An Experimental Computerized CBAC Search Project
STEPHEN R. HELLER.* RICHARD J. FELDMANN, and KENNETH P. SHAPIRO
Heuristics Laboratory, Computer Center Branch and Data Management Branch.
Division of Computer Research and Technology, National Institutes of Health.
Department of Health. Education and Welfare. Bethesda, Md. 20014
Received September 17, 1971
A literature search system for CBAC is described. Both SDI and retrospective searching techniques are available to the user. The system relies heavily on searching keywords from the CBAC text rasher then the text itself and employs batch/tape and time-sharing/ disk hardware.
An experimental Chemical Information System, consisting of the literature, substructure search, and property files is being tested at the Division of Computer Research and Technology of the National Institutes of Health.(1). One part of this experimental system, the literature project, is the topic of this paper.
THE DATA BASE
The major problem that confronts a new information center is what data base to use. Some data bases cover a limited area, some are broad, some contain only titles, some contain full abstracts, some are highly indexed with complex dictionary and index terms, some provide rapid literature indexing and abstracting, and some are slower but more thorough. One, or even two, data bases will not cover the needs of the broad biochemical/medical research complex at NIH .MEDLARS is already used extensively at NIH, but is highly medically oriented.
Among all the relevant sources of chemical/biological data bases, Chemical Abstracts Service has the largest computer-readable files as well as plans for greatly expanding its services in computer-readable data bases. Chemical-Biological Activities (CBAC), one of the journals published by Chemical Abstracts Service, was chosen as the first experimental data base. While CBAC does not cover as many journals as one might prefer, it does cover a wide range of chemical and biochemical, and some clinical, journals. Its outstanding features from the user viewpoint appeared to be the availability of the full abstract on computer tape and the large number of searchable data elements which include author, title, location of work, CODEN, abstract, molecular formula and REGN (Chemical Abstracts Registry Number).
THE SEARCH PROGRAMS
Computer text-searching programs run either in batch mode or are interactive. The frequency of appearance of CBAC (fortnightly) makes it reasonable to do current awareness or SDI (Selective Dissemination of Information) searching in the batch mode.
Retrospective searching is a different matter. If a scientist is writing a review or needs to find references for a broad study, a batch search is probably quite acceptable since a good deal of extraneous information can be tolerated (or may even be desirable to be sure of complete coverage). If he wants to explore a few references to a particular research topic, however, a narrow request is probably served best by interactive searching.
Another valuable use of an interactive retrospective search system is to establish a properly worded SDI profile for the particular data base. This is especially true for an open-vocabulary system such as CBAC.
At DCRT, experiments are being undertaken with three search systems: Batch SDI, Batch Retrospective, and Interactive Retrospective.
Batch SDI. Through the generosity of the National Science Library/National Research Council of Canada, DCRT has obtained and substantially modified the CAN/ SDI system. The CAN/SDI system consists of four programs: reformatting, profile generation, searching, and printing.
The first and last parts have been completely rewritten and considerable changes have been made to the middle two.
While the CAN/SDI reformatting program would leave a given data base essentially in the original form, the DCRT modification using the CBAC data base makes a very substantial change and thus it may be preferable to call this first program a reprocessing program.
The CBAC tapes are read in and one record is built for each abstract (originally CBAC had about 40 to 50 records) which consists of tags of data and pointers to tines. tags within the record. For example, the tag for REGN contains a list of all the registry numbers in this abstract separated by delimiters. The advantage of this is that when searching for REGN, one need only look at the tag containing the REGN and not at any other part of the record. The same is true, of course, for the other data tags in the abstract record (keywords, authors, etc.).
In the reprocessing, two operations are performed on the title and abstract sections of CBAC. First a word is read in and if it is a "trash" word (words that are not of value for searching purposes), it is thrown out. If it is not a trash word, then a check is made for duplication in that abstract
The title Digest section of CBAC is converted into keywords. The 140 words in the above aigest and title are passed against the "trash" list and duplicates removed. the following 19 keywords ret'ain and appear on the computer tape as follows:
MEOlANISM
ACTION
DCYCLOSERINE
TRANSPORT
SYSTEMS
DALANINE
LALA\i'INE
GLYCINE
ACCUMULATION
ESCHERICHIA
COLI
TRANSPORTED
DALANINEGLYCINE
SYSTEM
LINElVEAVERBURK
EFFECTIVE
INHIBITOR
ANTAGONIZED
EFFECT
Note that all punctuation marks are removed.
Figure. 1. CBAC abstract before and after reprocessing step
and if duplication is found, then the duplicate word is not put into the keyword tag field. This keywording of titles and abstracts reduces the size of the field being searched to 15 to 45% of its original size (Figure 1).
The profile generation program is simply a program that "compiles" profiles for the search program. The program checks for syntax and is quite simple to use. The modifications in this program consist of adding two new data elements that can be searched--namely, REGN and molecular formula--and allowing for a wider variety in the number of hits or references to be retrieved for a given search. Normally from one to ninety-nine references can be requested for each profile, but special codes allow the user to get additional references. The program is accessible through a terminal-based text editor system (WYLBUR) and this allows users to code and check their own profiles if they so desire at any of approximately 250 remote terminals serviced by the DCRT/Computer Center Branch of the National Institutes of Health. Normally, one can get 5- to 15-minute turn-around time with this program, and thus it is very convenient to use.
Figure 2. Print program output for the abstract in Figure 1
To aid the user in profile formulation, a Profile Design Manuals was written. The 80-page manual, modelled after one written for the CAN/SDI project, contains seven sections. Included in the manual are a description of CBAC, numerous and varied examples of coded profiles, detailed instructions on coding, the "trash" list, the list of journals covered by CBAC, and sample output.
As a further aid to the user, listings of the REGN and corresponding names or molecular formula from CBAC were made available. Of the approximately 54300 unique REGN, about 3500 entries contained the same name with different REGN. The computer program used to pull this list from the data base assumed an error free format. In this case and in many instances in the reprocessing/ reformatting of the six years data base, a great variety of errors were encountered. However, it was very pleasing to find the errors fell off sharply as a function of time between 1965 and 1970 CBAC tapes.
There are seven types of terms that can be searched: titles, authors, location of work, CODEN, REGN, Molecular Formula, and keywords. The last is the section that DCRT adds to the data base. Authors and keywords can be right truncated. For economic reasons only titles terms can be left truncated. CODEN and REGN of course, are not truncated. The logic is AND, OR, and NOT.
It is worth noting that of seven term types, keywords account for 95% of the terms used, followed by REGN 3%, all others 2%.
The third program is the search program, and this was modified to search for REGN and molecular formula in the reformatted tapes. The search program has been modified so that it no longer searches for phrases in the abstract section, but rather a keyword or a combination of keywords. For example, to search for "nuclear magnetic resonance" the user must request "nuclear' and "magnetic' and "resonance." Clearly, combinations out of context are a distinct possibility although it has been surprising that this has not yet been found to have occurred.
The last program is the print phase which prints the abstract essentially as it appears in the hard copy printed version of CBAC (Figure 2).
Batch Retrospective. The same programs and tapes are used for SDI and batch retrospective searching. The reprocessing of the tapes and modifications to the other SDI program enable us to reduce very substantially the cost of retrospective searching of CBAC. In 10.9 CPU minutes on an IBM 370/165, we are able to search an approximately 89,000 document CBAC file for 77 profiles. Because of the efficiency of the batch sequential search in our multi-programming 360/370 system, we feel that development of an inverted file search system on the 360/ 370 containing full abstracts is unnecessary.
Interactive Retrospective. In addition to the batch 360/370 program previously described, we are experimenting with a streamlined CBAC search system which is interactive and is accessible via teletype terminals from a timeshared DEC PDP-10 computer at DCRT.
The streamlined Interactive Retrospective CBAC search uses an inverted file created by reading in the title, digest author, and REGN sections of the CBAC tapes on the PDP10, eliminating "trash" words and duplicates as in the 360 reprocessing, collecting all keywords, author names and REGN along with their pointers to the CBAC volume, issue, and abstract numbers that they came from. The file is then sorted, each word on the sorted list is reduced to a 36-bit code by a technique called hash coding. The hash coding involves the use of a 3rd degree polynomial equation based on Fermat s last theorem. This hash word, along with the set of pointers for the word, are stored on disk packs. The search program simply asks the user to type in a subject word, author name, or REGN. The program then hashes the input word, compares the hashed word to those stored on the disk and, usually within 1 to 2 seconds, replies to the user that either the word is not in the file or that there are N references to that word. The user may then request the references immediately or type in further word(s) to use for further searching. The program will automatically search for the Boolean AND conjunctions for all terms typed in.
CHEMICAL SUBSTRUCTURE INDEX--A NEW RESEARCH TOOL
An example of the POP- 10 program input and output is:
Type word for search specification
DNA
Found 175 references to word DNA
References Y/N?
N
Type word for search specification
RNA
Found 133 references to word RNA
References Y/N?
N
Reference conjunction
Conj No. |
No. Abstracts |
Words |
1 | 20 | DNA, RNA |
2 | 133 | DNA |
3 | 175 | RNA |
Thus, in a matter of minutes of connect time and some 2 to 3 seconds of CPU time, the user identified 20 references that contain DNA and RNA. Even further refinement would have been possible if the user wished to narrow his question. It is important to note that only the CBAC digest number is available here as output, as opposed to the full abstract, title, author, etc.
This program has been very helpful for generating SDI profiles or retrospective search requests for the 360/370 programs because the user can quickly check the existence in the system of a variety of words and their possible synonyms (necessary because CBAC is free text and does not have rigidly limited vocabulary).
USER REACTION TO A BATCH
SDI SEARCH SYSTEM
Preliminary studies indicate that the users agree on several points:
The PDP-10 search is impressive, it is useful in its own right and is a strong inducement to try the 360/370 SDI and retrospective searches. Its main flaw is no abstract.
The full abstract (available via the 360/370 programs) is the best part of CBAC. Users would be pleased if all of CA (and BA) contained full abstracts on tape. (Indeed some even said they would not consider using CA condensates or BA as they currently exist.)
Computer searching of data bases is valuable, but CBAC covers too narrow a range.
A number of users have expressed surprise to find journal articles retrieved that they had not seen previously.
CONCLUSIONS
The testing of CBAC is still in an experimental stage with a limited number of users. It is too early to determine if it will go into production status and, if so, whether it will need further refinements to achieve an optimal balance of the economic and informational needs of the NIH user community.
With the expansion of the size of the CBAC data base beginning in January 1972 (by about 40%), the range of coverage will obviously become greater and should please users. However, it remains to be seen if the intellectual value of increased coverage will offset the finanical cost of the larger data base.
Quite clearly when CA with complete abstracts are available in computer readable form, a searchable keyword section would have great economic value.
ACKNOWLEDGM ENT
The authors thank Inez Gaffney and James Heilik of the National Science Library, National Research Council, for their assistance. We also thank Keatha K. Krueger, Victoria Ragen, and Billie Mackey for their aid in the Pilot CBAC Study.
LITERATURE CITED
(l) Feldmann, R. J., Heller, S. R., Shapiro, K. P., and Heller, R.S., "An Application of Interactive Computing--A Chemical Information System, "J. Chem. Doc., to be published.
(2) Heller, S. R., "Profile Design Manual," DCRT/CIS, CBAC Literature Project.