RIAO '88 TRIP REPORT

• Document Type: CREST
• Collection: General CIA Records
• Document Number (FOIA) /ESDN (CREST): CIA-RDP93B01194R001700050003-9
• Release Decision: RIFPUB
• Original Classification: K
• Document Page Count: 7
• Document Creation Date: December 27, 2016
• Document Release Date: February 13, 2013
• Sequence Number: 3
• Publication Date: April 27, 1988
• Content Type: MEMO

Declassified and Approved For Release 2013/02/13: CIA-RDP93B01194R001700050003-9
27 April 1988
SUBJECT: RIAO '88 Trip Report
On March 21-24, 1988 a conference was held at M.I.T. on the
theme "User-oriented Content-Based Text and Image Handling." The
.RIAO is an international conference organized by the Centre de
Hautes Etudes Internationales d'Informatique Documentaire. It was
the first conference to be devoted entirely to the field of
Information Retrieval and the response was overwhelming. The 700
attendees were more than double what the sponsor expected, causing
a last minute shift to larger rooms for the presentations.
Attached is a complete list of all the papers delivered at the
conference and each attendee was given 3 bound volumes containing
the complete text of all the papers that were presented. The
bibliographies alone (listed for each paper) are extremely
valuable sources for the history and state-of-the-art of this
subfield of computer science worldwide. It would be impossible to
review all of the 75 papers presented, but I have attempted to
list some of the more interesting ideas given in three different
categories: Methodologies that were measured and clearly shown to
work, methodologies that are currently being evaluated with no
measurements available yet, and miscellaneous systems that do not
fit into the other two categories.
Most of the attendees were from academia, government, or from
companies such as Boeing that have the same information retrieval
problems as everyone else. Conspicuously absent from the
conference were all the commercial marketers of text/search
retrieval software that were in force at the CD-ROM conference a
few weeks earlier. At the beginning of one of the sessions,
someone summed up the situation quite accurately when they said
that in the beginning of the field of IR (Information Retrieval),
there was a lot of theory but no applications. Now there are a
lot of applications (commercial) with no regard to theory. IR
research has shown clearly that Boolean type retrieval of full
text databases yields only 16% of relevant documents even when
queried by an expert (DIALOG, SAFE, etc). This conference
demonstated that there are many methodologies that can be employed
to increase this rate of recall and precision. In fact, each
Declassified and Approved For Release 2013/02/13: CIA-RDP93B01194R001700050003-9
 Declassified and Approved For Release 2013/02/13: CIA-RDP93B01194R001700050003-9
methodology tends to yield a different set of valid hits, thereby
pointing to solutions that employ multiple, different types of
approaches simultaneously. AI techniques and knowledge bases were
clearly shown to be of use in textual processing especially when
they employed techniques that used feedback from users in an
interative fashion to refine the results.
Category 1: Proven Methodologies
a) Attribute-based model described in "A Technique to Improve
the Precision of Full-Text Database Search," by David Hsiao, Naval
Postgraduate School.
b) "Towards a Friendly Adaptable Information Retrieval
System," by Shih-Chio Chang, GTE Laboratories. Uses PROLOG to
take relevance feedback until it meets the needs of the user.
c) "Inverted Signature Trees: An Efficient Text Searching
Technique for Use with CD-ROMs," by Alan Tharp, North Carolina
State University. A new data structure called inverted signature
trees is introduced and compared with textsignatures and the B+
tree. This new data structure facilitates rapid access to all
sentences in a text file that contain specific search words while
still maintaining search words in alphabetical order. Using the
12,000 page World Book Encyclopedia stored on a CD-ROM, a 4 word
search using text signatures would require 3.8 minutes, whereas it
would only take 6 seconds using inverted signature trees. If a
word is not present in the file, the hashing procedure used to
form the signatures returns this result immediately since the
inverted lists need not be accessed. For the B+ tree, four
seconds would be required if the entire B+ tree index is stored in
memory, otherwise it would take 10 seconds.
d) "Who Knows: A system Based on Automatic Representation of
Semantic Structure," by Lynn Streeter, Bell Communications
Research.. I found this approach (which is purely mathematical as
opposed to linguistic) to be one of the most interesting of the
conference. The actual application, "Who Knows" was developed so
that anyone calling Bell for technical information would be
directed to the appropriate internal organizations. For instance,
what if someone wanted to know something about-Artificial
Intelligence and who was doing what. In Bell, as in our own
Agency, there might be many different people/organizations working
in this area. Based upon an arbitrary natural language request
from a user, the system returns a ranked list of all pertinent
organizations containing the who, what, where and how that a user
wants to know. When Bell did a controlled test ofthis method
which uses vector mathematics to place relevant documents in the
same vector space, they found it to be more than 5 times more
accurate than traditional keyword matching systems, even when the
keywords were weighted. "Who Knows" is written in C and runs
under a standard Unix system. A typical query takes 2-3 seconds on
a Sun 3/75. The system was also able to predict, based upon an
analysis of 263 new technical abstracts, the actual department
Declassified and Approved For Release 2013/02/13: CIA-RDP93B01194R001700050003-9
 Declassified and Approved For Release 2013/02/13: CIA-RDP93B01194RO01700050003-9
PAGE 3
that produced each one.
(e) "An Application of Artificial Intelligence Techniques to
Automated Key-Wording," by James Driscoll, Univ. of Central
Florida. The JCS (Joint Chiefs of Staff) uses a computer program
to store, categorize and retrieve observations/lessons learned
from military exercises, wargames and real-world military
operations. The system requires interraction with keyword experts
for the documents to be catalogued correctly. Using Artificial
Intelligence techniques to automate this manual keywording process
proved not only to be more accurate than any given keyword
expert's performance (as measured by a team of 5 experts
collaborating together), but it also cut the indexing time per
document down from 7 minutes to 15 seconds.
(f) "Looking for Needles In a Haystack or Locating
Collocational Expressions in Large Textual Databases," by Yaacov
Choueka, Bell Communications Research. This paper was delivered
under the session titled "Automatic Thesaurus Construction." It
addresses the problem of collocational expressions in generating
an automatic thesaurus. A collocational expression is defined as
a sequence of two or more words whose exact meaning or connotation
cannot be derived directly from the meaning of its components.
Examples: west bank, big board, the west coast. Using ten million
words of the New York Times News Service (9 1/2 months of all the
wire stories) a thesaurus was generated of collocational
expressions using techniques described in the paper which
contained a high degree of precision and recall.
(g) "Conceptual Information Retrieval From Full-Text," by
Richard Tong, Advanced Decision Systems. This paper describes
RUBRIC, a system developed for an unnamed client (obviously from
the intelligence community based on the "terrorism" examples
used). It's goal is to "understand" the text of documents in
sufficient detail that an automatic determination can be made as
to whether a user should analyze it. The system replaces the
parsing of text employed by other natural language efforts with an
evidential accrual and reasoning mechanism. The results are -
impressive: 68% precision and 68% recall based upon 730 articles
taken from the-Reuters newservice. ADS has two other operational
systems which run on a SUN. One contains 2000 rules in the
knowledge base and the average retrieval speed is between 1 and 10
seconds against a...1.00.MB collection of mixed documents (manuals,
almanacs, and short reports).
Category 2: Methodologies Currently Being Evaluated
a) "Intelligent Search of Full-Text Databases," by Susan
Gauch, Univ. of North Carolina. An expert system reformulates a
user's Boolean query depending upon whether whether too many or
too few hits were obtained. A hierarchical thesaurus is used as
one of the techniques for expanding a user's Boolean query when
too few hits are obtained.
Declassified and Approved For Release 2013/02/13: CIA-RDP93B01194RO01700050003-9
 Declassified and Approved For Release 2013/02/13: CIA-RDP93B01194R001700050003-9
L)
PAGE y
b) "About Reformulation in Full-Text IRS (Information
Retrieval Systems)," by Christian Fluhr, Centre National de la
recherche Scientifique (France). An expert'system is used as a
reformulation strategy manager to translate a request for
information from very large textual databases in unrestricted
natural language. Reformulation of both the query (which is often
expressed in terms different from those found in the documents
that should be returned) and the database texts themselves so that
they converge is accomplished via an expert system which applies
various reformulation techniques such as lemmatization, misprints
and spelling correctors, explicitization, syntagmatic extensions,
and weighted comparisons.
c) "The TINA Project: Text Content Analysis," by Dr. Schwarz,
Siemens, West Germany. The U.S. Department of Commerce U.S.
Patent and Trademark Office is conducting a large scale test of
TINA Natural Language programs on 195,000 patent abstracts. This
software combines morphological software with syntactic analysis
of free-text to aid a user in modifying his search question or in
indexing a given document he may be adding to the system. Future
plans include the addition of semantic processing of free-text and
the integration of parallel processing techniques for the
retrieval and matching algorithms.
d) "Tex-Nat: A Tool for Indexing and Information Retieval," by
J.M. Lancel, CAP SOGETI INNOVAtION. This tool for Indexing
documents (not using the traditional stop list) and retrieving
them in spite of spelling errors via a phonetic equivalent is
available on PC-DOS and UNIX machines and will be ported to the
IBM/VM and PS\0S2 environments.
e) "Conceptual Information Extraction and Retrieval From
Natural Language Input," by Lisa Rau, AI Program, GE Research and
Development Center, N.Y. SCISOR (System for Conceptual
Information Summarization, Organization and Retrieval) Is a
prototype system for taking texts in constrained domains and
storing them in conceptual format which can be interrogated by
natural language. The system is written in Common Lisp and
processes stories at the rate of a few seconds. per sentence. An
on-line news service is currently being connected to enable
extensive testing of the system. Other potential uses of the
system are structured database generation from text and automatic
summaries of complex events.
f) "Automatic Thesaurus Construction by Machine Learning from
Retrieval Sessions," by U. Guntzer, Technical University of
Munich. TEGEN is written in PASCAL usinga knowledge-based
programming method running on a CYBER 990 computer. It is is
presently being used to build a thesaurus from a controlled
vocabulary learnt by the computer by analyzing the searches
carried out by users over a period of time based upon both
implicit and explicit feedback. The system is presently being
used with a collection of 130,000 documents of the Faculty of
Declassified and Approved For Release 2013/02/13: CIA-RDP93B01194R001700050003-9
 Declassified and Approved For Release 2013/02/13: CIA-RDP93B01194R001700050003-9
Mathematics and Computer Science. Results of a manual analysis of
the thesaurus entries generated will be reported on at a later
date.
g) "A Logic Programming Approach to Full-Text Database
Manipulation," by R. Marshall, Loyola College. NASA's ENVIRONET
database was captured in the form of facts and rules using.-first
order predicate logic (PROLOG) to facilitate user's retrieval of
information in text, mathematical formula-based numerical
extraction, interpolation and extrapolation, and tables and
graphs. Also, the system understands the difference between an
aerodynamics expert or a computer engineer retrieving information
about "RAM". The system runs on a MicroVAX II at Goddard Space
Flight Center in Greenbelt, MD.
h) "Interactive Knowledge-Based Indexing: The MedlndEx
System," by Susanne Humphrey, National Library of Medicine. This
paper describes the MedlndEx System which uses a knowledge base to
provide interactive assistance in indexing the periodical medical
literature. The system is in prototype form and the results of
evaluation programs are not yet available.
Category 3: Miscellaneous Topics
a) "Structure of Information in Full-Text Abstracts," by
Elizabeth Liddy, Syracuse University. Proves that discourse
linguistics can be used to extract a structure from free-text
documents such as scientific abstracts, thereby making it possible
for these documents to be handled successfully in information
retrieval systems. The automatic detection and use of this
structure Is being implemented on the Connection Machine at
Syracuse University.
b) "Using English for Indexing and Retrieving," by Boris Katz,
AI Laboratory, M.I.T. START (Syntactic Analysis using Reversible
Transformations) is an attempt to understand English text and
index the knowledge contained within it (for use in generating an
AI knowledge base) and also to generate English text from
retrieved portions of the knowledge base.
c) "Information Aids for Technological Decision-Making: New
Data Processing and Interrogation Techniques for Full-Text Patent
Databases," by W.A. Turner, CDST/ CNRS, France. This technical
alert system was developed by CNRS to provide French Industry with
a competitive edge in international markets. It exploits the data
in full-text patent databases. Its premise is that pre-established
classification schemes are not useful for describing the
scientific and technological state of the art at any given moment.
Through co-word analysis techniques they have been able to detect
subject areas in which a statistically significant number of
patents have been filed during a given period of time. The user
is able to obtain activity profiles for any research topic (i.e.,
conductive thin film) which shows, by country, whether or not that
country's "innovative activity" is weaker or stronger than
Declassified and Approved For Release 2013/02/13: CIA-RDP93B01194R001700050003-9
 Declassified and Approved For Release 2013/02/13: CIA-RDP93B01194RO01700050003-9
expected..
d) "Browsing and Authoring Tools for a Unified Medical
Language System," by Henry Kamorowski, National Library of
Medicine. Because available commercial' knowledge engineering
environments have only rudimentary display and browsing tools,
IntelliCorp's KEE and Interlisp-D were combined with their own set
of lower level layout and display routines to provide.a "fish-
eye view" of a semantic net. BATS (Browsing and Authoring Tools
for Semantic networks) was written so that it could easily be
ported to another environment, i.e., the MAC II.
e) "The Informatics Calculus: A Graphical Functional Query
Language for Information Resources," by Richard Epstein, West
Chester University, PA. A doctoral thesis which describes the
design of a query language which uses informatics calculus to meld
traditional database and hypertext technologies.
f) "Implementing a Distributed Expert-Based Information
Retrieval System," by Edward Fox, VPI, VA. For four years, Fox et
al at VPI have been developing a distributed expert-based
information system written in C and PROLOG called CODER. Various
modules (Query Formulator, Search, Problem Description Builder,
Browse, Lexical) work together under a Retrieval Strategist
(Planning and Coordination Heuristics) to process a user's request
at various levels simultaneously. Future plans include an
evaluation of 3 different approaches to accessing data on a CD-
ROM: TOPIC (based on RUBRIC) is an AI type of retrieval, Personal
Librarian which is based on statistical algorithms to rank
output, and Window Book, a hypertext system.
g) "Universal Multilingual Information Interchange System," by
Suban Krishnamoorthy, Framingham State College, MA. Once in
awhile some unknown person from some unknown place comes up with a
truly revolutionary concept which largely goes unnoticed for a
number of years. Such is my impression of this paper which was
given by this native Indian (who had to cope with several of the
fifteen different Indian languages as he moved around India as a
student together with several European languages plus English as a
graduate student and professor). He contends that a unified
solution to handle and interchange information in any (variable)
number of scripts and languages does not exist. He then describes
a .complete design for just such a universal.multilingual
communication system down to the minutest detail, including
character sets, language coding, I/O devices, multilinual keyboard
design, font design, etc. He concludes by stating that the design
of his system is simple, practical and viable now and that a
universal computer communication system (CCS) such as the one he
shows (with no linguistic barriers or boundaries) will be the
future CCSs of the world and have enormous social impact.
h) "Automatic Recognition of Sentence Dependency Structures,"
by Timothy Craven, University of Western Ontario. A system for
creating automatic abstracts from documents. Based on results, it
Declassified and Approved For Release 2013/02/13: CIA-RDP93B01194RO01700050003-9
 Declassified and Approved For Release 2013/02/13: CIA-RDP93B01194RO01700050003-9
appears to have more immediate utility as a means for extracting
structured information from long texts for use in online
databases.
i) "The Utah Retrieval System Architecture (URSA)," by Lee
Hollaar, University of Utah. Using a distributed information-
retrieval system using Apollo workstations, windows, and
specialized backend processors, a 40 GB database with 120 parallel
searches and an aggregate search speed of'150 MB/second"is
currently being implemented and will cost about $2-3 million.
(There were a number of other papers describing experimental
hardware architecture solutions such as VLSI , neural networks and
parallel processing.)
Declassified and Approved For Release 2013/02/13: CIA-RDP93B01194RO01700050003-9