatively weak technique for the selecting function because attributes of records other than names are not well served by alphabetic order. Not all significant attributes are names and alphabetizing may yield an ordering of limited usefulness. Filing by date is better for currency and a classified arrangement, as on the shelves, is better for bringing related subject materials together. "If the names of the classes, in a natural language, are used to arrange them, we do not get a helpful order. In fact, names scatter classes in a most unhelpful chaotic order. It will give us an order like algebra, anger, apple, arrogance, asphalt, and astronomy" (Ranganathan 1951).

Selection involves the principle of sorting (or partitioning) the file of catalog records into at least two sets: those selected and those not selected. The traditional approach to the evaluation of retrieval system performance, using recall and precision as measures, is based on how closely the partitioning achieved by the selection process matches an ideal partitioning based on the perceived relevance and non-relevance of the individual documents.

ONLINE SEARCHING

Although the objective of retrieval in the sense of selecting is supposed to be the identification of "relevant" items, relevance is a subjective and situational phenomenon. In practice, therefore, what all mechanical systems retrieve is the records matching some objective attribute, such as subject headings or word occurrence, to some measurable degree. A virtue of online catalogs capable of searching on keywords and of performing Boolean (i.e. postcoordinate) searches is that they can be used easily and very flexibly to create partitions of the file of records in an astronomical number of combinations on an ad hoc basis. Indeed, one of the difficulties with current online catalogs is that the size of the retrieved set is difficult to predict: For 30-40% of searches zero records are retrieved and in the other searches hundreds or even thousands of records are commonly retrieved (Markey 1984, Matthews & Lawrence 1984). It is within these retrieved sets that alphabetic ordering by heading, traditional in card catalogs, has been continued as the conventional ordering. The catalog may report, say, 600 records found and is normally programmed to sort the records into alphabetic order of main entry in order to display the first screenful.

There is, however, no logical or technological necessity for the partitioning to be limited to only two sets. There could be any number of sets. One view, strongly held in some circles, is that partitioning should be taken to the limit, with as many subsets as there are records. This yields a ranking of each individual record in decreasing order of some computed measure taken to be an indicator of probable relevance ("document ranking"). Of course, in this case each set has a population of just one record and so there is no opportunity for any alphabetic ordering. More generally the scope for alphabetic ordering arises to the extent that the number of records within sets is large enough to need ordering for locating specific records. Later we shall show that alphabetic ordering is not necessarily to be preferred, even when feasible.

CARD CATALOGS RECONSIDERED

The technology underlying online catalogs lends itself to the dynamic selection of sets of records in many different ways. The computer could also be programmed to sort any set of retrieved records for display in any definable order. A characteristic of card, book-form, and microform catalogs is the stability of the ordering of the records. A strength of card catalogs compared with book-form and microform catalogs is that individual records can be continuously added and removed--and even rearranged. Nevertheless, although feasible in principle, one would not want to rearrange the order of the cards. To subdivide a dictionary catalog into separate name, title, and subject heading files or to rearrange cards under each subject headings from alphabetic to, say, reverse chronological or call number order would be possible but would be a tedious and expensive undertaking. One could not afford to switch lightly from one order to another, let alone on an ad hoc basis. Instead, holes in the cards and metal retaining rods are used to restrain the cards to the chosen ordering.

Although card catalogs do not permit the creation of arbitrary sets in the same way that online catalogs do, some of the same capability exists in ways that may not be obvious:

1. The set of records filed under any one name, title, or subject heading would be the same as the set retrieved using the same name, title, or subject heading used as an exact search in an online catalog;

2. Subdividisions of Library of Congress Subject Headings are used to create a multiplicity of smaller subsets. The practice of adding subdivisions to create complex Library of Congress Subject Headings (LCSH) has the same kind of effect, achieved prior to the search (pre-coordination), as post-coordinate searches which use subject keywords using a Boolean AND;

3. Boolean AND searches can be performed in a card catalog, if one term is an established heading, by scanning the cards filed under that heading for the attributes specified in the rest of the AND statement;

4. Similarly, Boolean AND NOT searches can be handled, if the first term is an established heading, by scanning under the heading containing the first attribute and rejecting those records that contain the second; and

5. Boolean OR searches can be achieved in a card catalog by looking under two different headings, at two disjoint sets.

Nevertheless, for all that one may argue that card catalogs have, in principle, some of the functionality associated with online catalogs, the dominance of a single, linear alphabetic ordering remains a basic difference. Alphabetic ordering is needed not only to locate the records for known-item searching and to locate the headings and subheadings needed for selecting (e.g. subject searching), but also to determine the proper location in which to file each record in the first place. Since it is a human searcher who must locate each record, alphabetic ordering has emerged, with little challenge, as the preferred ordering not only for headings but also for the subset of records under each heading or subheading. Alternatives, such as arranging cards under each heading in chronological order (Uhlirz 1884), reverse chronological order to present the searcher with the most recent material first (Merrill 1934, Kleist 1945, Werking, Miller & Whaley 1986), or, in special cases, categorical arrangements, have merits, but alphabetic ordering is almost universally adopted. (For the slow development of alphabetizing, which was little used before printing, see Daly 1967 [and other cites appended]).

ORDERING RECORDS IN ONLINE CATALOGS

The prevalent alphabetic ordering of catalog records in card catalogs has been imposed on online catalogs, which typically store their records by an identification number based on the order in which records were received. But the change of technology from card to electronic means that alphabetic ordering has now become a matter of choice rather than of practical necessity: The computer could be programmed to provide any one or combination of a variety of orderings: chronological, reverse chronological, by language, by proximity of holding library, by loan status, by any attribute in or implicit in the data available to the computer. In brief, in online catalogs, unlike card catalogs, records do not have to be presented in alphabetic or, indeed, in any other order. Further, no one single consistent choice needs to be made because each set of records retrieved has to be newly ordered for display. The retrieved records could as well be arranged (and rearranged) in any of a variety of ways at choice.

More generally, and perhaps, more importantly, the flexibility of an online catalog to create retrieved sets at will also enables the online catalog to create subsets at will from the retrieved set. The sorting can range from the standard form of two sets--retrieved and not-retrieved--to the newer orthodoxy of strict document ranking. Better still, online systems can be made to subdivide a retrieved set into smaller subsets as desired or create aggregations of closely ranked records. These subsets or aggregations and, also, the records within them can be arranged at will in several different ways using whatever attributes are present in the records. A form of filtering is achieved. Choosing some other ordering first to create subsets prior to alphabetic ordering can make a substantial difference as the examples below will show.

Meanwhile it should be noted that partitioning, sorting, ordering, and filing are logically the same operation. Any sorting or ordering, as in alphabetic filing, is an example of partitioning. Since the locating and selecting functions of retrieval are achieved by partitioning, it follows that online catalogs (and online bibliographic systems generally) should be viewed not as a single retrieval process, but rather as a series of consecutive retrieval processes. In a typical online catalog the first retrieval process establishes (partitions off) the set to be retrieved. This first retrieval process may well be iterated one or more times. What may be overlooked is that, after records have been selected, there is always another retrieval process which orders (partitions) the retrieved set into some specified arrangement for display, usually, but not necessarily, alphabetic order of main entry.

The significance, even the existence, of this secondary retrieval process is masked in two cases. Firstly, the primary retrieval process may be allowed to pre-empt the second, as when retrieval techniques based on strict document ranking are adopted. Even here it would be a feasible option to allow the aggregation and rearrangement of sets of closely ranked documents should one wish. After all, one's confidence in the ranking depends on how perfectly the criteria used for ranking represent one's interests and the initial search may not have reflected all of one's preferences fully. Secondly, the ordering of records for display is easily overlooked if it is in (or a practice simply carried over from) a card catalog, where it is, in effect, a necessity.

The choice, and flexibility in choice, of ordering of records becomes more important for an online system than in a card catalog because of the mechanics of displaying records online. Users generally follow the default display option: They normally start by inspecting the first screen of records as selected for presentation by the system. Typically, alphabetic order of main heading is the only option provided unless the user modifies the search itself. Users may or may not choose to page on through successive screens of records, but, ordinarily, they do not. The consequence is that the first few records displayed can be expected to be the basis upon which the user will decide whether to display more, modify the search, attempt a new, different search, or abandon searching (Bates 1984, Berger 1992). The basis provided for this decision by conventional online catalog design is, therefore, influenced by how main entries happen to be spelled.

FILTERING

Bibliographic retrieval is normally a matter of searching. Rather unpredictable searches are made of a relatively stable set of records. Some other document retrieval systems, notably the screening of electronic mail and the analysis of electronic news services, are more a matter of filtering: Rather unpredictable flows of documents are compared with relatively stable search specifications. The secondary retrieval process of ordering retrieved sets of bibliographic records for display can be viewed as a form of filtering.

Tendency of existing online bibliographic systems and web-search engines to yield too many items is well-known. A plausible design ideal is that, regardless of the query and regardless of the size and contents of file, each search will ordinarily yield a small set of records, those most closely matching the searcher's preferences. This goal requires that the system know what preferences of the searcher should be used to reduce large sets. (This is an automated extension of what skilful use of search modifiers can achieve (Buckland & Florian 1991).

In the first example a subject keyword search for "Dresden" in the MELVYL catalog retrieves over 440 records (FIND SUBJECT DRESDEN, searched Dec 29, 1991). The 440 records are unlikely to be of equal interest to any catalog user. Browsing 440 records becomes tedious and alphabetic ordering by main entry is very unlikely to present anyone's preferred records first. In post-processing the workstation has been set to make three assumptions about the user's preferences: That, other things being equal, the user would tend to prefer recent material to older; English language works to foreign; and books held in the Berkeley campus to books held only on other campuses. These three preferences, operating on relevant fields in the downloaded records, are used to sort the records into a three dimensional array of subsets. Table 1 is a tabular analysis of the retrieved set.

The searcher may choose to "zoom in" on the last decade. See Table 2.

Any subset (the contents of any cell) can be selected for inspection, but a default display could reasonably start, in either table, with the subset in the top left corner since that represents the preferred value for all three preferences. Thereafter, the searcher can inspect other subsets. The default ordering (ranking) of subsets would reflect the relative importance of the preferences. For example, if location mattered more than language and language more than recency of publication, extending the display would, in this case, proceed cell by cell from top to bottom in the left-most column, then down the each column, from left to right, successively. Records are alphabetized within each subset containing two or more records. Even though the same pool of records is being drawn on, the effects of such filtering can be striking, especially if, as is common, only the first few records of any large retrieved set are inspected.

As an example of the effects of filtering, imagine that someone on the Berkeley campus wants a general work on chemistry. Consider and compare the following examples of the first few records found with and without filtering. A search for general works on Chemistry in the MELVYL catalog would be correctly expressed as FIND XSU "CHEMISTRY" (i.e. an exact search for the Library of Congress Subject Heading "Chemistry" without any subheadings or truncation). MELVYL's retrieves 1,494 records and, following standardized alphabetic ordering, displays the following initial selection of records. It is a motley assortment ranging in publication date from 1807 to 1981. Three out of six are not held at Berkeley. Not one seems appropriate as a general introduction to present-day chemistry. (Searched December 29, 1991. Records lightly edited for concise presentation.):

1. 31st general assembly of the International Union or Pure and Applied Chemistry, Leuven, Belgium, 1981 : Report of the delegation from the National Academy of Sciences. [Washington, D.C.] : National Academy Press, [1982?]. Two holdings, one at Berkeley.

2. Accademia del cimento, Florence. Essayes of natural experiments made in the Academie del Cimento ... [1684]. A facsimile ... New York : Johnson Reprint Corp., 1964. Seven holdings, none at Berkeley.

3. Accum, Frederick, 1769-1838. System of theoretical and practical chemistry / by Fredrick Accum. Philadelphia : Kimber & Conrad, 1808. [Microprint reproduction]. One holding, at Berkeley.

4. Accum, Friedrich Christian, 1769-1838. System of theoretical and practical chemistry ... [2nd American ed.]. Philadelphia : Kimber and Conrad, 1814 ([United States] : Merritt). [Microprint reproduction]. One holding, at Berkeley.

5. Accum, Friedrich Christian, 1769-1838. System of theoretical and practical chemistry ... 2nd ed. [London etc.] The author, 1807. One holding, not at Berkeley.

6. Acids and bases; a collection of papers by Norris F. Hall [and others] ... [Easton, Pa..] The Journal of chemical education [c1941]. Two holdings, neither at Berkeley.

The same 1,494 records, if filtered on the basis of preferences for the English language, for recency of publication, and for items held on the Berkeley campus (as half of the preceding list is not) before alphabetic ordering by main entry, yields the following first few. All held at Berkeley. All are modern, general introductions to Chemistry in English published within the previous three years.

1. Chang, Raymond. Chemistry . 4th ed. New York : McGraw-Hill, c1991.

2. Feigl, Dorothy M. Foundations of life : an introduction to general, organic, and biological chemistry / Dorothy M. Feigl, John W, Hill, Erwin Boschmann. 3rd ed. New York : Macmillan Pub. Co., c1991.

3. McQuarrie, Donald A. General chemistry / Donald A. McQuarrie, Peter A. Rock. 3rd ed. New York : W.H. Freeman, c1991.

4. Oxtoby, David W. Principles of modern chemistry / David W. Oxtoby, Norman H. Nachtrieb. 2nd ed. Philadelphia: Saunders College Pub., c1990.

5. Bodner, George M. Chemistry, an experimental science / George M. Bodner, Harry L. Pardue. New York : Wiley, c1989.

6. Sherman, Alan. Chemistry and our changing world / Alan Sherman, Sharon J. Sherman. 2nd ed. Englewood Cliffs, N.J. : Prentice Hall, c1989.

The second set of titles promises a much greater likelihood of user satisfaction in normal library use as a selection of general works on chemistry. All of the titles on the second list could eventually have been found using the first list, but only after examination of well over a thousand records.

In practice inexpert searchers are more likely to use the simpler search command FIND SUBJECT CHEMISTRY, which generates a subject keyword search and retrieves not only the 1,494 of the previous search but also another 21,000 records for works with any LSCH that includes the word "chemistry." The first few records displayed for this search are:

1. 1,3-dipolar cycloaddition chemistry / edited by Albert Padwa. New York : Wiley, c1984. Nine holdings, one at Berkeley.

2. 20 Jahre Zentralinstitut fur Physikalische Chemie : Festkolloquium ... Berlin : Akademie-Verlag, 1978. One holding, not at Berkeley.

3. 25 years of structural chemistry in Turku University : from 1,3-dioxane to natural products ... Helsinki : Suomalainen tiedeakatemia, 1990. One holding, not at Berkeley.

4. 25 years of the Research Institute for Soil Science and Agricultural Chemistry of the Hungarian Academy of Sciences. Budapest : [s.n.], 1974. One Berkeley holding, shelved off campus.

5. 31st general assembly of the International Union or Pure and Applied Chemistry, Leuven, Belgium, 1981 : Report of the delegation from the National Academy of Sciences. [Washington, D.C.] : National Academy Press, [1982?]. Two holdings, one at Berkeley.

6. 50 [i.e. Piat'desiat] let: Sovetskaia khimicheskaia nauka i promyshlennost'. Moskva, Khimia, 1967. One holding, not at Berkeley.

This selection is more modern than the first list shown, but in this case all of them are rather exotic works of narrowly specialized or historical interest. Again, half are not held at Berkeley. Not a single one seems to be a suitable find for anyone, especially any non-chemist, who was simply seeking a book about chemistry. Yet these are the first few found by a simple, typical search in an online catalog of standard design. The results are exacerbated by the large size of the MELVYL data-base, but the catalog of any large, long-established library will tend to yield such results of this kind if standard catalog practice is followed.

Filtering and sorting can, in principle, operate on any data in the catalog record and on anything implicit in the records. For example, a searcher may reasonably want an introductory college level treatment of a subject. This intellectual level is not coded in catalog records, but it might be inferred. For example, the three undergraduate libraries in the University of California (Berkeley Moffitt Library, UCLA College Library, and the University of California San Diego Cluster Library) maintain collections generally at this level. A stated preference for titles held at one or more of these three libraries but not limited to the copies in those libraries would tend to bring forward general introductory works. Filtering as before but with this additional preference for works selected by one or more undergraduate library applied to the 22,341 records retrieved by FIND SUBJECT CHEMISTRY generates a display of records, all reflecting Berkeley holdings, which begins with:

1. Atkinson, Daniel E. Dynamic models in chemistry : a workbook of computer simulations using electronic spreadsheets / Daniel E. Atkinson, Douglas C. Brower, Ronald W. McClard. Marina del Rey, Calif. : N. Simonson, c1990.

2. Chang, Raymond. Chemistry. 4th ed. New York : McGraw-Hill, c1991.

3. Feigl, Dorothy M. Foundations of life : an introduction to general, organic, and biological chemistry / Dorothy M. Feigl, John W, Hill, Erwin Boschmann. 3rd ed. New York : Macmillan Pub. Co., c1991.

4. Manahan, Stanley E. Environmental chemistry / Stanley E. Manahan. 4th ed. Chelsea, MI : Lewis Publishers, c1990.

5. Minkin, V. I. (Vladimir Isaakovich). Quantum chemistry of organic compounds : mechanisms of reactions / V.I. Minkin, B.Ya. Simkin, R.M. Minyaev. Berlin ; New York : Springer-Verlag, c1990.

6. Practicing to take the GRE chemistry test. 2nd ed. Princeton, N.J. : Educational Testing Service for the Graduate Record Examinations Board, c1990.

Again, with this second pair of lists, the difference between the first, conventional ordering and the second, filtered ordering of the same results of the same subject search is remarkable. The contrast indicates that the routine adherence to alphabetic ordering is a serious disservice. Online catalog displays can be customized to provide differently tailored service to different terminals (Wellings 1990). Our experience indicates that filtering and ordering options can be very useful, especially when they can be chosen dynamically by the individual searcher.

SUMMARY

Alphabetic ordering is necessary for locating headings, subheadings, and individual records filed in a card catalog. Since only one, fixed ordering arrangement can be supported, alphabetic order has dominated, despite the weakness of this arrangement, for the selecting (e.g. subject searching) role of the catalog.

Alphabetic ordering has been carried over into online catalogs even though the technological constraints are different. Searching can be seen as implying multiple retrieval processes: Partitioning to retrieve; and ordering for display. A choice of alternative orderings could be supported, including orderings based on multiple criteria such as date, language, and convenient availability. Subordinating the traditional alphabetic ordering by main entry to other forms of filtering using non-topical attributes can yield useful results. The carrying over of the alphabetic ordering of the card catalog to the online catalog appears to have been at a high price for users. Future online catalogs and online bibliographies should allow users to specify their own personal preferences as routine filtering devices for ordering the display of any or all of their search results. The result is not only a significant enhancement in service but also an indictment of the traditional dominance of alphabetic ordering.

Endnote: The tables and examples are derived from work on the development of a prototype adaptive online catalog (Otlet's Adaptive Search Information Service, OASIS) in 1992. A workstation (DECstation 5000/200) acted as a front-end to the University of California's MELVYL (tm) online catalog containing some 7 million titles held in the libraries of the nine campuses of the University. The workstation ordinarily serves transparently, as if a dumb terminal connected to MELVYL, but, through pre-processing and post-processing, capabilities not available in MELVYL can be demonstrated. A command FILTER downloaded a set of records from MELVYL into the workstation and then sorts them according to the user's preferences. See Buckland, Butler, Norgard and Plaunt (1992).