Better Things for Better Chemistry Through Multi-Media

Michael Lesk

Dennis Egan

Dan Ketchum

Carol Lochbaum

445 South St
Morristown, NJ 07960-1910


Will chemists prefer online delivery to traditional printed paper for primary journal information? Will online delivery let them do what they need faster and better than paper? Surprisingly, for all the attention given to machine-readable text, there is relatively little hard evidence that it is easier or quicker to use than traditional paper. In our experiments, we have compared paper with both image display and Ascii display, to see what features of systems will produce improved performance. We find better performance on searching tasks for both kinds of display than with paper, and comparable performance on reading tasks. We find that the chemists are heavily graphic-driven; quick delivery of graphical information is perhaps more important than manipulation of the text.

1. Introduction

In the CORE project (Chemistry On-line Retrieval Experiment), Cornell University is experimenting along with Bellcore, the American Chemical Society, Chemical Abstracts Service, OCLC, and University College London with the provision of primary journal material in electronic form [Lesk 1991]. In the course of this project we have been accumulating and displaying to chemists the contents of the American Chemical Society primary journals, a group of twenty of the most important journals for chemical researchers. The material is available both as typesetting tapes, and thus in Ascii, and also as scanned page images. Thus, we have the opportunity to compare the relative advantages of both image and text display for chemical material.

In past experiments, it has been surprisingly common for hypertext systems to be no more effective than paper when measured in actual experiments. Such results can be found in [Gordon 1988], [McKnight 1989], [Marchionini 1988], and [Kaindl 1992].

2. Data Preparation

The goal of the data preparation stage of the project is to arrive at two different computer representations of each article. One is the scanned images of each page of the article; the other is the combination of the marked-up Ascii text of the article and its tables, with bitmap images of the equations, figures, and other graphics.

The bitmap images of each page of the article are obtained by scanning at 300 dpi either the printed pages (for the articles published in 1991 or later), or the microfilm of the back issues (roughly 1980-1990). The scanning of the paper is done at Bellcore on an Improvision sheetfed scanner connected to a Sun-3 at about 20 pages per minute; the scanning of the microfilm is done either at Bellcore on a Mekel microfilm digitizer, at about 30 pages per minute, or by one of two outside contractors (STS Information Systems or Zuma Imaging). The microfilm is 16mm blipped cartridges (both vendors have commented on the high quality of the film).

The Ascii text is more complex to obtain. Typesetting tapes in a proprietary format are received from ACS and CAS, containing respectively the data used to print the primary journals and the related parts of Chemical Abstracts. The data is in a fielded format, which is converted to an SGML stream format. The SGML tags used are an extension of AAP Electronic Manuscript Format, enhanced to include the special items such as CAS Registry numbers that are included in the chemistry documents. The character set also must be expanded beyond what AAP provides; ISO codes are used where possible, but additional symbols are necessary. For display purposes, a set of bitmaps suitable for use as X window system fonts are created.

With the intent of providing hypertext-like links between articles on the basis of citations, some special marking of the text is done. Each reference is parsed into journal, volume, and page. If it appears to be a reference to an article in one of our twenty journals, a special code is inserted uniquely identifying the article (consisting of the journal, the volume, the page number, and the first two letters of the senior author's name). The same code is placed at the head of each article. In this way, any full text search system can retrieve the actual articles easily.

The typesetting tapes contain only the textual components of the articles. They must be supplemented with the figures that were printed with the original articles. These are found by analyzing the bitmaps of the scanned pages and looking for graphical material. This process has been described before [Lesk 1990]. and operates briefly as follows. The items being sought are both ``Figures'' and ``Schemes.'' ``Figures'' have numbers and captions; ``Schemes'' do not. They must be sorted out because each is referred to separately in the text; that is, a reference to Scheme 1 may appear before a reference to Figure 1, but the actual graphic for Figure 1 may appear earlier on the page (since, on most pages, the figures appear at the top of the page, while the schemes are in the middle of the columns).

  1. The page image is analyzed to determine whether it is tilted or not; sometimes the page, as fed through the scanner, was angled relative to the rollers and scanning mechanism. This is done by finding the left edge of each line on the page, and attempting to push a straight line against the dots this represents. The slope of that straight line is the skew of the page. It is not adequate to simply fit a straight line to the dots, because many pages contain figures and thus have a ragged left margin.
  2. The page is then rotated to bring it as exactly upright as possible. This is done with the conventional shifting algorithms rather than by full trigonometry.
  3. The page is then sorted into columns. This is done by looking for rivers of white space which run from top to bottom of the page. The standard format is double column, with full-width material used for wide figures and tables, and with titles and abstracts centered across the entire page.
  4. Each column is then examined for the precise bitmap of the word ``Figure.'' This is the most reliable way to distinguish figures from schemes. Since all the articles in each journal are typeset to the same format, it is possible to simply do an exact-match search for the right bit sequence.
  5. The number of dark bits on each horizontal scan line of the image are counted. This function is then autocorrelated at each possible shift and the first peak in the result is taken as the line spacing (that is, when shifting the image down by some number of pixels causes the dark lines to fall on top of each other, that's a multiple of the line spacing; the first such pixel count is the line spacing itself). This calculation is done on the bottom third of the column only (since the top is likely to be full of figures and distorted).
  6. The autocorrelation at a shift corresponding to the linespacing is then calculated for each area of the column, using a sliding window. This function is then examined: it is large if there is text, and small if there is a graphic. Potential graphics are then marked.
  7. The list of potential graphics is then improved. Tables, for example, may look like graphics, being very thinly printed. However, in ACS style they have several horizontal lines and can be rejected on that basis. Equations are also sparsely printed; they are rejected as being mostly centered, but with a label at the right margin. The list of instances of the word ``Figure'' is then used to decide which graphics are figures and which schemes.

The output of this process is a list of graphical items for each page; it takes about one minute per page on a DEC 5000 to do this. The lists for each page of an article are combined and compared with the list of references to graphics found in the article, and an attempt made to assign the numbers to particular figures and schemes. The article may also refer to tables and equations. The tables can be displayed normally and are taken from the typesetting tape. To simplify the job of the display program, each table is given twice in the Ascii file: once with each cell identified with the AAP tags, and once with the table formatted for constant-width fonts and ready to toss on the display.

Equations are more complex. Again, it is desired to simplify the work of the display program, and so the equations are displayed as bitmaps rather than as a complex typesetting problem. To get these bitmaps, the actual input equation (which is available on the typesetting tape) is converted to eqn, then Postscript, and then to bitmap form. All bitmaps are stored in Tiff Group IV; figures and page images are stored with one image per file, but equations are stored with multiple images per file, one file per article. This is to minimize overhead, since there tend to be many equations in each article which has any, all of them small (in the Journal of Physical Chemistry, for example, the average article with equations has a dozen equations, and a few have over 100 equations; but a fifth of the articles even in this journal have no equations at all).

When all is said and done, the average article is about 7 pages long, and the images of the pages in the article run to 500KB. It contains about 5 figures, and the figures take about 50KB per article. The text of the article is about 40Kbytes, so the text plus images is about 20% of the size of the full page images. The articles range widely in length: some very short errata may be only a few hundred bytes, while a single long article in Chemical Reviews can be over 500KB.

Unfortunately, there are two other large storage demands. First, the text must be processed for full-text retrieval. Depending on what kind of package is used for this purpose, an amount of space ranging from a quarter of the original text size to something comparable to the original text size will be required for the index files. More important, the low resolution of computer screens compared to paper makes it valuable to have alternate resolution versions of each page. In principle, these images could be computed on the fly, but at present that is too slow and so we store alternate versions of each page. The low resolution version presents a number of temporarily painful choices:

  1. People with large screens may be able to see an entire page at a resolution equivalent to 100dpi, while people with somewhat smaller screens will prefer 75 dpi, even though this is harder to read;
  2. Grey-scale can be used via anti-aliasing to improve readability, but more space is needed;
  3. Space can be saved by efficient compression algorithms, but response time is also important.

Some of these problems are going away as processors and algorithms for image decompression improve. On the latest processors, just becoming available, it is possible to decompress the full size images in 1 second, and thus we should be able to go to on-the-fly creation of the reduced-size images, and tailor them to the user's screen.

Finally, the cost of magnetic storage has meant that at both sites, Cornell and University College London, jukeboxes are being used to store the actual images. At Cornell a WORM jukebox is used; at University College London a magneto-optical jukebox is in use. This introduces additional delays in obtaining images, although not too serious at present.

3. Interfaces

The purpose of this project is to present alternate ways of accessing chemical information. Several different interfaces have been built. For the purposes of this paper most of the effort will be spent describing the interfaces which were used in a formal experiment. Three ways of getting at information were presented to the users.

  1. One possibility is the original paper format of the journals. This form is easy to read, familiar to the users, but has very few aids for searching. For searching users must rely on Chemical Abstracts, which can also be provided on paper. It is, however, a bulky publication which requires some familiarity and study to use most effectively.
  2. The users can look at the scanned images of the pages. This resembles the printed pages, of course, so that the appearance is familiar. However, the resolution is not as good as on paper, and various constraints would seem to make reading less comfortable: the computer screen can't be moved, the illumination and contrast may be less comfortable, and the user must choose between a poor-resolution image in which a whole page can be seen at once and a better quality image in which only a bit of the page can be seen at a time.
  3. The users can see an Ascii display of the material. In principle, this should subsume the previous version, since in theory one could rerun the original typesetting software and produce bitmaps matching the original pages (something like this is done for the equations). However, we would think that a reformatting on the spot could do better than just reproduce the original page, since it could be sensitive to window size and to the context, for example by highlighting words that matched query terms.

Although we have several different interfaces, e.g. the Sceptre interface from OCLC, and the WAIS interface from Thinking Machines, we ran experiments using two particular interfaces, the SuperBook interface for Ascii and a simplistic interface named Pixlook for the images. The interfaces differ in both search engine and in display form, but the primary differences are in the display format. Searching, in both cases, relies on the full text of the articles.

The major difference in search capability is that SuperBook searches for multiple terms as co-occurrences within a paragraph, while Pixlook searches for co-occurrences within an article. For some queries, e.g. liquid crystal, with two common words, paragraph co-occurrence produces a much more manageable list of hits. Other differences include the ability of Pixlook to support fielded searching on eight fields of the article (title, author, author affiliation, abstract, Chemical Abstracts indexing, text, references and captions) while SuperBook has been arranged to support fielded searching only on the title, author, and reference fields. However, searching on the other fields was not common in practice. Pixlook also supports the Boolean searching operations ``and'' and ``or,'' but again these are lightly used, with the alternative coordinate indexing searching being more common.

Samples of Pixlook operation are shown in Figures 1-3. Figure 1 shows a simple search plus the list of matching titles; Figure 2 shows a sample page, and Figure 3 then shows a sample page image with a ``zoomed'' window enlarging part of the text. Note that the Pixlook interface is basically pretty routine: the displays will normally look just like the printed pages.

Similarly, some SuperBook examples are shown in Figures 4-6. Figure 4 shows a simple search on ``porphyrin structure'', with highlighted text words; note the figure and footnote symbols in the right margin. Figure 5 then shows the footnote popped up, and Figure 6 the figure popped up. Note that the focus is on the text. Key aspects of the SuperBook interface include:

  1. A window to type searches, which are taken either as single word searches or as requests for co-occurrences in the same paragraph.
  2. A structured table of contents, arranged in hierarchical format. In our context this normally follows the hierarchy of Chemical Abstracts, which groups articles into 80 areas; and within each article, according to the section headings of the article. When a search is done, the number of hits is displayed for each item shown in the table of contents. It is possible to use other hierarchical arrangements in the table of contents; for example, we have also used a breakdown by journal and year of publication.
  3. A text window showing part of the text, with matching terms highlighted, and with icons to indicate the presence of footnotes, tables, or graphics.
  4. Pop-up windows to display the extra items \- footnotes, tables, schemes, figures or equations.
  5. Control buttons to change the text being read, or to enter or leave SuperBook.

SuperBook has been developed and tested on a wide variety of documents, although originally designed for textbooks. Unlike normal hypertext systems, it always maintains the text as a linear sequence, and so the user always can have a sense of position in the document. This position is shown both in the table of contents window and at the top of the text window, where the hierarchical headings are shown leading down to the place in the text represented in the window. Thus, SuperBook is not as well adapted as it might be to this collection, which consists of a large number of individual articles with weak relationships between them. As a result, even though it performed fairly well in our experiments, it can probably be tuned to do better on this particular collection.

As mentioned, other interfaces are also available. Figure 7 shows the WAIS interface (from Brewster Kahle of Thinking Machines) which can display both the Ascii text (but with no structure) or the full page images. All the interfaces run on the X windows system and can be accessed remotely over the Internet.

4. Experiments

We evaluated the different interfaces by experiments carried out by chemistry students at Cornell University. The experiments were based on a set of 1000 articles (the contents of Journal of the American Chemical Society for the first twelve issues of 1988). Three dozen students were divided into three groups: 12 were given the journal issues on paper; 12 were given Pixlook; and 12 had SuperBook. All of these groups were asked to do five different kinds of tasks, made by two chemistry professors. They were all given roughly half an hour of training, and then asked to do tasks for three sessions of two hours each.

The five different kinds of tasks represent a spectrum of difficulty. The table below shows the tasks.
CitationFind a particular fact in a known article
SearchFind a particular fact in an unknown article
BrowsingScan an issue for any of 8 topics.
EssayWrite paragraph on a topic.
TransformationFind analogy to a structural transformation.
Note that for the citation and browsing tasks, the student knew what to read. A citation task, for example, might be something like .DS In the article `Total Synthesis of Ginkgolide B' by E. J. Corey, M. Kang, M. C. Desai, A. R. Ghosh and I. N. Houpis, JACS v. 110, p. 649-651, what is reported as a medically important property of ginkgolide? .DE while by contrast a search task might be .DS What is the calculated P-O bond distance in hydroxyphosphine? .DE For browsing, the student was given an issue of the journal, and told to find out whether there was anything of interest on a list of 8 topics, each something like ``bridgehead halides'' or ``phospholipids.'' In each case four topics did appear in the issue and four did not; the student was scored on how many were correctly identified. The essay question involved writing a paragraph or two on a topic such as ``vibrational spectroscopy'' and was scored by graders both on the quality of the essay and on the quality of the list of references supplied with the essay. Finally, the analogous transformation question consisted of two chemical structures, and a request to explain how to turn the first into the second. Neither structure appeared in the file, but some analogy to the transformation does appear.

The details of the results have been given by [Egan 1991], so only a summary will be presented here.

  1. On the tasks where the students did not know what to read in advance, those using the computer systems did a great deal better. For example, on the search task, the majority of the students with the paper journals gave up and did not answer the question in more than half the trials. Fewer than 10% of the students with computer searching gave up.
  2. On the tasks where the students were told what to read (the citation task and the browsing task), paper and computer systems performed comparably. For example, in looking up articles and finding a fact in them, it took about five minutes to read the article and find the fact in any mode. The advantage of being able to highlight search terms did not seem to give the users of the SuperBook display any advantage. The students were all working on a 1152x900 bitmap screen; this apparently provides an adequate resolution level for straightforward reading tasks.
  3. People with computer searching available found more to read. For example, on the browsing task, although the overall performance rate was comparable, the errors made by those with paper were largely recall errors (failure to find relevant articles) while the errors made by students with the computer systems were largely precision errors (calling documents relevant which are not relevant).
  4. Availability of graphical information was important. On the analogous transformation task, for example, a number of users of SuperBook found the correct article but did not find the answer because they did not pop up the diagram to look at. In interviews the chemists also talked about the importance of structural diagrams and other visual cues.
  5. Although people would of course like their data to arrive as quickly as possible, they spent more than a minute reading each page once they had it. Thus, the bitrate demand from such a system is likely to be bursty \- some periods of high activity delivering lists of hits or pages, then low activity while the article is being studied. This suggests online availability of the data on a network, rather than a need to have duplicated copies of the database for each user.

In general, there was no trade-off between speed and accuracy. Errors came from difficulty of doing the task, not from rushing through it. Students were still learning how to do some tasks, such as the browsing task, on the third session of the experiment (i.e. after more than 4 hours of experience).

We also gathered the comments of the students using the system, who generally liked it but had many specific suggestions, often oriented towards making the systems easier to use. For example, requiring the ampersand instead of the word AND for specifying Boolean combinations did not seem friendly. It is difficult to scan long articles in image format, and some way of indicating where specific words can be found would be desirable (this is also quite difficult, requiring guessing on the basis of the article length).

We also talked to a number of chemists who were trying experimental versions of the system. Perhaps because these were the most motivated of the chemists, they tended to be very interested in information retrieval systems and thus to be familiar with the online systems that already exist. Their comments were less focussed on useability and more on functionality. They asked, for example, for the ability to save searches, and to search on CA substance index terms. They also asked for some new facilities, for example the ability to quickly move from an article to an article that it cites or that cites it. And they wanted better ways of downloading information so that they could assemble their own databases from tables in the articles they found.

Observation of the actual performance of the students on the tasks also leads to suggestions for improving the systems. For example, the students got into trouble doing the citation task with SuperBook when they wandered out of the article they were supposed to be reading. Unlike a paper journal, it is possible in the computer files to jump a thousand pages forward without realizing it. We can and have modified the interface to make it more obvious which article is being read, so that future users will know what to do.

Critics of electronic information delivery sometimes mention the importance of serendipity, so we tried to measure it. The students doing the browsing task had been given an issue of the journal, and a list of eight topics, and asked to list the ones which appeared in the issue. The students could either simply flip through the issue (as most students with the paper copy did) or try to search for the eight topics (as most students with SuperBook did). When they were done, the issue was taken away, and then they were given another list of topics to judge, testing their memory of what they noticed about the issue besides what they were looking for. The average performance on this second task was 60-65% correct, with no significant difference between those with paper and those with the electronic forms. Thus, serendipity was as high with computer access to material as with paper, refuting the fears of many about electronic libraries. Note that in no case is this performance very good: since the students would have achieved 4/8 right guessing at random, reaching 5/8 is not great (although it is statistically significant).

Perhaps most interesting is the importance of graphics to the various chemists. It is possible that the interface they would like most would be one in which the articles were represented by images, with the text being hidden under icons. For example, we put together such a search system as shown in Figure 8, which we referred to as the ``comic book'' version of the Journal of the American Chemical Society. Although the difficulty of searching pictures makes entirely image-based interfaces hard to build, we should look more carefully at interface designs that rely on presenting large amounts of visual material very rapidly, rather than slow, precise access to particular sentences. Most information needs that scholars have are relatively vague, and some combination of browsing and searching is more typical than the very focussed searching that information systems are often designed to meet.


Chemists use journal articles in many ways, from finding the address of an author to just looking for something interesting. Some chemists like to read words; others like to look at structural diagrams; some probably even like reading mathematical equations. Any system which serves them, whether based on paper or on electronics, must cater to a very wide spectrum of needs. The lessons we think we have learned, and are trying to apply in our software design, are:

  1. Some sense of context must be retained. Simply putting somebody into a gigabyte of text and letting them wander around is not as good as letting someone have a guide to where they are in a collection.
  2. Rapid presentation of text and pictures so that people can skim is important. People are good at looking quickly at a large amount of material.
  3. Searching without computers is much inferior to searching by computer.
  4. Graphical content is important. Text-only systems are not likely to satisfy chemists.

We were not able to settle whether image or ascii interfaces are preferred. The ascii text is needed for doing searching, but it seems that people can read the image text at comparable rates; apparently the familiarity of the format and the immediate visibility of the graphics compensates for the inadequate resolution of the computer screens.

But the main message is that performance overall was as good reading on screen as on paper, and better searching on screen. Either image or ascii display will do, but computer systems can compete with paper for performing chemical tasks.


This project is a collaboration led by Cornell University and including, in addition to Bellcore, the American Chemical Society, Chemical Abstracts Service, OCLC and University College London. Among the people who have done much of the work on the overall project are Jan Olsen, Howard Curtis, and Rich Entlich of the Albert Mann Library, Cornell University; Lorrin Garson of the American Chemical Society; Jim Lundeen and Lorraine Normore of Chemical Abstracts Service; Martin Dillon, Martha Lindeman, Eric Sigman, Mark Bendig and J. Charles Crabb of OCLC; and David Golds and Peter Kirstein of University College London. The collaborators are grateful for the support of Apple Computers, Digital Equipment Corporation, Springer-Verlag, Sun Microsystems Inc., and Sony Corporation of America.

[Egan 1991]. D.E. Egan, M.D. Lesk, R.D. Ketchum, C.C. Lochbaum, J.R. Remde, M. Littman, and T.K. Landauer; "Hypertext for the electronic library? CORE Sample Results," Proceedings of Hypertext 91, San Antonio, Texas (1991).

[Gordon 1988]. S. Gordon, J. Gustavel, J. Moore, and J. Haney; "The effects of hypertext on reader knowledge representation," Proc. Human Factors Society 32nd Annual Meeting pp. 296-300, Santa Monica (1988).

[Kaindl 1992]. Hermann Kaindl, and Holger Ziegler; "HIS \- An Information System about Hypertext on Hypertext," SIGLINK Newsletter 1 pp. 1-8 (March 1992).

[Lesk 1990.]. M. E. Lesk; "Full Text Retrieval with Graphics," NATO AGARD Conference Preprint No. 487, ``Bridging the Communication Gap'' pp. 5-1 to 5-7, Trondheim, Norway (Sept. 5-6, 1990.).

[Lesk 1991]. M. E. Lesk; "The CORE Electronic Chemistry Library," Proc. 14th ACM SIGIR Conference pp. 93-112, Chicago (October 13-16, 1991).

[Marchionini 1988]. G. Marchionini, Marchionini88, and B. Schneiderman; "Finding facts vs. browsing knowledge in hypertext systems," IEEE Computer 21 pp. 70-80 (1988).

[McKnight ]. C. McKnight, A. Dillon, and J. Richardson; "A comparison of linear and hypertext formats in information retrieval," Proc. Hypertext II, York, England