Bellcore performs all this scanning, creates a page-image file, and also selects from the pages the graphics, to mix with the text file (which is discussed later in the Workshop). The user is always searching the ASCII file, but she or he may see a display based on the ASCII or a display based on the images.

LESK ill.u.s.trated how the program performs page a.n.a.lysis, and the image interface. (The user types several words, is presented with a list-- usually of the t.i.tles of articles contained in an issue--that derives from the ASCII, clicks on an icon and receives an image that mirrors an ACS page.) LESK also ill.u.s.trated an alternative interface, based on text on the ASCII, the so-called SuperBook interface from Bellcore.

LESK next presented the results of an experiment conducted by Dennis Egan and involving thirty-six students at Cornell, one third of them undergraduate chemistry majors, one third senior undergraduate chemistry majors, and one third graduate chemistry students. A third of them received the paper journals, the traditional paper copies and chemical abstracts on paper. A third received image displays of the pictures of the pages, and a third received the text display with pop-up graphics.

The students were given several questions made up by some chemistry professors. The questions fell into five cla.s.ses, ranging from very easy to very difficult, and included questions designed to simulate browsing as well as a traditional information retrieval-type task.

LESK furnished the following results. In the straightforward question search--the question being, what is the phosphorus oxygen bond distance and hydroxy phosphate?--the students were told that they could take fifteen minutes and, then, if they wished, give up. The students with paper took more than fifteen minutes on average, and yet most of them gave up. The students with either electronic format, text or image, received good scores in reasonable time, hardly ever had to give up, and usually found the right answer.

In the browsing study, the students were given a list of eight topics, told to imagine that an issue of the Journal of the American Chemical Society had just appeared on their desks, and were also told to flip through it and to find topics mentioned in the issue. The average scores were about the same. (The students were told to answer yes or no about whether or not particular topics appeared.) The errors, however, were quite different. The students with paper rarely said that something appeared when it had not. But they often failed to find something actually mentioned in the issue. The computer people found numerous things, but they also frequently said that a topic was mentioned when it was not. (The reason, of course, was that they were performing word searches. They were finding that words were mentioned and they were concluding that they had accomplished their task.)

This question also contained a trick to test the issue of serendipity.

The students were given another list of eight topics and instructed, without taking a second look at the journal, to recall how many of this new list of eight topics were in this particular issue. This was an attempt to see if they performed better at remembering what they were not looking for. They all performed about the same, paper or electronics, about 62 percent accurate. In short, LESK said, people were not very good when it came to serendipity, but they were no worse at it with computers than they were with paper.

(LESK gave a parenthetical ill.u.s.tration of the learning curve of students who used SuperBook.)

The students using the electronic systems started off worse than the ones using print, but by the third of the three sessions in the series had caught up to print. As one might expect, electronics provide a much better means of finding what one wants to read; reading speeds, once the object of the search has been found, are about the same.

Almost none of the students could perform the hard task--the a.n.a.logous transformation. (It would require the expertise of organic chemists to complete.) But an interesting result was that the students using the text search performed terribly, while those using the image system did best.

That the text search system is driven by text offers the explanation.

Everything is focused on the text; to see the pictures, one must press on an icon. Many students found the right article containing the answer to the question, but they did not click on the icon to bring up the right figure and see it. They did not know that they had found the right place, and thus got it wrong.

The short answer demonstrated by this experiment was that in the event one does not know what to read, one needs the electronic systems; the electronic systems hold no advantage at the moment if one knows what to read, but neither do they impose a penalty.

LESK concluded by commenting that, on one hand, the image system was easy to use. On the other hand, the text display system, which represented twenty man-years of work in programming and polishing, was not winning, because the text was not being read, just searched. The much easier system is highly compet.i.tive as well as remarkably effective for the actual chemists.

+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ ERWAY * Most challenging aspect of working on AM * a.s.sumptions guiding AM's approach * Testing different types of service bureaus * AM's requirement for 99.95 percent accuracy * Requirements for text-coding *

Additional factors influencing AM's approach to coding * Results of AM's experience with rekeying * Other problems in dealing with service bureaus * Quality control the most time-consuming aspect of contracting out conversion * Long-term outlook uncertain *

To Ricky ERWAY, a.s.sociate coordinator, American Memory, Library of Congress, the constant variety of conversion projects taking place simultaneously represented perhaps the most challenging aspect of working on AM. Thus, the challenge was not to find a solution for text conversion but a tool kit of solutions to apply to LC's varied collections that need to be converted. ERWAY limited her remarks to the process of converting text to machine-readable form, and the variety of LC's text collections, for example, bound volumes, microfilm, and handwritten ma.n.u.scripts.

Two a.s.sumptions have guided AM's approach, ERWAY said: 1) A desire not to perform the conversion inhouse. Because of the variety of formats and types of texts, to capitalize the equipment and have the talents and skills to operate them at LC would be extremely expensive. Further, the natural inclination to upgrade to newer and better equipment each year made it reasonable for AM to focus on what it did best and seek external conversion services. Using service bureaus also allowed AM to have several types of operations take place at the same time. 2) AM was not a technology project, but an effort to improve access to library collections. Hence, whether text was converted using OCR or rekeying mattered little to AM. What mattered were cost and accuracy of results.

AM considered different types of service bureaus and selected three to perform several small tests in order to acquire a sense of the field.

The sample collections with which they worked included handwritten correspondence, typewritten ma.n.u.scripts from the 1940s, and eighteenth-century printed broadsides on microfilm. On none of these samples was OCR performed; they were all rekeyed. AM had several special requirements for the three service bureaus it had engaged. For instance, any errors in the original text were to be retained. Working from bound volumes or anything that could not be sheet-fed also const.i.tuted a factor eliminating companies that would have performed OCR.

AM requires 99.95 percent accuracy, which, though it sounds high, often means one or two errors per page. The initial batch of test samples contained several handwritten materials for which AM did not require text-coding. The results, ERWAY reported, were in all cases fairly comparable: for the most part, all three service bureaus achieved 99.95 percent accuracy. AM was satisfied with the work but surprised at the cost.

As AM began converting whole collections, it retained the requirement for 99.95 percent accuracy and added requirements for text-coding. AM needed to begin performing work more than three years ago before LC requirements for SGML applications had been established. Since AM's goal was simply to retain any of the intellectual content represented by the formatting of the doc.u.ment (which would be lost if one performed a straight ASCII conversion), AM used "SGML-like" codes. These codes resembled SGML tags but were used without the benefit of doc.u.ment-type definitions. AM found that many service bureaus were not yet SGML-proficient.

Additional factors influencing the approach AM took with respect to coding included: 1) the inability of any known microcomputer-based user-retrieval software to take advantage of SGML coding; and 2) the multiple inconsistencies in format of the older doc.u.ments, which confirmed AM in its desire not to attempt to force the different formats to conform to a single doc.u.ment-type definition (DTD) and thus create the need for a separate DTD for each doc.u.ment.

The five text collections that AM has converted or is in the process of converting include a collection of eighteenth-century broadsides, a collection of pamphlets, two typescript doc.u.ment collections, and a collection of 150 books.

ERWAY next reviewed the results of AM's experience with rekeying, noting again that because the bulk of AM's materials are historical, the quality of the text often does not lend itself to OCR. While non-English speakers are less likely to guess or elaborate or correct typos in the original text, they are also less able to infer what we would; they also are nearly incapable of converting handwritten text. Another disadvantage of working with overseas keyers is that they are much less likely to telephone with questions, especially on the coding, with the result that they develop their own rules as they encounter new situations.

Government contracting procedures and time frames posed a major challenge to performing the conversion. Many service bureaus are not accustomed to retaining the image, even if they perform OCR. Thus, questions of image format and storage media were somewhat novel to many of them. ERWAY also remarked other problems in dealing with service bureaus, for example, their inability to perform text conversion from the kind of microfilm that LC uses for preservation purposes.

But quality control, in ERWAY's experience, was the most time-consuming aspect of contracting out conversion. AM has been attempting to perform a 10-percent quality review, looking at either every tenth doc.u.ment or every tenth page to make certain that the service bureaus are maintaining 99.95 percent accuracy. But even if they are complying with the requirement for accuracy, finding errors produces a desire to correct them and, in turn, to clean up the whole collection, which defeats the purpose to some extent. Even a double entry requires a character-by-character comparison to the original to meet the accuracy requirement. LC is not accustomed to publish imperfect texts, which makes attempting to deal with the industry standard an emotionally fraught issue for AM. As was mentioned in the previous day's discussion, going from 99.95 to 99.99 percent accuracy usually doubles costs and means a third keying or another complete run-through of the text.

Although AM has learned much from its experiences with various collections and various service bureaus, ERWAY concluded pessimistically that no breakthrough has been achieved. Incremental improvements have occurred in some of the OCR technology, some of the processes, and some of the standards acceptances, which, though they may lead to somewhat lower costs, do not offer much encouragement to many people who are anxiously awaiting the day that the entire contents of LC are available on-line.

+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ ZIDAR * Several answers to why one attempts to perform full-text conversion * Per page cost of performing OCR * Typical problems encountered during editing * Editing poor copy OCR vs. rekeying *

Judith ZIDAR, coordinator, National Agricultural Text Digitizing Program (NATDP), National Agricultural Library (NAL), offered several answers to the question of why one attempts to perform full-text conversion: 1) Text in an image can be read by a human but not by a computer, so of course it is not searchable and there is not much one can do with it. 2) Some material simply requires word-level access. For instance, the legal profession insists on full-text access to its material; with taxonomic or geographic material, which entails numerous names, one virtually requires word-level access. 3) Full text permits rapid browsing and searching, something that cannot be achieved in an image with today's technology.

4) Text stored as ASCII and delivered in ASCII is standardized and highly portable. 5) People just want full-text searching, even those who do not know how to do it. NAL, for the most part, is performing OCR at an actual cost per average-size page of approximately $7. NAL scans the page to create the electronic image and pa.s.ses it through the OCR device.

ZIDAR next rehea.r.s.ed several typical problems encountered during editing.

Praising the celerity of her student workers, ZIDAR observed that editing requires approximately five to ten minutes per page, a.s.suming that there are no large tables to audit. Confusion among the three characters I, 1, and l, const.i.tutes perhaps the most common problem encountered. Zeroes and O's also are frequently confused. Double M's create a particular problem, even on clean pages. They are so wide in most fonts that they touch, and the system simply cannot tell where one letter ends and the other begins. Complex page formats occasionally fail to columnate properly, which entails rescanning as though one were working with a single column, entering the ASCII, and decolumnating for better searching. With proportionally s.p.a.ced text, OCR can have difficulty discerning what is a s.p.a.ce and what are merely s.p.a.ces between letters, as opposed to s.p.a.ces between words, and therefore will merge text or break up words where it should not.

ZIDAR said that it can often take longer to edit a poor-copy OCR than to key it from scratch. NAL has also experimented with partial editing of text, whereby project workers go into and clean up the format, removing stray characters but not running a spell-check. NAL corrects typos in the t.i.tle and authors' names, which provides a foothold for searching and browsing. Even extremely poor-quality OCR (e.g., 60-percent accuracy) can still be searched, because numerous words are correct, while the important words are probably repeated often enough that they are likely to be found correct somewhere. Librarians, however, cannot tolerate this situation, though end users seem more willing to use this text for searching, provided that NAL indicates that it is unedited. ZIDAR concluded that rekeying of text may be the best route to take, in spite of numerous problems with quality control and cost.

+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ DISCUSSION * Modifying an image before performing OCR * NAL's costs per page *AM's costs per page and experience with Federal Prison Industries *

Elements comprising NATDP's costs per page * OCR and structured markup *

Distinction between the structure of a doc.u.ment and its representation when put on the screen or printed *

HOOTON prefaced the lengthy discussion that followed with several comments about modifying an image before one reaches the point of performing OCR. For example, in regard to an application containing a significant amount of redundant data, such as form-type data, numerous companies today are working on various kinds of form renewal, prior to going through a recognition process, by using dropout colors. Thus, acquiring access to form design or using electronic means are worth considering. HOOTON also noted that conversion usually makes or breaks one's imaging system. It is extremely important, extremely costly in terms of either capital investment or service, and determines the quality of the remainder of one's system, because it determines the character of the raw material used by the system.

Concerning the four projects undertaken by NAL, two inside and two performed by outside contractors, ZIDAR revealed that an in-house service bureau executed the first at a cost between $8 and $10 per page for everything, including building of the database. The project undertaken by the Consultative Group on International Agricultural Research (CGIAR) cost approximately $10 per page for the conversion, plus some expenses for the software and building of the database. The Acid Rain Project--a two-disk set produced by the University of Vermont, consisting of Canadian publications on acid rain--cost $6.70 per page for everything, including keying of the text, which was double keyed, scanning of the images, and building of the database. The in-house project offered considerable ease of convenience and greater control of the process. On the other hand, the service bureaus know their job and perform it expeditiously, because they have more people.

As a useful comparison, ERWAY revealed AM's costs as follows: $0.75 cents to $0.85 cents per thousand characters, with an average page containing 2,700 characters. Requirements for coding and imaging increase the costs. Thus, conversion of the text, including the coding, costs approximately $3 per page. (This figure does not include the imaging and database-building included in the NAL costs.) AM also enjoyed a happy experience with Federal Prison Industries, which precluded the necessity of going through the request-for-proposal process to award a contract, because it is another government agency. The prisoners performed AM's rekeying just as well as other service bureaus and proved handy as well. AM shipped them the books, which they would photocopy on a book-edge scanner. They would perform the markup on photocopies, return the books as soon as they were done with them, perform the keying, and return the material to AM on WORM disks.

ZIDAR detailed the elements that const.i.tute the previously noted cost of approximately $7 per page. Most significant is the editing, correction of errors, and spell-checkings, which though they may sound easy to perform require, in fact, a great deal of time. Reformatting text also takes a while, but a significant amount of NAL's expenses are for equipment, which was extremely expensive when purchased because it was one of the few systems on the market. The costs of equipment are being amortized over five years but are still quite high, nearly $2,000 per month.

HOCKEY raised a general question concerning OCR and the amount of editing required (substantial in her experience) to generate the kind of structured markup necessary for manipulating the text on the computer or loading it into any retrieval system. She wondered if the speakers could extend the previous question about the cost-benefit of adding or exerting structured markup. ERWAY noted that several OCR systems retain italics, bolding, and other spatial formatting. While the material may not be in the format desired, these systems possess the ability to remove the original materials quickly from the hands of the people performing the conversion, as well as to retain that information so that users can work with it. HOCKEY rejoined that the current thinking on markup is that one should not say that something is italic or bold so much as why it is that way. To be sure, one needs to know that something was italicized, but how can one get from one to the other? One can map from the structure to the typographic representation.

FLEISCHHAUER suggested that, given the 100 million items the Library holds, it may not be possible for LC to do more than report that a thing was in italics as opposed to why it was italics, although that may be desirable in some contexts. Promising to talk a bit during the afternoon session about several experiments OCLC performed on automatic recognition of doc.u.ment elements, and which they hoped to extend, WEIBEL said that in fact one can recognize the major elements of a doc.u.ment with a fairly high degree of reliability, at least as good as OCR. STEVENS drew a useful distinction between standard, generalized markup (i.e., defining for a doc.u.ment-type definition the structure of the doc.u.ment), and what he termed a style sheet, which had to do with italics, bolding, and other forms of emphasis. Thus, two different components are at work, one being the structure of the doc.u.ment itself (its logic), and the other being its representation when it is put on the screen or printed.

SESSION V. APPROACHES TO PREPARING ELECTRONIC TEXTS

+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ HOCKEY * Text in ASCII and the representation of electronic text versus an image * The need to look at ways of using markup to a.s.sist retrieval *

The need for an encoding format that will be reusable and multifunctional +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++

Susan HOCKEY, director, Center for Electronic Texts in the Humanities (CETH), Rutgers and Princeton Universities, announced that one talk (WEIBEL's) was moved into this session from the morning and that David Packard was unable to attend. The session would attempt to focus more on what one can do with a text in ASCII and the representation of electronic text rather than just an image, what one can do with a computer that cannot be done with a book or an image. It would be argued that one can do much more than just read a text, and from that starting point one can use markup and methods of preparing the text to take full advantage of the capability of the computer. That would lead to a discussion of what the European Community calls REUSABILITY, what may better be termed DURABILITY, that is, how to prepare or make a text that will last a long time and that can be used for as many applications as possible, which would lead to issues of improving intellectual access.

HOCKEY urged the need to look at ways of using markup to facilitate retrieval, not just for referencing or to help locate an item that is retrieved, but also to put markup tags in a text to help retrieve the thing sought either with linguistic tagging or interpretation. HOCKEY also argued that little advancement had occurred in the software tools currently available for retrieving and searching text.

She pressed the desideratum of going beyond Boolean searches and performing more sophisticated searching, which the insertion of more markup in the text would facilitate. Thinking about electronic texts as opposed to images means considering material that will never appear in print form, or print will not be its primary form, that is, material which only appears in electronic form.

HOCKEY alluded to the history and the need for markup and tagging and electronic text, which was developed through the use of computers in the humanities; as MICHELSON had observed, Father Busa had started in 1949 to prepare the first-ever text on the computer.