American Journal of Computational Linguistics ~i crofiche 77 
:-$ THE FINITE STRING 
NEWSLETTER OF THE ASSOCIAT.IOlj FOR COMPUTATIONAL LINGUISTICS 
George E Heidorn kps accepted appointment as Editor-of 
AJCL His term begins with the first .issue fgr 1979 
The experiment With a microfiche journal is over. The 
next issue of the journal will be on paper, with possible 
microfiche supplement. 
The re~iting Editor apologizes to the membership for the 
long delay in release of the present material for publica- 
tion. Grave persona1 difficulties interfered with all of 
Hays"s routine activities during the period, and voluntary 
activities necessarily suffered most 
Released for publication March 25, 1979, 
AMERICAN JOURNAL OF COMPUTATIONAL LINGUISTICS is'puhljshed 
by the Association fpr. Comfiutational Linguistics 
EDITOR David G Hays, 5048 Lake Shore Road, Hamburl:, IJew 
York 147775 
EDITORIAL ASSISTANT William Benzon 
MANAGING EDITOR Donald E Walker, Artificial Intelligence 
Center, SRI International, Menlo Park, California 94025 
TECHNICAL ADVISOR Martin Kay, Xerox Palo Alto Research 
Center 
American 
Journal 
Cornput at ional 
Linguistics Microfiche 77:2 
CONTENTS 
ACLt MINUTES OF THE 16TH ANNUAL BUSINESS MEETING ...... 3 
............. 
SECRETARY-TREASURER'S REPORT 7 
OFFICER! FOR 1979 ................... 9 
OFFICERS 1963-1979 .................. 10 
......... 
NSFr SUPPORT FOR COMPUTATIONAL LINGUISTICS 12 
NEWSX SHORTNOTES .................. 14 
ARISTREPRINTREQUEST ................. 16 
.. ......... SUMMERLINGUISTICSATTE)CAS ... 18 
PH D PROGRAMS IN COMPUTATIONAL LINGUISTICS ... 19 
JOURNAL: COMPUTATIONAL LINGUISTICS AND COMPUTER LANGUAGES . 21 
SUMMARY OF RESEARCH ON COMPUTATIONAL ASPECTS OF EVOLVING 
................. THEORIES RaymondD . Gumb 27 
TAXONOMY: INFORMATION SCIENCES 
Editors of Information Systems . .. 31 
MACHINE AIDS TO TRANSLATION A CONCISE STATE OF. THE ART 
................ BIBLIOGRAPHY WayneZachary 34 
REVIEWS~ ON HUMAN COMMUNICA.TION. 3D ED. BY COLIN CHERRY 
............ 
William L . Benzon .... 41 
ABH~~NGIGKEITSGRAMMATIK. BY JURGEN KUNZE 
................. 
KennethF . Ballin 47 
GLANCING. REFERRING AND EXPLAINING IN THE DIALOGUE SYSTEM 
...... 
. 
HAM-RPM W . Wahlster. A . Jameson. and W Hoeppner s;b 
A CRITICAL LOOK AT A FORMAL MODEL FOR STRATIFICATIONAL 
.............. 
. 
LIN~~JISTICS Alexander T Borgida 68 
ASSOCIATION FOR COMPUTATIONAL LINGUISTICS 
MINUTES: 16th Annual Business Meeting 
26 July 1978 
University of Illinois, Urbana, Illinois 
Jonathan Allen, President, presiding 
MINUTES OF THE PREVIOUS MEETING 
Allen noted that the minutes 
of the previous meeting had been published 
in the Finite String, Volume 14, Number 3-, Microfiche 65 of the American 
Journal of Com~utational Linauis tics. 
FINANCIAL STATUS 
Don Walkw, Secretary-Treasurer, reviewed the financial status of the 
Association, a copy of which is attached to these Minutes. He presented 
income and expehses both for 1'977 and for 1978 through 21 July. The 
balances of $6,332.19 for 1977 and $7,060.63 for 1978 constitute ajs9et9 
of $13,39r.82. 
However, the major costs for AJCI, for the current year 
are yet, to be incurred, and dues for AFIPS have not yet been billed, At 
the same time, the $1,000 advance to cover costs of the TINLAP-2 Meeting 
is likely be returned, along with a portion of the excess of income over 
expenses, which will be shared with ACM/SIGART. 
The indebtedness of the Association ta the Center for Applied 
Linguistics, which housed the previous Secretariat, has been paid off. 
Of the total of $13,486:06, funds from the previous Secretariat provided 
$9,913.92; $3,572.14 was taken out of income for 1977. 
I 
Walker remarked that the income item from AFIPS for 1977 of $2,365, 
which refledted a disbursement of surplus funds from the National 
Computer Conference, was unlikely to be repeated again soon. AFIPS is 
in the process of creatAfig 2 new journal, ABACUS, modeled after 
Scientific American, and future surpluses probably will be used to 
defray or at least to backup the startup ,costs. The journal is expected 
to be self-sustaining, and might eventually show a profit. 
The current balance in the TINLAP-1 Account is $109.41; approximately 75 
copies remain. 
MEMBERSHIP 
Walker reported membership figures for 1977 of 500 individual and 201 
institutional, for a total of 701. The current figures for 1978, 
through 21 July, are 405 individual and 208 institutional, for a total 
of 613. A slightly more detailed breakdown is attached to these 
Minutes. 
EDITOR 'S REPORT 
Dave Hays, Edito-r of the AJCL, announced thaG he was resigning, 
effective at the end of tb year. The next issue will include the 
TINLAP Proceedings; the final issue of the current year will contain a 
complete index for the five years of its publication. The Jourrlal wa3 
established in 1974 as an NSF-sp,onaored experiment in microfiche 
pub1icat;ion. Anticipating the implications of George Heidornta survey 
(see below), Hays remarked that this mode of publication is likely to be 
replaced by microprocessor technology and might never repeive a full 
scale trial. 
Allep expressed the gratitude of t,he Association to Hays for his 
devotion, his constructivenes~, and his tireless efforts in establishing 
and sustaihing the A JCL . This tribute was affirmed by the members. 
Allen then announced that Heidorn, currently Associate Editor, would 
replace Hays as Editor in January. 
A NEW FORMAT FOR THE AJCL 
Heidorn presented the results of his survey of the qembership regarding 
a new format for the AJCL. Of 513 questionnaires mailed (to both 
current members and to members ~ho paid for 1977 but not yet for 1978); 
212 were ret;'urned. The responses favored crea'tion of a hard copy 
edition, similar in format to the Communications of the ACM, with or 
without an acoompanying microfiche version. Most members felt that such 
a change would encourage a wider readership and an increase in the 
submission of techical articles, A more comprehensive report by Heidorn 
is included elsewhere in this iswe. 
Allen reported that the ExecubSve Committee, after reviewing Heidorn's 
findings, has decided to issue the Journal in the new format with both 
hard copy and microfiche versions sent to each membep, beginning with 
the first issue of 1979. The microfiche also may contain appendixes for 
technical articles, program listings, and other material of interest to 
the membership but less appropriate for Inclusion in hard copy form, 
like the list of members. 
NEXT MEETING 
Allen stated that a decision about the time and place of the 1979 
meeting had not yet been made. Asiloalar, near Monterey in California, 
is being considered, possibly around the time of the next International 
Joint Conference on Artificial IntelLlgence, which wj-11 be held in Tokyo 
from 20 to 24 August. In the discussion, members expressed concern that 
the choice should not discourage attendance by graduate students. 
For 1980, the Executive Committee recommended that a third TINLAP 
meeting be held. The President, Vice President, and Secretary-Treasurer 
constitute an interim opmmittee to investigate this possibility and to 
negotiate with SIGART, a3 appropriate. Allen also remarked that an 
offer already had been received from Aravlnd Joshi to host TINLAP-3 at 
the Universi ty of Pennsylvania. 
NOMINATIONS FOR OFFICERS FOR 1979 
Aravind Joshi, reporting for the Nominating Committee (Joshi, Petrick, 
and Chapin), announced the following nominat5ons for officers for 1979: 
Nominating Committee: Jonathan Allen, MIT, 
Executive Committee: Stanley Roaenschein, RAND 
Secretary-Treasurer: Donald Walker, SRL International 
Vice President: Bonnie Lynn Webber, University of Pennsylvania 
President: Ronald Kaplan, Xerox Palo Alto Research Center 
No additional nominations were received from the floor. Allen called 
for a vote, which was unanimous, and the alate was declared elected. 
NEW BUSINESS 
Carol Lane presented a resolution supporting the ratification of the 
Equql Rights Amendment to the U.S. Constitution. After extensive 
discussion and after motions to amend and to table were defeated, the 
members affirmed the following substitute resolution by a vote of 20 to 
13: 
WHEREAS, inclusion in the Constitution of these United States 
is the basic unalienable right of every citizen; 
WHEREAS, the Association for Computational Linguistics views 
as intolerabl.; the selective exclusion of over one-half 
the pophation of this country; 
WHEREAS, the Equal Rights Amendment, writing women into the 
Constitution, must be ratified by three-fourths of the 
states (38) prior to its incorporation 
THEREFORE, BE IT RESOLVED, that all future conventions, 
rneetlngs, and conferences of the Association for 
Computational Linguistics will, for the duration of time 
durlng which the Equal Rights Amendment is under 
consideration by the several states of the United States, 
be held only in those states that have ratified the Equal 
Rights Amendment. 
RESOLUTIONS 
HavQng set out in advance the syntax of his report for the Resolutions 
Committee, Ron Kaplan expressed the gratitude of the AssocYation to Dave 
Waltz, his session organizers, and the Unlverdty of Illinois for the 
organization and conduct of the meeting; to the National Science 
Foundation, and particularly to Carol Ganz Brown, for its financial 
support; to the current officers for their constructive-efforts during 
the first part of their elective term (and with encouragement to 
contlnue these efforts for the rest of the year); to the retiring 
members of the AJCL Editorial Board--Robert Barnes, Fred Damerau, Gary 
Martins, John Olney, and Naomi Sager--for five years of effective 
service; to Dave Hays for his countless hours and fruitful endeavors in 
realization of the AJCL; and to George Heidorn for his willingness to 
become the new editor. 
Dave Hays called attention to the efforts of Martin and Iris Kay in the 
preparation of the AJCL ~ibliography, and they were duly included in the 
list of resolutions. 
The members, affirmed these sentjments enthusiastically, and Allen 
directed the Secretary-Treasurer to express the appreciation ~f the 
Association formally to Dave Waltz., 
The meeting adjourned. 
Donald E. Walker 
Secretary-Treasurer 
Attachments: Financial Status, Membership Status, Officers for 1979 
FINANCIAL STATUS 
ASSOCIATION FOR COMPUTATIONAL LINGUISTICS 
Income: 
Dues 
Back Issues 
Meetings 
AFIPS 
Mail ing Labels 
Interest 
TINLAP-1 Account 
.TL Production $ 
AJCL Bibliography 
AJCL Editorial 
Neeting Expenses 
AFIPS Dues 
Secke tariat Services 
Postage 
Supplies 
Printing 
TINLAP-1 AccounL 
Center for Applied 
Linguistics 
Balance : $ 6332.19 
Assets: 
Savings 
Checking 
Petty Cash 
Center for Applied Linguistics Account 
Debt (as of 3-14-77) $13486.06 
Paid (out of 1976 funds) 9913 .92 
(out of 1977 funds) 3572.14 
TINLAP-1 Account 
Current Balance $109.41 
STATUS 
Individual 
US 
Foreign 
Institutional 
US 
Foreign 
Special 
TOTALS 
ASSOCIATION FOR COMPUTATIONAL LINGUISTICS 
Officers for 
- 
Dr. Ronald M. Kaplan President 
Xerox Palo Alto Research Center 
415:494-4436 
3333 Coyote Hill Road 
Pal~ Alto, CA 94304 
Professor Bonnie Lynn Webber Vlce President 
Computer and Informat'ion Scimce 
235: 243-8540 
The Moore School 
University of Pennsylvania 
Philadelphia, PA, 19 104 
Dr. Donald E. Walker 
Artificial Intelligence Center 
SRI International 
Menlo Park, CA 94025 
Dr. Jerry Rw Hobbs 
Artificial Intelligence Center 
SRI International 
Menlo Park, CA 94025 
Dr. Bertram C, Bruce 
Bolt Beranek and Newman 
10 Moulton Street 
Cambr i.dge , MA 02 13 8 
Dr. Stanley J. Rosenschein 
Rand Corporation 
1700 Main Stsreet 
Santa Monica, CA 90406 
Dr. Stanley R. Bet~ick 
IBM Watson Research Center 
P.O. BOX 218 
Yorktown Heights, NY 10598 
Dr. Paul G. Chapin 
Linguistics Program 
Natl onal Science Foundation 
Washington, DC 20550 
Professor Jonathan A1 len 
Electrical Engineering 
MIT, 36-575 
Cambridge, MA 0213,g 
Dr. George E. Heidorn 
IBM Watson Research Center 
P.O. Box 21-8 
Yorktown Heights, NY 10598 
Secre tary-Treasurer 
415:326-6200~307I 
Execut 3ve Committee 
(1977 - 1979) 
4 15 :326-6200x2229 
Execut ive Commi t tee 
(1978 -- 1980) 
617:491-1850 
Executive Committee 
(1979 - 1981) 
213 :393-0411 
Nominating Committee 
(1977 - 1979) 
914: 945-2175 
Nominating Committee 
(1978 - 1980) 
202: 254-6326 
Nominating Committee 
(1979 - 1981) 
617: 253-2509 
Editor, AJCL 
91 4: 945-2776 
OFFICERS 
ASSOCIATION FOR MACHINE TRANSLATION AND COMPUTATIONAL LINGUISTICS (1963-1968) 
ASSOCIATION FOR COMPUTATIONAL LINGUISTICS ( 19 68-1 979) 
President 
V ice-Pres 
S ec-Treas 
Executive 
Committee 
--------- 
Editor (FS) 
Nominating 
Committee 
--------- 
President 
Vice-Pres 
Sec-Treas 
Executive 
Committee 
-------I- 
Editor (FS) 
Editor (MTCL) 
Nominating 
Comrni ttee 
President 
V ic e-Pr es 
Sec-Treas 
Executive 
Committee 
-----1-1- 
Editor (FS) 
Editor (AJCL) 
Nominating 
Comrni ttee 
--I--&.--.- 
President 
V ic e-Pres 
Sec-Treas 
Ekecutive 
C ommi t te e 
- ----C--- 
Editor (AJCL) 
Assoc Editor 
Nominating 
Committee 
196_7 
Yngve 
Hays 
Josselson 
Rho de s 
Garvin 
Lehmann 
Roberts 
See 
Oetttnger 
Lamb 
1967 
Kuno 
Walker 
Jos sels on 
Sat terthwait 
Hockett 
Pendergraf t 
Roberts 
Yngve 
Garvin 
Hays 
Lieberman 
xu 
Fri edman 
Simmons 
Jo~seLson 
Wall 
Robinson 
Montgomery 
Roberts 
Kay 
Plath 
Walker 
I9i5 
Joshi 
Petrick 
Roberts 
Mart ins 
Rieger 
Nash-Webbe.r 
Hays 
S l mmons 
Barnes 
Woods 
m 
Hays 
Alt 
Josselson 
Sebeok 
Garvin 
Lehmann 
Roberts 
Y ng ve 
Oe ttinger 
Lamb 
1968 
Malker 
Mersel 
Jos se 1s on 
Sat terthwai t 
Fromkin 
Pendergraf t 
Roberts 
Yngve 
Garvin 
Kuno 
L ieberman 
1972 
S immons 
Fromkin 
Roberts 
Wall 
Robinson 
Chap in 
Roberts 
Kay 
Plath 
Friedman 
m 
Petrick 
Grimes 
Roberts 
Diller 
Rieger 
Nash-Webber 
Hays 
Joshi 
Barnes 
Woods 
ll96.5- 
Lehmann 
Garvin 
Josselson 
Sebeok 
Hockett 
Kuno 
Roberts 
Yngve 
Hays 
Lamb 
wi9 
Kay 
Plath 
Jos selson 
Sat terthwai t 
Fromkin 
Montgomery 
Roberts 
Yngve 
Garvin 
Kuno 
Walker 
XEi. 
Barnes 
Woods 
Roberts 
Mart ins 
Robinson 
Chapin 
Roberts 
S immo ns 
Plath 
Friedman 
Dxz 
ehapin 
Allen 
Walker 
Diller 
Hobhs 
Nash-Webber 
Ways 
Heidorn 
Jos hi 
Petrick 
Woods 
19.66 
Garvin 
Oe tt inger 
Josselson 
Sebeok 
Hocket t 
Prendergraf t 
Roberts 
Yngve 
Hays 
Lieberman 
197_Q 
Plath 
Friedman 
Josselson 
Wall 
Fromkin 
Nontgomery 
Roberts 
Y ng ve 
Kay 
Kuno 
Walker 
39.39 
Woods 
Wall ' 
Roberts 
Mart ins 
Jos hi 
Chapin 
Hays 
Simmons 
Barnes 
Friedman 
1978 
Allen 
Kap lan 
Walker 
Diller 
Hobbs 
Bruce 
Hays 
Heidorn 
Jos hi 
Petrick 
Chapin 
President 
V ice-Pres 
Sec-Treas 
Executive 
Committee 
---- ...--...- 
Editor (AJCL) 
B~soc Editor 
Nominating 
Committee 
---".---- 
Kap lan 
Webber 
Walker 
Rosenschef n Ros ens.chein 
Hobbs 
Bruce Bruce 
Heidorn 
~l'en Allen 
Petrick 
Chapin Chapf n 
Allen 
FS =Them- 
MTCL = Hachine ,- - -tics 
AJCL = uw. J-L pf ComDutational 
American Journal of Computational Linguistics 
N S F SUPPORT FOR COMPUTATIONAL LINGUISTICS 
Pau1.G. Chapin, Director of the Linguistics Program of the 
National Science Foundation, announced the follow~ng grants 
for research of obvious relevance to computational linguist,ics. 
LANGUAGE UNIVERSALS ARCHIVING PROJECT 
Charles A. Ferguson and Joseph Greenberg 
Stanford University - $49,200 - 13 months 
COMPUTER STUDIES IN FORMAL LINGUISTICS 
Joyce Friedman 
University 'of Michigan - $40,000 - 24 months 
(The Intelligent Systems Program of NSF awarded the same amoun~) 
COMPUTATIONAL COMPLEXITY OF GRAMMAR G NL RECOGNITION PROBLE'MS 
William Rounds 
University of Michigan - $44,600 - 24 months 
American Journal of Computational Linguistics 
During the -Fiscal Year 1978, the Division of Science Information 
of NSF issued at least the following grants for supportof 
research relevant to computatibnal Linguistics. 
CORRELATION OF LANGUAGE STRUCTURE WITH INFORMATION 
Zellig- S. 'Harris 
University of ~erinsyivania - $163,329 - 29 months 
INTEGRATED 'MAN/MACHINE INTERFACE FOR NETWORK RESOURCE UTILIZATION 
Martha E. Williams 
University of Illinois - $174,432 - 24 months 
American Journal of Computational Linguistics 
The Microprocessor Standards Comittee of the IEEE formed a 
subcommittee to work on standsrds for the five high-level 
languages widely used in microprocessor applications:, Basic, 
Fortran, Cobol, PL/M, and Pasqal. The object foreseen was the 
iaentification and endorsement of appropriate standards. The 
first meeting was called in October by Bruce Ravenel of Language 
Resources, 1311 Lombard StrBet, San Francisco, CA 94109. 
What fields need linguists or persons with some linguistic &now- 
ledge? What are the "demands and perspectives" of present and 
possible employing institutions? Anyone with information 
relevanf to these issueb can correspond with Dr. Walther Kindt, 
~akuZ'tPt fur Linguistik und Literaturwissenschaft, Universitat 
Bielefeld, Postfach 8640, 48'00 Bielefeld 1, Germany. 
The de-tection of orthographic errors in keyboarding* of Swedish 
text is the topic of work undertaken by Rolf. Gavare, 
Department 
of Computational Linguistics, GBteborgs Universitet, Norra 
Allegatan 6, S-413 01 GBteborg, Sweden, who invites correspondence. 
MTt ENGL~SH AND* THAI VIA MONTAGUE GRAMMAR 
The work of Joyce Friedman is being applied in work on transla- 
tion. Kurt Godden, 1408 E. 27, Lawrence, KS 66044, invites 
contact. 
American Journal of Computational Linguis ties 
ASSOCIATION FOR WOMEN IN COMPUTING, 
AWC was founded in 1978 to promote communication, professional 
development and advancement, and education.. Membership is open 
without restriction. The Corltespondent of AWC is Anita Cochran, 
58137 Bell Laboratories, Murray Hill, NJ 07974; 201-582-7817. 
A program . tu check spelling in English text has been written by 
Ralphe E. Gorin, A1 Laboratory, Computer Science Department, 
Stanford University, CA 94305, with additions by William Plurmner 
and Jerry Wolf of BBN and Richard Johnsson and Philip Xarlton 
of Carnegie Mellon University. 
ONLINE BIBLIOGRAPHY IN SIGN LANGUAGE AND  ELATED AREAS 
The Syracuse Information Retrieval Experiment system has been 
adapted by Jim Bourg of the Gallaudet College Library and is 
being used by William C. Stokoe, Linguistics Research Lab, 
Gallaudet College, Kendall Green, Washington, DC 20002, for 
a pseyho- and sociolinguistic bibliography. Anyone with a 
console may inquire about access; no charge is levied at present. 
American Journal of Computational Linguistics 
Request for reprints 
ANNUAL REVIEW 
INFORMAT ENCE 
AND 
Volume 14 of ARIST is in preparation. Authors of chapters 
need help in finding relevant recent publications. They will 
appreciate receivirrg offprints from authors at their respec- 
tive addresses. 
If'the appropriate chapter writer is not apparent, write tb 
Martha E. Williams 
Editor, ARIST 
R.R. No. 1 
Monticello, Illinois 61856 
CHAPTER TOPICS AND WRITERS 
COMPUTER. ARCHITECTURE FOR 
NATURAL LANGUAGE AND IR 
APPLICATIONS 
Professor P. Bruce Berra 
(& Ellen Oliver) 
Syracuse University 
Syracuse, New York 13210 
RETRIEVAL TECHNIQUES 
Professor Michael McGill 
School of Information Studies 
113 EucZid Avenue 
Syracuse University 
Syracuse, New York 132i0 
COST ANALYSIS OF SYSTEMS AND 
SERVXCES 
Mr. Colin Nick 
Applied Communication Research 
P.O. Box 5849 
Stanford, California 94305 
LIBRARY AUTOMATION 
Els. Mary Jane Probst Reed 
Associate Director for 
Research and Planning 
Washington State Library 
Olympia, Washington 9,8504 
Mr. Hugh Vrooman 
Illinois State Library 
Centennial Building 
Springfield, Illinois 62796 
INFORMATION SYSTEMS 
LAT@N WRICA 
Professor Tefko Saracevis 
School of Library Science 
Case Western Reserve University 
Cleveland, Ohio 44106 
(Additional authors for this 
and following chapters on 
next frame.) 
ARIST REQUEST FOR REPRlNTS 
IlJFORMATION SYSTElIS IN LATIN 
AMERICA (continued) 
Gilda Braga 
Institute Brasileiro de infor- 
macao em Ciencia e Technologia 
Av. General Justo 171; 4O 
Rio de Janeiro, Brazil 
Alvaro Qui.jano Solis 
Biblioteca Daniel Cosio 
Villegas 
El Colegio de Mexico 
Camino A1 Ajusco No. 20 
Apartado Postal 20-671 
Mexico 20, D.F. 
Mexico 
INFORMATION SYSTEMS FOR 
CONSUMER CONCERNS 
Profess-or Vivian Sessions 
School of Library Science 
McGill University 
3459 McTavish Street 
Montreal, Quebec H3A 1Y1 
Canada 
COMPUTERS AND PUBLISHING 
Mr. David Staiger 
American Institute for Aero- 
nautics & Astronautics 
1290 Avenue of the Americas 
New York, New York 10019 
EDUCATION AND TRAINING 
FOR ONLINE SYSTEMS 
Ms. Judy Wanger 
1523 Sixth Street, Suite 12 
Santa Monica, California 90401 
DATA BASE MANAGEMENT SYSTEMS 
Dr. Ronald Wigington 
(& Michael A. Huffenberger) 
Chemical Abstracts Servic~ 
2540 Olentangy River Road 
Columbus, Ohio 43202 
SYSTEI4S DESIGN- -PRINCIPLES 
AND TECHNIQUES 
Dr. Ronald Wyllys 
Graduate School of Library 
Science 
University of Texas 
Austin, Texas 78712 
FUNDAMENTAL PRINCIPLES AND 
THEORIES OF INFORMATION SCIENCE 
Dr. Pranas Zunde 
School of Information and 
Computer Science 
Georgia Institute of Technology 
Atlanta, Georgia 30332 
American Journal of Computational Linguistics 
Summer Linguistics at Texas 
The University of Texas at Austin announces 
a special Summer Linguistics Program 
which will consist of a substantial offering 
of graduate courses given by our own 
faculty members and a distinguished list of 
visitbg scholars. In addition to these 
courses, workshops and seminars (for 
credit as conference courses by arrange- 
ment with individual faculty members) 
will be available on topics such as syntac- 
tic universals, conditions on rule applica- 
tion. fip;ij- and scope, formal vs. substan- 
tive explanation in phonology, etc. Ure in- 
vite applications from graduate students in 
linguistics and allied disciplhes such .as 
education, English, foreign languages, psy- 
chology, philosophy, anthropology, and 
others, 
The list of courses and instructors will in- 
clude the following: 
LIN 380K Generative Phonology-Robert 
T. Harms 
IAN 380L Transformational Grammar- 
Jorge Hankamer and Ivan Sag 
LIN 381M Phonetic Theory-Peter MacNeilage 
LIN 384 Outline of Turkish Syntax-Jorge 
Hankryner (tentative) 
LIP 393 Semantics-Robert E. Wall 
IJN 393 Seminar in Phonetics and 
Phonology-Bj6m Lindblom 
&IN 393 Seminar in Syntax and Semantics- 
Emmon Bach and Barbara Partee 
LIN 394 New Directions in Historical. 
Linguistics-Robert D. King 
LIN 398 Seminar in Linguistic 
Variation- John Baugh 
In ad& tion to the above-listed faculty 
members, the following scholars will be 
available for individual consultatbu: Lee 
Baker, Peter Cole, David Decamp, Polly 
Jacobson, Lauri Karttunen, W. P. Leh- 
mann, Fritz Newmeyer, Susan Schmerling, 
and others. Several intensive Oriental and 
European language courses will also be 
taught as a part of the regular UT summer 
session. 
Low cost accommodations will be availa- 
ble in housing cooperatives. Classes will 
begin June 13 and exams will end July 20; 
the Program will thus not conflict with the 
LSA Linguistic Institute in Salzburg. 
Tuition and fees: For Texas residents, the 
price of summer courses is $64.70 for one 
three-hour course, $89.90 for two three- 
hour courses. For out-of-state residents, it 
Ps $159.70 for one three-hour course, 
$304.90 for two three-hour courses. In ad- 
dition, there is a $10.00 property fee, 
refundable at the end of the course. 
For application materials, pkase complete 
the detachable section and mail it to: 
Summer Linguistics Program 
Department of Linguistics 
University of Texas at Austin 
Austin, Texas 78713 
DEADLINE: May 1,1979 
American Journal of Computational Linguistics 
PHD PROGRARS IN 
COnPUTATIONAL LINGUISTICS 
During the summer of 1978, Alan K. Melby wrote to many 
American universities, asking about the graduate work 
in computational linguistics. He has supplied a copy 
of his list of affirmative answers, presented here in 
the casual format of its compilation. Melby points out 
that MIT and Yale did not respond but would be considered 
by a student planning work on computers and language. 
California C. C. Cheng 
Linguistics 
Ken Wexler University of Illinois 
School of Social Sciences Urbana, IL 61801 
University of California 
Irvine, CA 92717 Kansas 
(cognitive science, mathema- 
(tical linguistics 
Electrical Engineering Dept . 
University of California 
Los Angeles, CA 9002~1 
(some work in A1 
Clara Bush 
Depqrtment of Linguistics 
Stanford University 
Stanford, CA 94305- 
(under Prof. Terry Winograd- 
Connecticut 
David Michaels 
Room 230 H.R. Monteith Bldg. 
University of connecticut 
Storrs, CT 06268 
(analysis and synthesis of 
(speech--Haskins Laboratory 
(connect ion 
G. K. Krulee 
Department of Linguistics 
David Dinneen 
Department of Linguistics 
University of Kansas 
Lawrence, KS 66045 
Emon Bach 
Department of Lfnguistics 
University of Massachusetts 
Amherst, NA 01002 
Michigan 
Joyce Friedman 
Computer and Communication Sciences 
University of Michigan 
Ann Arbor, I41 '48108 
(with qualifications--ask JF 
Minnesota 
Center for Research in Human 
Learning 
University of Minnesota 
Minneapolis, MN 55455 
- 
(psychologists with interest in A1 
~orthwes tern University 
Evanston, IL 60201 
(with Computer Science Dept. 
PHD PROGRAMS IN COMPUTATIONAL LINGUISTICS 
New York 
DavidLHays 
State University of New York 
Buffalo, NY 14214 
Lewis Levine 
Department of Linguistics 
Washington Square 
New York, NY 10003 
(students can work under 
(Naomi Sager 
Simon Belasco 
Department of Linguistics 
Pennsylvania State University 
University Park, PA 16802 
ified "yes" 
Rhode Island 
J. 3. Wren 
Box E 
Brown University 
Providence, RI 
Texas 
Department of Cinguistics 
University of Texas 
Austin, TX 78712 
Department of Linguistics 
Georgetown University 
Washington, DC 20007 
American Journal of Computational Linguistics 
COMPUTATIONAL LINGUISTICS AND COMPUTER LANGUAGES 
EDITORS: T, FREY, T, VAMOS 
PUBLISHER: COMPUTER AND AUTOMATION INSTITUTE, BUDAPEST 
. . *. 
EDITORIAL BOARD$ €3. DOMOLKI, E. FARKAS, F. KIEFER, T. LEGENDI. 
A. MAKAI, F. PAPP, G SZEP, D. VARGA 
CONTENTS OF NUMBER lit 
I. Nemeti: On a property of the category of partial algebras 
Gy. Revesz: A note on the relation of turing machines to 
phrase structure grammars 
P.B. Schneck: A new program optimization 
B. ~omijlke Formal description of software components 
E. Santa-Toth: by structured abstract models 
G. Fay: 
H . Heiskanen : 
T. Legendi: 
Cellular design principles, a case study of 
maximum selection in codd-icra cellular space /I/ 
Semantic theory from a systematical viewpoint 
Callprocessors in computer architecture 
Gy. Hell: Mechanical analysis of Hungarian sentences 
One double-issue or two issues per year of ca. 350 pp., 20.5 x 28.5 cm. 
1977 (numbers 12 + 13) will be published 1978/1979. 
Price per issue: Hfl. 42,- + postage. 
Send orders to: 
John Benjaxning B. V. 
Amstewk 44 / P. 0. Box 52519 
1007 HA Amaterdrm / The Netherlands 
~cl.: (030 j 738156 I ~dnr 15798 jw 
CaMw: BENPER /Amatdam 
American Journal of Computational Linguistics 
New Journal 
DISCOURSE PROCESSES 
A MULTIDISCIPLINARY JOURNAL 
EDITOR: ROY 0. FREEDLE ABLEX PUBLISHING CORPORATION 
Educational Testing Service 355 Chestnut Street 
Princeton, New Jersey 08540 
 orw wood, New Jersey 07648 
609-921-9000, ext. 2651 201-767-8450 
Personal subscription. $19.50. Institutions: $45~00. 
The role of culture-specific schemata in the comprehension and 
recall of stories 
Walter Kintsch and Edith Greene 
A code in the node: The use of a story schema in retrieval 
Jean M. Mandler 
An experimental investigation of contingent query schemes 
Catherine Garvey-and Mohamed BenDebba 
Inference and coherence 
Edward J. Crothers 
How to catch a fish: The memory and representation of common 
procedares 
Arthur C. Graesser 
American Journal of Computational Linguistics 
BOOKS: 
IGOR A. MELICUK. STUDIES IN DEPENDENCY SYNTAX. Ann Arbor: 1979 
Karoma Publishers, Inc. 
6 by 9 inches 
163 +ixpp., pap 
ISBN, 
nly $4.50 
720-001-2 
............ Foreward (by Paul T. Roberge, editor) v 
.... Preface .................... xiii 
Dependency Syn-tax ......... +. ............ 1 
... The Prediicative Construction in Dyirbal ....... 23 
Types of Surface-Syntactic Relations ............ 91 Bibliography ....................... 151 
Abbreviations and Symbols . . Lr- ............ 162 
I. MEL~CUK. R. RAVI~. AUTOMATIC TRANSLATION, 1964-1970. 
Departenlent de linguistic et Case postale 6128, Succursale IlA1! 
philogie Montreal, P.Q. N3C 3J7 
Universite de Montreal 
Ce volume qui est une suite au guide bi bliographique analytique de 
la T(raduction) A(utomatique), publie a Moscou en 1967, recense 
1357 ouvrages de la T.A. et les domaines voisins, parus entre 1964 
et 1970, aussi bien en U.R.S.S. qu'en Occident. 
Pout la plupart 
des ouvrages recenses, on a donne un r@sume detaille (en russe) qui 
se veut une description suffisante du contenu. Le guide est desti- 
ne aux 1 inguistes, aux traducteurs et aux informaticiens, chacun y 
pouvant trouver des rsnsei griements uti Tes. Prlx approximati f du 
volume: $16. 
This vol urn, a conti nuation of the Analytical Bi bl iographi cal Guide 
to A(utomatic) T(ranslation), published in Moscow in 1967, includes 
1357 works in and on TA, as we1 1 as neighboring domains, which ap- 
peared between 1964 and 1970 either in the Soviet Ur-tion or in the 
West. For most of the pyblicati-ons listed, a detailed abstract is 
provided, which is intended to adequately represent the contents of 
the corresponding item. The guide is addressed to linguists, trans- 
1 ators and information processing spec? a1 i sts a1 1 of whprn wi 11 hope- 
fully find in it useful data. Approximate price will be $161. 
Books cont. 24 
J.D. Goldstein, D.W. Lakamp, A. Pietrzyk. INFORMATION SERVICES ON 
RESEARCH IN PR~GREsS: A WORLDWIDE INVENTORY. 
National Technical Information 
Service, US Dept. of Commerce 
5285 Port Royal Road 
$ringfield, VA 22161 
462 pp. , softcover 
PB-282 02>/AS 
$14.50 hard copy 
$3.00 microfiche 
Part I: World Trends in Information on Research in Progress* An 
Overview 
Part 11: Profiles of Information Systems and Services on Research 
in Progress 
Part 111: Indexes to o'rganization and system names, organization 
and system acronyms, persons respinsible and subject 
coverage given 
Starr Roxanne HPltz, Murray Turoff. THE NETWORK NATION* HUMAN 
COMMUNICATION VIA COMPUTER. Addison-Wesley: 1978 
Hardbound ISBN 0-201-03140-X c. 826.50 
Paperbound ISBN 0-201-03140-8 c. $14.50 
The Nature of Computerized Conferencing 
Potential Applications and Impaces of Computerized Conferencing 
Proj ecting the Future 
F.W. Lancaster. MWARD -PAPERLESS INFORMATION SYSTEMS. Academic 1978 
179 pp., $J3.50 
ISBN 0-12-436050- 6 
In this nontechnical presentation the author first describes the 
paperless inform~tion/comrrnication systems currently being developed 
for intelligence agencies. Then existing scientific and technological 
communication systems are critiqued. Finally the intelligence 
design is reform~lat~ £or the science and technology environment. 
L. Bourrelly and E. Chauraqui. THE DOCUMENTATION SYSTEM SATIN 1. 
VOL. 1 - GENERAL DESCRIPTION AND USER'S MANUAL (1974) 398 PP. 
VOL. 2 - SYSTEM GENERATION AND SET~UP INSTRUCTTONS (1978) 397 PP. 
SATIN 1 is implemented at the "Laboratoire dlInformation pour les 
Sciences de l'Homrnel' in Marseille and is designed specifically for 
the-documentation processes in the humanities and social sciences, 
archeology, geography, history, etc., and permits the representation 
of information about artifacts of many kinds (paintings, sculputures 
etc.) in addition to standard document content. 
Books cont . 
25 
Herbert E, Bruderer. SPRACME - TECHNIK - KYBERNETIK. 
AufsHtze zur- Sprachwissenschaft, maschinellen Sprachverarbeitung, 
- 
kiinstlichen ~ntelligenze und Computerkunst. ~erlag Linguistik, 
MCfnsingen/Bern 1978, 187 Seiten, 39 Schweizer Franken (Foreign orders 
must be prepaid: Surface 44 Swiss Francs, Air 50 Swiss Francs), 
ISBN 3-85784-000-5 
Herbert E. Bruderer. HANDBOOK OF MACHINE TRANSLATION AND MACHINE- 
AIDED TRANSLATION. AUTOMATIC TRANSLATION OF NATURAL LANGUAGES AND 
MULTILINGUAL TERMINOLOGY DATA BANKS. New York 1978, 
North-Holland Publishing Company 959 Pages, ISBN 0-7204-0717-6. 
Foreward by Prof. K. Bauknecbt, Department of Computer Science, 
University of Zurich. English translation by the Commission of 
the European Communities, Brussels. c.1600 item bibliography. 
Herbert E. Bruderer (Hg. led. ) . AUTOMATISCHE SPRACH~~BERSETZUNG. 
(in preparation) 
Wissenschaftliche Buchgesellsrhaft, Darmstadt 1979. etwa 450 
Seiten (Wege der Forschung, Band 272),. 41-45 DM. ISBN 3-534-06312-0, 
nur fur Mitglieder der Buchgesellschaft. 
American Journal of Computational Linguistics 
YALE A. I . PROJECT: RESEARCH KEPORTS AVAILABLE 
Send orders to: M.S. Barham, Adm. Asst., A.I. Project, Department of 
Computer Science, Yale Univers~ t y , J 0 Hl llhouse Ave . -320DL 
New Haven, CT 06520 
< > # 78 - Riesbeck, C.K. 
& Schank, ROC= - Comprehension by Computer: Expectation-Based 
Analysis of Sentences In Context. 
< > # 88 --Lehnert, W.G. - The Process of Question Answering 
< > #I16 - Cullingford, R.C. Script Application: Computer Understanding 
of Newspaper Stories 
< > $227 - Schank, R.C. 6 
Carbonell, J.G. - Re: The Gettysburg Address 
< > #I28 - Schank, R.C., 
Wilensky, R., 
Carbonell, J.G., 
Kolodner, J.L. & 
Hendler, J-A. - Representipg Attitudes: Some Primitive States 
< >" #I31 - Lehnert, WcG. - Representing Physical Objects In Memory 
< >-#137 - Charnfak, E. - On The Use Of Framed Knowledge In Language 
Comprehension 
< > 8139 - Riesbeck, C.K. 
& Charniak, E. - Micro-SAM and Micro-ELI: Exercises in 
Popular Cognitive Yech-snics 
< > #140 - Wilensky, R. - Understanding Goal-Based Stories 
Ph-D. Thesis 
< > #I41 - Schank, ROC. - Inference in the Copceptual Dependency 
Parad igm: A Personal History 
< > #142 - Kolodner, J4Lr - Memory Organization For Katural Language 
Data-Base Inquiry 
< > #I43 - Schapk, RbC. & 
Birnbaum, LaA= - Real-Time I~tegrated Parsing 
< > 8144 - Schank, R.C. & 
Lebowi t z, bf - Big Words 
<- > #I45 - Schank, B.C. - Interestingness: Controlling Inferences 
< 
> fib6 - Carbonell, J.G. 
Cullingford, R.E. & 
Gershman , A.Y. - Knowledge-Based Fachine Translation 
American Journal of Computational ,Linguistics 
SUMMARY OF RESEARCH ON COMPUTATIONAL ASPECTS OF 
EVOLVING THEORIES 
Raymond D. ,Gumb 
Temple University 
The concent of an evolving theory [3] is a natural 
extension of concepts in free tense lopic with equality. 
In the semanticr of free tense logic, an individual can 
have a property at one time Mat tt does not ab another, 
the domain of discourse can vary with the passage of time 
as individuals are born an8 die, an8 individual terms can 
refeq at one time but not at another. understand in^ a 
theory to be a set of sentences in the languape of free 
tense logic with equality, an evolvinq theory is an indexcd 
eet of theories ordered In time. Intuitively, an evolvinp 
theory mip;htreoresent the life work of a thinker, where 
the individual theories in the evolvinp theory correspond 
to chapters of the life work written at different ti&es. 
The evolving theory reflects chan~es in the thinker's 
views with the passape of tfme. 
Evolvinp theories have been studied from both semantic 
and  roof theoretic ~erspectives, and various concents 
such as the semantic conce~t of satiafiability and the 
woof theoretic concept of consistency have been extended 
to anply to evolving theories [3]. 
The semantics is given 
Ih terms of metaphor theory which stands intermediate 
between Leblanc'e truth-value semantics and moddl theory 
and translates readily into both. 'The deductive system 
far (a class of) evolving theories Is celled the forest 
method because a tree is generated for each point in time. 
The forest method, in effect a jqeneralization of Kripke 
tableaux construct ions, is mechanlzable, The forest 
method is correct with res~ect to the semantics 131. 
The forest' method can be &n~lled whemer the1 restriction 
on the temporal order relation has the com~utable Krinke 
closure ~ro~erty, a property of m roper ties of relations 
whlch has been characterized model theoretically in [?I . 
The eornput'able Krlpke closures are a subclass pf the 
monotonic closures r7], which have closure properties much 
like transitivity (a computable Kripke and monotonic closure). 
A preservation theorem giving the syntactic form of 
axiornatizatione of the first-order monotonic cloeures [7] 
suagests a peneralizatton of the Roy-Warshall transitive 
closure algorithm. The preservation theorem might also 
be used to determine (much as an entry in an engineer in^ 
handbook) whether, given a property of relations Pr, there 
is a Pr-closure algorithm. 
Evolving: theories can be based in other intensional 
loales such as as modal [4] and intuitionistic lopics. 
Some (but notr?all) of the loeics underlying kinds of evolvin~ 
theories where the forest method is ap'alicable can be 
axiomatized 11, 6J* 
In certain (but not all) cases, the forest method 
(or restrictions of) can be used to give effective 
proofs of me extended Joint consiatencg theorem, a result 
whlch incornorates the Craip and Lyndon interpolation 
lemmas and.the Robinson joint consistency theorem. 
Roughly, the theorehi nates that two theories TI an8 T2 are mutu- 
ally inconsistent just- in case there is a separating sentence 
F such that F (not F) is a logical consequence of TI (T2) 
and F "talks about'" only individual@ and relation8 Chat 
both TI an8 T do. Effective proofs of the theorem are 
2 
well-known in standaydl first-order logic, and sjmilar results 
hkwe been established' In free lq~ic with equality [5] 
and in a family of free modal lo,qics with. equality[4]. 
Since the  roof Is effective, a (de~th first) algorithm 
can be e~trlctea from it for constructing, piven a closed 
forese for the union of TI and T2, a senarating sentence F. 
Evolving theordes, outfitted with additional devices 
to enhance their plausibility, appear to be a natuial for 
representinp ttemporal knowledge. When the extended joint 
canal-stency theorem holgs for the representation langulsge 
of a knowleaae base and when an effective proof has bean - 
given, the separating sentence algorithm might be a 
useful tool for ~in~ointing inconsistencies in the knowleifve 
base 2. 
Only very restricted versions of the al~orithms 
mentioned. above have been programmed, those in LISP [2] 
and SNOBOL, 
American Journd of Computatioaal Linguistics 
TAXONOMY: I NFORMAT I ON SC I ENCES 
The following taxonomy is that used by the journal INFORMATION 
SYSTEMS. 
1 qcncrnl aspects 
analysis, modbllinq, dcscrl ptlon and evaluation 
of informatlon systems 
analysis 
dcslan 
nodelling 
dcscrlption 
implementation 
3 data basc systems 
3.1 qlobal aspects, global desion 
3.2 system analysis for FBS, userdbmands 
3 -3 fcasibllity studlc~i:, evaluation of DRS, 
summary of expcrlenccs 
3.4 formal description of data basc 
data bascJanquaqcs 
systems and 
3.5 data modcls, informatlon modcls 
3.5.1 hierarchic DM 
3.5.2 network DM 
3.5.3 relational DM 
3.5.4 others 
3.6 data def in3 tion languaqcs (DDL) 
3.7 data translation 
3.8 procedural data manipulation lanfiuaqcs (DML) 
3.9 non-procedural (descriptive) DML, tcrmlnal lanqu-acres 
(interactive lanquaqcs),natural lanquaae intcriaccs fl 
3.10 dialog functions, computer assistance, computer 
guidance, dialog support; 110-functions 
3.11 implementation aspects 
3.12 architecture of D~S, interfaces 
3.13 distributed DBS 
3.14 file rfianagement systm 
3.1 5 data . structures, operat2ons upon da'ba structures 
3.15.1 data structures 
3.15.2 operations 
3.19.3 pictorial data structures and operations, 
data bases in computer graphics pnd CAD 
3.F6 storage structurcs, access path, access methods, 
search strategies 
3.W reorqanization, selforqanislng structures, 
optirnlzation 
3.18 storage technoloqy, speciallzcd hardware for DBS 
(data - base processors) 
3.19 security, integrity 
3 -20 privacy 
3.21'descrlption of realizeet data base systems 
3 2 . I DBTG-based DBS 
5. ZI . z re~atlonal-Elm DBS 
3.21.3.1 ADABAS 
3-21 -3.2 DMS-I1 
3..11 -3.3 DMS-I100 
3.21.3-4 IDS 11 
3.21 -3.5 IDMS 
3.21-3,Q IMS 
3.21 -3.7 SYSTEM 2000 
3.21.3.8 TOTAL 
3.21 .3.9 others 
4 method and model base systems 
4.1 method base systems 
4.2 model base, systems 
4.3 description-of reakzed systems 
5 planning and decision support systems 
5.1 gorecasting systems 
5.2 pJannlng information systems 
5.3 decision support systcms- 
5.4 control systems 
5.5 description of realized systems 
6 question answering systems, coqnitivc methods 
I representation of knowledge, 
6.2 problem solving 
6.3 natural language systems 
6.4 pattern processing 
6.5 deduction and inference 
6.6 artlf icial intelligence methods, cognit.ivc methods 
6.7 description of realized systems 
7 - document retrieval systems 
7.1 indexinn, classification, thesaurus problems 
7.2 qvaluation measures 
7.3 documentation services 
7.4 library automation 
7.5 description of realized systems 
8 dlskributed systems 
8.1 distributed processing 
8.2 distributed storing 
8.3 distributed control 
8 4 data c~mmunication, protocols 
8.5 architecture, topoloqy 
8.6 descript on of realized systems 
9 specla1 application orlentcd lnforrnatlon systems 
management information systems 
macro economic information systems 
information systems in public adrninistxatlon 
information systems in medicine 
technical information systems 
information systems in jurisprudence 
ecoloqy information systems 
description of realized systems 

systems. This reflects the development of MAT as a more practical al-terna- 
h. 
tive to pure machine translation and its position as one of the only areas 
of computational lingustics that has progressed to the point of wide scale 
appl ication. 
A few other ground rules have been followed. First, since the 
field continues to change so rapidly and this is a state-of-the-art survey, 
few references over a decade old are included. Second, where a single 
MAT system is described almost identically in several research reports 
and/or publications in several languages, only a single reference is given, 
in Engl ish where possibTe. Third, works deal ing stric'tly with hardware 
advances such as new graphics display techno1 ogy, micropragramnabl e di spl ay 
fonts, or mu1 ti1 ingual printers are a1 so excluded. 
This bib1 iography was compiled from a much larger one containing 
nearly 750 entries and is intended to\ provide a concise summary of the 
current work in the field with strong focus as stated above on machine 
aided translation systems that are presently in opera ti^^.* 
* ThiQ research was conducted under a subcontract to Analytics from 
Chase, Rosen and Wallace, Inc. with funds provided by the Office of 
Research and Development, Central Intel 1 igence Agency. 
ALPAC 
(Automatic Language Processing Advisory Committee) 
1966 Lanquage and Machines - Computers in Translation and Linguistics 
National Academy of Sciences, National Research Counci 1 : Washington, 
D.C. 
American Mathehatical Society 
1968 "Research on Machine Aids to- an Editor of Scientific Translations." 
NTIS document PB-177 775. 
Eerzon, V.E. 
1971 "Some Techniques for Formalizing the Process of Establishing 
U-Re1 ationq Between Sentences i'n a Corrective Text," in Nauchn6- 
Tekhniceskaya Informatsia, Seriya 2, No. 8. 
"Bisby, R. and Kay, M. 
1972 ThelMIND Translation System: A Study in Man-Machine Collaboration. 
NTIS Document Ad 749 000, Wnd C0r.p. : Santa Monica, Cal if. 
Bruderer, H. E. 
7 977 "The Present State of Machine Aided Trans1 ation," Overcoming the 
Language Barrier, in Commission of-the European Communities, 
pp 529-556. 
Bruderer , H. E. 
In Press Handbook of Machine and Machine Aided Translation, Aorth 
Holland: New York (published in German in 1975). 
Burge, John 
1978 "The TARGET lj'roject's Interactive Mu1 ti1 ingual Dictionary, 
It 
Project Technical Report No. 13. Depts. of Modern Languages and 
Computer Science; Carnegi e-Me11 on University. 
Charniak, E. and Milks, Y.: eds 
1976 Computati onal Semantics, North Holl and: New York. 
Chevalier M., Danserau, J., and Paulin, 6. 
1978 TAUM-METO: Description du Systeme. University de Montreal, 
Montreal , C$nada. 
Comi ssion of the European Comrnuni ties, 
19f7 
Overcoming the Language ~arrier : Thi rd European Congress on 
-1nformat5on Systems and Networks. Vol . 1. Munich: Verlag 
Dokurnantations. 
Daley, Cd. D.H., and Vechino, Lt. Col. R.F., USAF 
1973 
"The West German Federal Bureau of Languages and Machine Aided 
Transl ation in Germany. " In Federal Linguist, Vo1. .5, No. 3-4; 
pp 14-18. 
Dubuc, Robert 
"TERMIUM: System Description;* in Mta: Journal of ~ranslatfon, 
Vol . 17, NO. 4, pp 203-219. 
Dubuc, Robert, and Gregoire, Jean-Francois 
1974 "Banque de Terminologie et Traduction," in Meta: Journal of 
Translation Vol . 20, No. 4, pp 180-184, (in French). 
Goetschal ckx, 3. 
1974 "Terminology and Documentation in International Organizations ,-" 
in Babel, Revue Internationale de la Traduction, Vol . 20, No. 4, 
pp 185-187. 
Grimes, Joseph E. 
"Computing in Lexicography," in The Linguistic Reporter, Vol . 12, 
NOS. 5-6, pp 1-4. 
Hann, Michael 
1974 "~r-inciples of Automatic Lemmatisatlo-n-," in ITL, Review of Appl ied 
Linguistics, Vol. 23, pp 3-22. 
Hirschberg , Lydia 
1965 
"Dictionaries Automatiques pour Traducteurs Hurnai ns ,Ii in Journal 
des Traducteurs, Vol. 10, No. 3, p 78, (in French). 
Hlavac, T. 
1973 
"Dealing with the Language Barrier by Means of Computer," in 
Kniznue and Ved. Inf. (Czcheoslovakia), Vol. 5, No. 2, pp 70-75. 
Ivanova, I.S. 
1969 "Problems of Automatic Thesarus Construction," in Nauchno- 
Tekhnfceskaya Informatsiya, Seriya 2, No. 1, pp 17-20 
Joint Publication,Research Service 
1977 
"Machine-Assisted Trans1 ation in West Germany," trans1 ation of 
article by-xarious authors, NTIS document JPRS 68726. 
Jordan, S.R., Brown, A., and Hqtton, F.C. 
1977 "Computerized Russian Trans1 ation at ORNL ," in Journal of the 
American Society for Informatior! Science, January, pp 26-33. 
Jossel son, H. 
1967 "Research in Automatic Russian-Engl ish Scientific and Technical 
~exicography," Final Technical deport, Wayne State University. 
NTIS document PB-7 80 930. 
Kay, Martin 
1976 "The Proper Place of Man and Machines in Translation," in 
American Journal of Computational Lingu+stics, Microfiche 46. 
Krol lman, Fredrick 
1971 "Lingui sti c Data Banks and the Technical Trans1 ator, " in Meta: 
Journal of Trqnslation, Vol. 16, Nos. 1-2, pp 117-124. 
Krollman, Fredrick 
1974 "Data Processing at the Trans1 ators Service," in Babel , Revue 
Internationale des Traducteurs, VO~. 20, No. 3, pp 121-129; 
Krol lman, Fredrick 
1977 
"User Aspects of an Automatic Aid to Translation as Employed in 
A Large  rans slat ion Service,'' In Overcoming the ~anguage Barrier&, 
by the Commission of the European Communities, pp '243-257. 
Lipprnan, Erhard 0. 
"An Approacn to Computer Aided Translation," in IEEE Transactions 
on Engineering Writing and Speech, Vol. 14, No. 1, pp 10-33. 
Lippman, Erhard 0. 
1975 "On-Line Generation of Termi~ological Digests in Language Trans- 
lation," in IEEE Transactions on Professional Communications, 
Vol . PC-18, NO. 4, pp 309-318. 
Lippman, Erhard O., and Plath, W.J. 
1970 "Time Sharing and Computer Aided Translation," in The Finite 
String, .Vole 7, No. 8, (microfjrhe). 
Loh, Shiu-Chang 
1976 "Translation of Three Chinese Scientific Texts into English & 
Computer," Association of Literary and Linguistic Computing 
Bulletin, Vol. 4, No. 2, pp 104-106. 
Loh, S.C., and Xong, L. 
1977- "compute; Translationu of Chinese. Scientific Journals," in Overcoming 
'the Language Barrier , in Commission of the European Comuniti es, 
pp 631-646. 
Luther, D.A., Montgomery, C., and Case, R. 
1977 
"An Interactive Text-Editing System in Support of Russian Trans- 
lation by Machine," in IFIPS National Computer Confereoce 
Proceedings, pp 789-790. 
Lyttle, E.G.9 et. al. 
1977 
"Junction Gramnar as a Base for Natural Language Processing," in 
American:Journal ofjComputationa1 Linguistics, 1975, No. 3, 
(mi crof i ?fie). 
Mathias, Jim 
1973 
"The Chinese-English Translation Assistance Group and Its 
Computerized Glossary Project," in Federal Linguist, Vol . 5., 
NOS. 3-4, pp 7-13. 
Dew Scientist 
1'977 "Many Hands Make Chinese Translator Work" in The New Scientist, 
Vol. 76, No. 1073, p 88. 
Price, James 
7 970 "Abstract of the Development of a Theoretical Basis for Machine 
Aids for ~ransiation from Hebrew to Engl i sh," in Hebrew -Computational 
Linguistics Bulletin, Vol 2., 65-83. 
Schmidt, R.2 and Vol lnhals, 0. 
1974 "The Use of the Lexicographical Branch of a Data Bank System 
to Produce a phrase01 ogi cal Technical Glossary," in ~achine 
Assisted -Translation in West Germany, JPRS document 69726. 
Schulz, J. 
1971, "Le Systems TEAM, une Aide a la Traductlon," in Meta: Journal of 
Translator, Vol . 16, Nos. 1-2.. (in French). 
Schulz, Joachim 
1974 "Lexicography with TEAM -- Automatic Dictionary Composition,'' 
in ~achi ne- ~ssi sted Transl ati on in WesZ ~ermany , JPRS document 
68726, pp 23-34. 
Shaffer, Richard A. 
1978 
"California Firm to Unveil a Computer that Processeb Words for 
Translators ," inr-The Wall Street Journal, 24 October 1978. 
Sinaiko, H.W. 
1971 
"Trans1 ation by Computer," in Science, Vol . 174, pp 1182-11 84. 
Smith, Raoul 
1978 "Computational Bilingual Lexicography: A La Recherche du mot 
Juste," pape? read at Foreign Broadcast Information Service 
Seminar on Computer Support to Transl ation, 
Stallings, W. 
1975 "The Morphology of Chinese Character: A Survey of Models and 
- - 
Applications," in Computer and the ~umanities-, Vol . 9, pp 13- 
24. 
Waf ker, Gorden, Kuno, Susumu, Smith, Barbara, Hold, Rol and 
1968 "Chinese Mathematical Text Anal vsis ," in IEEE Transactiol~s 
on Engineeritig Writing and ~~&ch,-~ol, 11, No. 2, pp 118- 
128. 
Warotamasikkhadit, U., Kanchanawan, N., and Londe, D. 
(no date) "The Design and Construction of a System to Transl iterate Thai 
by Computer," in 6th Austral ian Computer conference Proceedings, 
pp 833-839. 
Weaver, W. 
1955 "Trans1 ation," in Machine Trans1 atSon of Languages, W .N. Locke 
and A.D. Booth, eds., Technology Press of MIT and John Wiley 
and Sons: New York. 
Weber, Heintz Josef 
1976 
"Au tomatiche Lemmati seirung -- Zi el setzung und Arbei tsweisceines 
Linguistischen Identifikati~nsverfahrens," in Linguistische.. 
Berichte, Vol . 44, pp 30-47 (in German). 
Wilks, Yorick 
1973 "An Artificial Intelligence Approach to Machine Translation, 
If 
in Computer Models of Thought and Language, Roger Schank and 
Kenneth Colby,' eds., W.N. Freeman and Co.: San Francisco, 
pp 114-115. 
Unknown 
1976 "The Lexicography Informati on Sys tern (LEXIS) of the Bundeswher 
Language Service," in Machine Assisted "h-ansl ation in West 
Germany,NTIS Document JPRS 68726. 
young, Mary E. 
1978 
"Machine Translation (A Bib1 iography with Abstracts) ,'I available 
f rorn National Technical Information Service, NITS document 
PS-78f0448. 
American Journal of Comput~tional Linguistics 
ON HUMAN COMMUNICATION: 
A REVIEW, p SURVEY, AND A CRITICISM 
COLIN CHERRY 
3rd Ed. Cambridge, MA 
M.1.T PRESS. 1978 
'374 + xv pp. 
ISBN 0-262-03065-9 
Reviewed by 
WILLIAM L. BENZON 
Language, Literature and Communications 
Rensselaer Polytechnic Institute 
Troy, New York 12181 
The comments on the dust jacket for the third edition of Colin Cherry's 
On Human Communication come from reviews of previous editions whf-ch appeared 
in such diverse places as the Canadian Journal of Psychology, Physics Today, 
and Romance Phllolo~ ahd lndicllte that it fs a broad ranging and fascinating 
book. The range is certainly broad, so broad that I have included a table- 
of contents as an appendix to this review rather than attempting to summarize 
the contents of the hook. 
But it is no longer fascinating. Students of human c~mmunication 
have thought and debated much and even learned a little between the pub- 
lication of the first edition of On Hman Communication in 1957 and the pub- 
lication of the third edition in 1978. But very little of that material has 
become part of the substance of Cherry's book @though some of it is cited). 
Thus, while the first editf on may well hkve been "A Review, a Survey, and 
a Criticism'' (the book's subtitle), the third edition is not. 
Too much is 
left unconsidered. 
The most serious gap is in the consideration of natural languages. 
Cherry gives the impression that infomation theory, statistics, Fourier 
analysis, and perhaps a little logic are the most important fonnal tools 
for the analysis of natural language. That may have been true when Cherry 
mote the first edition of the book, but it is cqrtainly not true now, 
Cherry does make a few references to Chomsky, but none of the substance of 
the Chomskian revolution (not to mention post-Chomskian developments) has 
affected Cherry's treatment of natural language. The texture of theorizing 
and model building in lin@istics, psycholinguistics, and cognitive psychology 
has undergone considerable change since the Fifties, but little of that 
has become part of the substance of On Human Communication. 
Similarly, a great deal of work: has been dove on nonverbal commur~ication 
in the last two decades. While cherry alludes to some of this work, he 
midies no attempt t~ summarize any of the major lines of inquiry. As with 
modern linguistics, the material is too diverse to cover it a11 in the sort 
of survey which Cherry intends On Human Congmicatidn to be. What I find 
so disheamtening is that so little of thls material is mentioned at all, 
especially when one realizes that Cherry has added a new chapter ("~uman 
Communication: Peeling, Knowing, and Understanding") to ,the third edition 
for the purpose of talking ab~ut what is specifically human about human 
communication. That Cherry should devote ten pages to Zipfvs law while 
not even mentioning the work of Paul Ekman, Carroll Izard, aad Manfred 
CJJrnes (to mention only the work which comes mqst readily to my mind) on 
the expression and cmnunication of emotion is biearre, 
-Pinally, Cherry's treatment of semiotics is relatively insulated from 
most of the-semiotic research of the last tqo de&es. Semiotics is itself 
such a diverse and amorphous enterprise (as diverse and amorphous as the 
study OF human communication is) that it is perhaps unfsir to criticis* 
Cherry for shortchanging it. 
Rut Cherry introduced the t3piC of ritual 
into his final chapter 
and that is a subject on which semioticians have 
had a great deal to say (I am thinking of structural anthropologists such 
as Claude Levi-Stmuss, Edmund Leach, and Victor Turner). C,onsequently 
I am inclined to view Cherry's neglect of Semiotics perhaps more 
harshly than I otherwise would. 
No doubt Cherry catlld be charged with other sinssf oniinission, but the 
three I've mentioned are serious enough, It is equally beyond doubt that 
an edition of On Human Communication which $ncluded this material would be 
a very different book, not Colin Cherry's book at all. If Cherry had, 
attempted and achieved a synthesis of his material, then its value as a synthesis 
might well outweigh the dated nature of some of the elements of the synthesis. 
Rut Cherry wasn't after a synthesis; he simply wanted to see what was out 
there. There is now much out there which Cherry hasn't seen. 
Consequently 
On Human Communication is not a good guide. 
The person -who wants or negds 
a general introduction to the subject of human communication which reflects 
the current state of the art(s) will have to look elsewhere. 
APPENDIX: Abbreviated table of contents from On Human Communication 
Chapter 1: ,Communication and Organization - - An Essay 
1 'The Scheme of This 3ook 
2 What Is  communication"^ 
3 
\&at Is It That Wg Communicate? 
4 Some Difficulties of Description of Human Communication 
5 Co-operative and Non-co-operative Links 
6 Communication and Social Pattern 
? Group Networks 
8 Communication Is an Act of Sharing 
Chapter 2: Evolution of Communication Science - an Historical Xeview 
1 Language and Codes 
2 The Mathematical Theory of Communication 
3 
Brains - Real and Artificial 
4 On Scientific Nethod 
Chapter 3: On Signs, Lanwage, and Communication 
1 Language: Science and Aesthetics 
2 ~fiat Is a Language? 
3 
Toward a Logical Description of Language 
4 Features as the "General Co-ordinat es" of Szeech- 
5 
Statistical Studies of Language "Form" 
6 Wsrds and Meaning: Semantics 
Chapter 4: On Analysis of Signals, Especially Speech 
1 The Telecomrhunication Engineer Comes - onto the Scene 
2 Spectral Analysis of Signals 
3 Speech Representation on the Frequency-Time Plane 
4 The Specification of Speech 160 
Chapter 51 
On the Statistical Theory of Communication 
1 Doubt, Inf omation, and Discrimination 
2 Hartley' s Theory r " Infonation" as Logical "Instructions 
to Select" 
3 
When the Alternative Signs Are Not Equally Likely to Occur 
lc The Use of Prior Information: Redundancy 
5 Messages Represented as Vave Forms: "Continuous~nf omation 
6 Communication as Information, When Noise is Present 
? 
The Ultimate Capadty of a Noisy Channel 
8 Mandalbrotls Explication ofZipf'sLaw- Continued 
9 
Comments on Infomation Interpreted as hxropy 
Chapter 6: On the Logic of Communication (~yntactics, Semantics, 
and ~matics) 
1 "Signifies" - of Mental Hygiene 
2 Are Different Measures of " Infomat ion" Needed? 
3 
About "Semantic Inf ornat ion" 
4 
Syntactic, Semantic, and Pragmatic "Infomationw 
A Relationship 
5 Language, Logic, and Zxperirnent 
Chapter,?: On Cognition and Recognition 
- 
9. 
Recognition as Our Selective Faculty 
2 Sane Simple Philosophical Notes 
3 Recognition of Universals 
4 The Importance of Past Experience: Reality and Nightmare 
5 The Intake of Infonhation by the Senses: Some 
&uantitative Experiments 
6 The Search for Invariants, in Pattern Recognition 
7 
On the Srain as a "Machine" 
Chapter 8; Human Communication: l eeling, Knowing, and Understandine; 305 
1 Communication Is Always an Act of Sharing 306 
2 Signs of Cause versus Signs of Meaning 31 2 
3 The Importance of Ritual 314 
4 Spontaneous Speech. The Extraction of Veaning 318 
5 Human Language and Animal Signaling 329 
6 On Human Communication 334 
Appendix 339 References 344 
Index 365 
American Journal of Computational Linguistics 
ABHANGIGKEITSGRAMMATIK 
JURGEN KUNZE 
Zentralinstitut fur Sprachwisse~chaft 
Akademie der Wissenschaften der DDR 
Akademie-Verlag 
Berlin, DDR 
Reviewed by 
504 pp. 
1975 
KENNETH F. BALLIN 
Department of Linguistics 
SUNY at Buff a10 
Amherst, New York 14261 
JUrgen Kunze establishes his dependency grammar with 
four components. The syntactic is the most important. The 
three non-syntactic components are the paradigmatic component, 
theselectional component, and the assigning component. In the 
first chapter of his book ABHANGIGKEITSGRAMMATIK (Dependency 
Grammar) the reader gets introduced to some of the basic concepts 
useful in understanding the notions explicated later sn. Sub- 
ordination or dependency is introduced by way of a diagram, known 
as a tree, consisting of several connected points. A point or 
node that is connected to one closer to the top of the page is 
subordinate to it. This is called direct dependency. Indirect 
subordinat'ion is when two nodes are connected with one or more 
points in between them. These three nodes comprise a part tree. 
Kunz review 
48 
Obviously there are several part trees which combine to make a tree. 
If the bottom-most node of our lfttle part tree is not superordinate 
to any other point then the part tree is an end complex. 
Every 
node is an en& complex with itself as its only member. 
Once one decides to attach words to these nodes it changes 
from a conne- the dots game to some sort of meaningful diaexam 
T&e 
first step in this change is to bring order to the diagram. 
Since 
language is the object of study here and the language the book was 
written in procedes from left to right, the author has ordered his 
tree from left to right. This type of tree is known as a W tree, 
i.e. Ghere each node is attached to a word. The book deals with 
M trees. These are trees in which the nodes are connected to signal 
combinations (Merkmalkombinationen). A marked tree is one in which 
a11 the connections are subordinate relations on one kind or another 
from a set containing all the kinds of subordinate relations pos- 
sible. 
In making his investigatiods, Kunze has limited his field 
of study to modern day written German. This suffices as for in any 
pure theoretical investigation it is acceptable to assume the ob- 
served language is a set of given sentences. 
The practicability 
of his theory depends on finding a standard of correctness. 
In 
this case tapping the knowledge of a native speaker is of no help 
Four ways are suggested as possibilities for this standard of cor- 
rectness. 
The first is grammatical correctness in which all sen- 
tences are acceptable as long as they function as members of their 
classes, i.e. nouns as nouns, verbs as verbs. Second is a more 
refined gramatital correctness taking the meaning of the verb 
into account. 
Third is the suggestion of a very strict grammar 
Kunz review 
49 
bordering on grammar and semantics including semantic categories 
such as ABSTRACT or CONCRETE. The fourth consideration is a 
semantic grammatical correctness. Though this standard of correct- 
ness is needed to make the theory work, measuring correctness is 
not a major factor. 
There is, says Kunze, a base language and base structures 
that can express semantic and syntactic ambiguities. It is important 
when studying these structures to consider which categories and 
qualities are contained in it. A category is a variable with set 
value ranges, for example, in German, case. Qualities are restrictions 
imposed for ordering, appearance or non appearance of sentence frag- 
ments. Plainly-not all categories and qualities are in every 
sentence of a language. An expansion of the base language leads to 
a simplification of the descriptive system but also costs quite a 
bit as far as analysis is concerhed. 
On starting into the meat of the matter tha author writes 
that in no way can one expect such a simple tool as dependency 
trees to encompass the linguistic relations within the sentence 
that are conditioned through language. This inadequacy is evidenced 
by the following situation. Every grammatical structure has an or- 
dered dependency tree. It is however possible to have two different 
structures represented by the same tree. This is one of the 
principles for the representation of sentence structures using 
dependency trees. Reduction is another principle by which we get 
sentences like 'My friend will bring the book' from sentences 
like 'My friend will bring you the book tomorrow. An additional 
principle removes those nodes which were dropped from the latter 
sentence to arfive at the former This procedure is only permis- 
Kunz review 
sible when the middle steps are somewhat acceptable. 
An interesting concept is introduced by the author. He 
calles it the configuration criterion. It ways that one element 
is substitutable for another it it has all the same grammatical 
properties. 
This concept is used frequently in deciding what is 
dependent on what. 
The fourth chapter deals with the non-syntactic components. 
Paradigmatic categories are established. The categories are 
Gknus, commonly called gender, Person, and Number. Each of these 
three categories is established on the basis of data similar to 
the following example. I saw him. We saw him. These two sen- 
tences are syntactically equal but paradigmatically quite different. 
This illustrates the category of number, the first person singular 
changes to the first person plural. The author splits these 
categories again to account for the phenomenon of endings in 
German. It is possible to have a possessive pronoun with a mas- 
culine ending referring to a female person. Quasi categories are 
also established. These are tempus, modus, and case, and they 
are only quasi categories because they affect other parts of 
speech in a sentence. Kunze constructs a list which enumerates 
the category responsible for their relations, e.g. a noun is 
paradigmatically related to its apposite through case. 
The separation of the paradigmatic from the selectional 
is due to the ease with which the former are presentable. Selec- 
tional relations are more narrowly defined in this case than in 
generative grammar. 
As with the paradigmatic relations there 
are nine selectional relations, five of which belong to the in- 
Kune review. 
ventory and the other four don' t . 
There are demands made on a system of subordinate relations. 
The first of these is that the marked tree should be an adequate 
reeresentation of the syntactic structure of the sentence. Secondly, 
the subordination relations must allow all categories, qualities, 
and relations in the base structure To represent and differentiate 
the paradigmatic and selectional relations that can't be expressed 
through assigning. 
Affectation ways (Wirknngswege) are dashed lines connecting 
two nodes dominat+ed by a third node (see diagram). They represent 
other relations that exist between nodes aside from subordination. 
That these affection ways of both the paradigmatic and selectional 
relations must be represented through subordination relations is 
another demand made on the system. The last demand made is that 
the conditions for the paradigmatic and selective points (Vorgaben) 
must also be represented. 
The principle called the differentiation principle proves 
these last two are met. The system makes this determination by 
using a a knowledge of dependency trees, a fixed inventory of para- 
digmatic and selectional relations, and a fixed language base in 
a way which yields the required relations. 
The last concept developed by the author is that of bundles. 
Kunz review 52 
There are four types of bundles - a simple bundle, an elementary 
bundle, a complex bundle, and a complex implication bundle. A 
bundle is a tree used to represent not a sentence but a set of 
sentences, i.e. trees. In a complex bundle the paradigmatic and 
selective properties need only be given once. 
Chapter 8 is a discussion of some questions that were 
brought out as a result of this theory. 
American Journal of Computational Linguistics 
GLANCING, REFERRING AND EXPLAINING 
IN THE DIALOGUE SYSTEM HAM-RPM 
W, WAHLSTER, A, JAMESON, W, WOEPPNER 
Project: 'Simulat ion of Language Understand1 ng ' 
Germanisches Seminar der Universitat Hamburg 
von-Melle-Park 6, D-2000 Hamburg 13, West Getmany 
SUMMARY 
This paper tocusses on three components of- the dialogue system HAM-KYM, 
which converses in natural language about visible scenes. F~rst, ~t is 
demonstrated how the system's communicative competence is enhanced by its 
imitation of human visual-search processes. The approach taken to noun- 
phrase resolution is then described, and an algorithm for the generation 
of noun phrases is illustrated with a series of examples: Finally, the 
system's ability to explain its own reasoning is discussed, with emphasis 
on the novel aspects of its implementat~on. 
1, THE TREATMENT OF VISUAL DATA 
The natural language dialogue system HAM-RPMu converses with a human partner 
about scenes which either one or both are looking at directly (or have a 
photograph of). At present the system, which is implemented in FUZZY 
(~eFaivre 1977), 1s being tested on two domains: the interior of a living 
room and a traffic scene. 
Since it is assumed that both partners begin the dialogue with relative- 
ly little specific knowledge about the scene, most of the specific informa- 
tion used by the system during the conversation must be, obtained by a pro- 
cess more or less analogous to looking at the scene. we have found it worth- 
while to make the analogy qulte close, requiring the system to retrieve its 
visual data by doing something like casting a series of glances centered on 
various points in the scene. 
Fig. I is a schematic drawing of a section of our traffic scene, showing 
a tree with a parking lot in front of it. How easy is it to recognize the 
various objects in Fig. 1 when glancing at point A? CAR9 and CAR8 will be 
about equally easy to recognize as cars. TREE4 will probably be recognized 
more easily, since it is equally close to point A, and very large, and since 
there are no simi lar types of objects. On the other hand, CAR3 wi 11 be less 
easily recognizable, since it is farther away.  MAN^ is probably too far away 
to be recognizable as a man at all (he is recognizable only from the points 
nearest him, as is shown by the four arrows pointing away from h~m). 
Just this information is stored In HAM-RPM in a separate associative 
network corresponding to point A. In all, there are about a hundred such 
small networks (represented by the small dots in Fig. I), corresponding 
to possible glances at the scene. The statements about the nature of the 
various objests which are recognizable from the point in question are ordered, 
in a way characteristic of the FUZZY programing language, in terms of their 
recognizability, so that they will automatically be retrieved in that order2. 
T.he system's overall structure is described in (v. Hahn et al. 1978) as 
are the goals and methodological principles which guide the research 
wi thin the project. 
2, These networks are implemented as CONTEXTS in the sense introduced by tile 
l anguage CONN I VER. 
Fig. 1. "The man in front ot the tree" 
A simple example will show how the data stored in this way can be used by 
the system. When interpreting the defin~te description the man In front of 
the tree, assuming %hat TREE4 is the one meant, the system enters several 
CONTEXTS in front of TREE4, within each retrieving the internal names of the 
men recognizable from that point. It doesn't flnd  MAN^ until it has entered 
the CONTEXT corresponding to point 0. It then enters a couple more, and, fin- 
ding no further men, assumes that it has found the referent of the definite 
description. Information not only about the respective types of the various 
objects, but also about their other attributes is stored in a similar way. 
Why is it worth all this trauble to make the systan sensitive to the 
recognizability of the various facts about a scene from the various points 
within it? After all, the facts themselves could be stored very straight- 
forwardly. 
Our principal justification is that, for a dialogue system which is 
supposed to communicate effectively with a hunlan partner, the bare facts 
about the scene are less important than the way the partner himself would 
be likely ro perceive them. If only the facts thenlselves are known, intor- 
rnation may be lacking which is essential for the production of a connnunica- 
tively adequate response. For example, the definite description whose inter- 
pretation was just sketched was, strictly speaking, ambiguous, as there is 
a second man in front of the tree whom the system would have considered to 
be the referent of the description if MAN2 hadn't been there. Yet the system 
didn't even notice this ambiguity, since it stopped shortly after flnding the 
first man. 
To be sure, the resolution of such ambiguities could also be achieved 
by giving the system general Information on the recognizability of objects 
for human beings and letting it choose on that basis which of the potential 
referents of the description was the one which the partner was most llkely 
to have intended to refer to. Instead of doing this, we have made the system 
itself a model of its partner, so that instead of referring to a model, it 
only has to 'be itself' or 'act naturally' in order to comnunicate effective- 
ly'. 
In add~tion to the interpretation of ambiguous utterances, there are 
other situations in which this approach can be applied elegantly (Fig. 2). 
SITUAT ION l NFORMATI ON 
1) Interpretation of an 
ambiguous definite 
description 
2) Generation of a definite 
description 
3) Description of a part 
of a scene 
Fig. 2 
Which object the speaker is 
probably referring to 
Which reference points will 
be easy fqr the l i stener to f iad 
Which objects the speaker might 
be ~nterested in hearing about 
Two of the reports (v. Hahn 1978a, 1978b) which have been issued by the 
HAM-RPM group deal with the question of the nature of the relation be- 
tween the dialogue partner model and the human partner in some detail. 
When describing the location of arl object with reference to other objects, 
the systeii~ wi ll usually find a nun~bet of potential reference points; in 
general, it should mention those which are visually easiest for the listener 
to find. This is likely to happen if it itself finds these reference points 
partlculdrly easily. When answering a vague question, such as a request to 
describe what is on the other side of the street, the system will have to 
select among the many visible facts those which the listener might be inter- 
ested in hearing about. In many cases, these will be the visually most sali- 
ent facts. 
2. NOUN-PHRASE RESOLUTION 
Two of the components of HAM-RPM which make use of the visual data are 
those responsible for noun-phrase resolution, that is, the determination 
of the potential referents of a noun phrase, and noun-phrase generat ion, 
that is, the construction of noun phrases to identify objects uniquely. 
The procedures which resolve noun phrases wo~k on the shallow struc- 
ture af the input sentence. Th~s IS what IS obtained after multiple-word 
phrases and idioms have been replaced with canonical expressions, the words 
have been looked up in the lexicon, and a simple morphological analysis 
has been performed. 
A definite noun phrase is recognized within the shallow structure 
as a structure consisting of a definite article, possibly one or more 
attributes, a noun, and possibly a relative clause (Ritchie 1977). In a way 
reminiscent of Winogred's SHRDLU (Winograd 1972), processes involviqg seman- 
tics and pragmatics are activated in HAM-RPM as soon as possible during the 
analysis of the input sentence. 
The noun-phrase interpreter tries to find a unique referent for each 
definite noun-phrase by using the knowledge stored in the conceptual and 
referential semantic networks and performing visual search algorithms. For 
example, the definite description The prcture hanglng to the left of the 
red chair, referring to Fig. 3, is replaced with the internal object-name 
PICTURE1 in the shallow structure of the sentence. This strategy can save 
a good deal of unnecessary processing: if no object is found which satis- 
fies the description, there is no further parsing, but rather fed-back to 
Fig. 3 
the conversational partner. In the .case where more than one potential ref- 
erent IS found, the one mentioned most recently IS assumed to be the ref- 
erent. If none of them has been mentioned recently. the system asks the part- 
ner for further details, assuming, as it were, that he does have some parti- 
cular object in mind. These details take the form of a noun phrase, which 
may be either complete or elliptical. Further attributes of the intended 
object may be specified, it may be characterized in terms of Its spatial 
relations to other objects, or the noun originally used in the descript~on 
may be replaced wi th a more specific one. 
Not all noun phrases, of course, can be replaced immediately w~th a 
specific referent. One such case IS exemplified by the description the chalr 
rrl front of the red cupboard. Appl led to the scene i~ Fig. 3, the noun phrase 
the red cupboard cannot be replaced, because there is more than one red cup- 
board, but it cannot be ignored, either, because there is more than one chair. 
The entire noun phrase can only be interpreted when it is recognized that 
there is only one pair bf objects which stand in this relation to one an- 
other. 
Another case where a definite noun phrase can't simply be repidced 
directly by its referent is the generic description with definite article, 
as in the sentence The cha~ 1s somet-hrng to sit on. Lately we have been 
thinking about what formal features of a sentence might be helpful in recog- 
n I z i ng such descr ipt ions (see Grosz 1976) . 
Two clues which tend to favour a generic interpretation are the absence 
of any referential attribute and the presence of an adverb such as usually 
or normally. On the other hand, a generic interpretation becomes somewhat 
less plausrble if the noun phrase is the object of a local preposition, as 
In on the chair, if the sentence is in the past tense; or if the verb can 
be generally classified as one involving visual perception or spatial re- 
lations. We assume that, no matter how many weak inference rules of this sort 
are incorporated into the system, there will still be some ambiguities which 
can only be resolved by other means, including interaction with the speaker. 
A general goal in this connection is a sort of compatibility between 
noun-phrase resolut ion and noun-phrase generat ion, in the sense that the 
system should be able to understand any kind of noun-phrase that it can gen- 
erate, and vice versa. 
3. NOUN-PHRASE GENERATION 
The method we have developed for the inverse process, noun-phrdse generation, 
is distinguished from earlier approaches mainly in three respects. 
The f ~rst is its use of what might be called a iworst-case-firsti strat- 
egy. The second is the way it takes into cons~deration the ease with which 
the l~stener will be able to interpret the description it generates, when 
more than one uniquely identifying description is possible (~errmann G Laucht 
1976). The third is its use of complex spatial relations to deal with the 
'worst cases', that is, those in which several objects are indistinguishable 
on the basis of their properties alone. 
Let's exatnine a fewexamples of the behavior of the algorithm First, 
two trivial cases. 
Fig. 4 
The street in Fig. 4 has a proper name, and is thus referred to simply as 
Schluterstrasse. The tree is the only one in the discourse world, and hence 
is identified as the tree. The number of interesting possible strategies be- 
comes greater when the object to be described is one of several belonging 
to the same conceptual class. Consider for example MAN1 in Fig. 4. The sys- 
tem looks among its properties for one which distinguishes it from MAN2 and 
MAN3, and describes it as the small man. A similar process underlies the 
generation of the noun phrase the blg man with the umbrella to refer to 
MAN3. 
Note that the system uses redundant labels. This is a consequence of 
the sequential nature of its noun-phrase generation: First, the property 
'big' is found. When the system notices that there is another big man in 
the scene, it looks for a Furkher dist iw*-shing property and finds the um- 
brella. This property would in fact be adequate in itself, but the system 
doesn't attempt to find a minimal characterizing set of attributes. This 
sort of redundancy, which is often found in human beings, saves time both 
in the generation and in the interpretation of definite descriptions. 
HAM-RPM frequently uses negative characterizations of various kinds, 
as, for example, when MAN2 is described as the bry man without an umbrella. 
Now let's turn to some more complex problems of noun-phrase generation. So 
that the pictures don't get too cluttered, we will use examples from a simple 
GREEN 
SQUARE3 
Fig. 5 
domain of geometrical figures (fig. 5). Consider CIRCLE1 in Fig. 5. Note 
that there are two green circles in the scene. The presence of several ob- 
jects which are indistinguishable on the basis of their attributes alone 
is the worst case which can occur. The reason why we have spoken of a 'worst- 
case-f irst' strategy is that the system checks for this case early, rather 
than trying immediately to construct a simpler characterization such as those 
in the last few examples given. 
Informal observation shows that human beings also often notice the pres- 
ence of identical objects in a scene immediately. The only way to distinguish 
these two circles is by reference to spatial relations, for example, the 
green clrcle in front of the red square. 
We may note in passing two ways in which the form of a description may 
be constrained by the form of the question which is being answered. First, 
properties wh~ch have been presupposed in the question should not be men- 
t ioned i n a descr i pt ion. Cons i der the ques t ion Which square is red?. The an- 
swer The red sguare is clearly unacceptable, so instead the system answers 
  he square in front of the green triangle (= SQUARE I i n F tg . 5) . A second 
constraint of this sort is that the system sphould not produce circular 
Fig. 6 
descr~ptions. For example, when answering the question Where is the red 
triangle?, the system should not answer To the left of the green square 
whlch is to the rlght of the red triangle, a l though each ha1 f of thi s de- 
scription is perfectly natural when considered in isolation. 
It sometimes happens that objects chosen as spatial reference points 
In a description in turn have to be identified with the help of other ref- 
erence points. For example, CIRCLE2 in Fig. 5 is described as the green 
circle ln front of the green square which 1s to the left of the red triangle. 
As this example shows, the products of such recursive applications of the 
generation algor~thm can soon become difficult to understand. We have made 
the maximum allowable depth of recursion a parameter which can be set to 
varlous values for experimental purposes. 
Perhaps the most difficult problem in noun-phrase generation is the 
unique identification of an object when there are objects with exactly the 
same prsperties in its immediate neighbourhood (fig. 6). This is a task 
which often causes difficulties even for a human speaker. To set the stage, 
suppose that CIRCLE6 in Fig. 6 IS to be identified. The system first des- 
cribes its position relative to the other circles in its group: the r~ght- 
hand clrcle Then it identifies the group of which CIRCLE6 is a member with- 
63 
In the see-ne as a whole, distinguishing it first fra CROUP2: in ftont 
(ind to tl~c* left and then f ro~n GROUP3 : which 1s to the left of the square. 
Thus 
the complete descr ~ption is The right-hand czrcle in front and to the 
1 tjft wlil cdr rs to tllr left of th~ squnra. 
TO put the point more generally, complex scenes sometimes have a hier- 
archical.struct~re in which groups of s~milar objects serve as units which 
have to be identified in much the same way that objects themselves are. The 
remarks we have made about circular descriptions and recursion depth apply 
on the level of groups as well. 
Concluding this sketch of HAM-RPM's .noun-phrase 
interpreter and gener- 
ator, we would like to stress that all these algorithms are doma~n-independ- 
ent. 
4. EXPLANATION 
Although all of the examples d~scussed up to now have involved some sort of 
description of visible aspects of a scene, HAM-RPM frequently makes use of 
general knowledge and inference rules to draw conclusions. 
For example, the system might be asked Is the park3ng zone tarred?, 
where the parking zone in question, though part of the scene, is hidden 
from view. It would then try to answer the question using approximate infer- 
ences based on fuzzy knowledge (~ahlster 1978), concluding that the parking 
zone might very well be tarred, because a parking zone is in a sense a part of 
a street, and streets, like thoroughfares in general, are usually tarred. 
Inferences which stand on such shaky ground as this one are of limited use 
to the conversational partner unless the system can describe the reasoning 
which underlies them. 
Furthermore, not just any description will be satisfactory: the system 
ought to act in accordance with the following 'three maxims, as formulated 
by (Grice 1975): 
1. Hake your contribution as informative as is required. 
2. Don't make your contribution more informative than is required. 
3. Be relevant. 
Thus, describing an 
inference chain in every detail will not in general be 
cmunicatively adequate, if some of the inferences are essentially defini- 
t I I , aild t~tbllcc conccpt~~il I y tr i v id Only whcn the dialogue partner has 
rt.pcdtrhdly rccl~rcsted deta ils about inferences wi l l i t be sensible to mention 
all ol tllc~i~ 
Now I~bt's looh at the way we have tried to achieve these goals in HAM- 
RPM, uhi~lg the example just given. Three processes are essential. First, 
while the ruasorilng is being performed, a sort of trace of the Inference 
process is stored in a separa'te date base called INFERENCE-MEMORY. Second, 
after an explanation of the conclusion has been requested, this part of 
memory is traversed to find those of the assumptions used which are on a 
conr~~uo~catively appropriate level of detail. Finally, these assumptions are 
expressed in natural language. 
An cbbent~al role in the first two of these phases is played by the 
ineta-knowledge associated with each inference rule which is available to 
the systenl. As you can see from the two inference-rule definitions in Fig. 7, 
Apply the control knowledge coded in TRACE-PROCEDURE-DEMON7 
Don't use instantiations of premises w~th a degree of belief less than 0.3 
The degree of uncertainty of this rule is 0.5 
RULE. If youwant toshow (X IS Y) 
show that (X I SA 2) 
and (2 IS Y) 
META-KNOWLEDGE : 
Apply the control knowledge coded In TRACE-PROCEDURE-DEMON7 
Don't use instantiations of premises with a degree of be1 ief less than 0.4 
The degree of uncertainty of this rule is 0 8 
RULE. If you want to show (X IS Y) 
show that (X IS-PART-OF Z) 
and (Z IS Y! 
Fig. 7 
she 
but 11 piccc oF il~ctd-hnowludgc conccrns the degree ot uncertainty asso- 
ciated w~th the rule. The n~ost interesting piece of meta-knowledge ifi fh~s 
Situc7tion is th~i specitication of a particular FUZZY procedure demon. The~~e 
deil~ons iilrlorcc during the application of an inference rule global control 
regi~llcs speci tied by the programmer (~e~aivre 1977). In particular, one of 
the thinqs done by TRACE-PROCEDURE-DEMON7 is the storage of the reasoning 
bteps in INFERENCE-MEMORY. 
Suppose now that the assumptions at the top of Fig. 8 are represented 
in serilantic networks. Applying the two rules in Fig. 7 to them, the system 
bullds up the goal tree1 in Fig. 8. The internal trace which is built up 
by the procedure demon is shown at the bottom of Fig. 8. Note that the 
( (PARK I NG-ZONE I S-PART-OF STREET) 0 7) 
( (STREET I SA THOROUGHFARE ) 1 
( (THOROUGHFARE I S TARRED) 0.7) 
/ 
( (PARK1 NG-ZONE I S-PART-OF STREET) 
TARRED) . o 5) 
( ( (PARK I NG-ZONE 1 S TARRED) 5) 4 (PARKI NG-ZONE I S-PART-OF STREET). 7) ( (STREET I S TARRED ) 5) ) 0.8) 
( ( (STREET I S ~AARED) 5) (STREET IS THOROUGFAAE) I ) ( (THOROUCHFARE 1 S TARRED) 5). 0 5) 
Fig. 8 
entries in the inference memory are ordered in terms of the degree of un- 
certainty of the executed inference procedures. This means that the most 
uncertain entries will bementioned first in the explanation, and the mst 
1) The confl ict-resolution strategy which is used is one which favours more 
specific ones. 
tl.ivi,,l onuLl ~)~obdbly not rjt cll I. Ttli's t-rtlccts our hypothesis that degree 
ot uncrbl tairity is the irlost important factor deterir~ining the relevance of a 
-,tt8p ill 'In inl'crlv~cu ~h;lin, 4s tar as justificatilon of the conlusion is con- 
cer-ncd 
Our approach to explanation is distinguished from the previous efforts 
ill' Witlograd in h~s SHRDLU system (winograd 1972) and of the MYCIN group (~cott 
ct dl. 1977). In SHRbLU each theorem calls the fuhctions MEMORY and MEMOREND 
explicitly, which manipulate the inference memory. We have improved over this 
by int-ting the managelsent of the inference memory Into a higher process, 
which controls all reasoning processes. The structure of the inference rules 
thembelves is therefore not obscured by the presence of trace commands. Our 
dpproach generalizes the corresponding features of MYCIN, in which the con- 
ceptual co~nplex~ty of a rule is a logarithmic function of its certainty fac- 
tor and the goal tree is traversed in steps whose size is specified by a 
nulnerica l argument of the WHY command (~avi s et a1 . 1977). 
This approach is also related to recent research by Davis in his 
TE IRES IAS system (Davis 1977) and Eussman in his AMORD (~e Kleer et a1 . 1977) 
and EL (Stallman & Sussman 1977) projects, in which general problems of an 
explicit control of reasoning are explored, in that it is based on an explic- 
it representation of control and meta-knowledge, which typically is 'hidden 
away' In the interpreter and therefore is inaccessible to the inference sys- 
tem. 
The explanation facility of HAM-RPM is far from being complete. Ulti- 
mately, the system must understand exactly what the dialogue partner failed 
to comprehend, 
ACKNOWLEDGEMEN' 
This research IS currently being supported by the Deutsche Forschungsgemein- 
schaft. 
American Journal of Computational Linguistics 
A CRITICAL LOOK AT A FORMAL MODEL 
FOR STRATIFICATIONAL LINGUISTICS 
Alexander T. Borgida 
Department of Computer Science 
University of Toronto 
Toronto, Ontario M5S 1A7 
SUMMARY 
We present here a formalization 
of the straiificational model of linguis- 
tics proposed by Sampson C131 and inves- 
tigate its generative power. In addition 
to uncovering a number of counter- 
intuitive properties, the results present- 
ed here bear on meta-theoretic claims 
found in the linguistic literature. For 
example, Postal [llj claimed that strati- 
ficational theory was equivalent to 
context-free phrase-structure grammar, 
and hence not worthy of further interest. 
We show, however, that Sampson's model, 
and several of its restricted versions, 
allow a far wider range of generative 
powers. In the cases where the model 
appears to be too powerful, we suggest 
possible alterations which may make it 
more acceptable. 
1. Introduction 
Linguistic theories are at least partially interested 
in presenting the regularities found in natural languages. Given 
the current dominance of the Transformational Generative (TG) 
school in the field of linguistics, it seems necessary for 
theories competing for attention to possess a formal model, In 
addition to the advantages normally derived from presenting 
results through a formalism, such as precision, succinctness and 
verifiability, one can also comment on the veracity of meta- 
theoretic claims. It was using such formal arguments that 
Chomsky and his collaborators demonstrated the inability of 
finite automata and of context-free grammars to describe all 
natural language constructs. Similarly, the formal work of Peters 
and Ritchie [8,91 was important in uncovering inadequacies of two 
notions of TG theory namely, the "recoverability of deletions 
condition" and the "universal base hypothesis". 
Finally, since many generative linguists want grammatical 
theories which characterize natural languages, they fault any 
theory which is .too powerful" in the sense of being able to describe 
languages which clearly cannot be natural languages, such as non- 
recursive sets. Furthermore, computer scientists working on 
natural languages will have to give in the future more considera- 
tion to the work of linguists, especially on "exotic" languages, 
in order to be able to observe a wider range of phenomena. Such 
access will be facilitated if the formalismsin which the grammaTs 
are prese'nted lend themselves to computer implementation for 
purpose$ such as parsing, testing, etc. This entails, among other 
things,that linguists should avoid as much as possible features 
which make their grammars generate non-recursive sets, and hence 
it is one of the purposes of the present paper to point out such 
features and discuss possible ways of avoiding them. 
In this paper we will discuss one model proposed for the 
stratificational theory of linguistics. This theory, advanced by 
S. Lamb, H.A. Gleason Jr. and their collaborators (C51,[61,C71), 
advocates that langdages be described in terms of several sub- 
systems, known as strata. Each stratum has its own set of units 
and a tactics specifying the tlcorrectfl ("all~wable'~) structures 
that stratum. specific grammar might for example have strata 
corresponding roughly to semantics, syntax-morphology and phonology, 
although this is by no means standard. Furthermore, the strata 
are linearly ordered as levels, and there is a realization relation 
which connects adjacent strata by attaching to every well-formed 
structure on one stratum, zero or more accompanying structures 
on the adjacent strata. Note therefore that a particular 
utterance has simultaneous expression on each stratum. 
In this paper we examine the formal model for stratifi- 
cational linguistics proposed by Sampson (L131). This model uses 
rewrite grammars G1,G2, ... to describe the tactics, while the 
realization relation is essentially a rewrite system R acting as a 
transducer between the languages of-the tactics. More specifically, 
realization connects adjacent tactics G 
and Gj+l 
j 
by matching 
sentences u in the language generated by G 
with those sentences 
j 
v in the language of Gj+l which can be derived from u by using 
rules from R. 
An important property of the linguistic realization 
relation is the fact that' every structure on some stratum can 
have only a finite number of llrealizates" on the next stratum. 
This means that the rewrite system R must be constrained so that 
it has no recursive symbols. Such a rewrite system will be called 
acyclic. 
We investigate here the effect of acyclic rewrite systems 
acting as transaucers on axiom sets, varying the type of the 
derivations and rules allowed. 
We prove in this paper that regular languages are 
closed under transduction by acyclic rewrite systems, but that 
the linear context-free languages are mapped onto the recursively 
enumerable sets. This implies that stratificational grammars 
with non-selfembedding ta~ctics would be too weak while those with 
even one context-free tactics would be too strong. If the 
realization derivation is restricted to be in some sense "left- 
most", then we show that the transduction can be performed by 
a finite,state device known as an a transducer. Furthermore, if 
productions with null right-hand sides are not allowed in an 
acyclic rewrite system then all the derivations can be made left- 
most. 
This provides one possible method of restricting the genera- 
tive power of acyclic rewrite systems. 
By deriving a recursive characterization of the languages 
generated with n-strata in terms of (n-1)-stratal languages, we 
can show that if the realization is restricted to being leftmost, 
then the languages described are homomorphic images of the inter- 
sections of the languages generated by the tactics. In particular, 
this means that we can find natural families of stratificational 
grammars which generate far example the sets recognized in real 
time by nondeterministic multitape Turing machines. This result 
partially confirms a hitherto unproven claim by Sampson, and 
discredits Postal's Clll classiciation of stratlficational grammars 
as just another variant of context-free phrase-structure grammars. 
Finally, we investigate the use of ordered rules in 
linguistic grammars and prove that in several models they allow 
the generation of sets which are not even recursively enumerable 
a clearly unsatisfactory situation. 
The remainder of the paper is structured as follows, 
In Section 2, we present the formal definitions and notation to 
be used, including the formal model for stratificational grammars. 
In Section 3, we examine the properties of "acyclic rewrite systemsw, 
which form the principal novel component in our definition of 
stratificational grammars. We then return in Section 4 to examine 
the generative power of st'ratificational grammars and relate the 
results to linguistics. 
2. Definitions 
We repeat here some important definitions from (121, and 
assume that the reader is familiar with the other basic notions 
of formal language theory. 
+ 
A vocabulary V is a finite set of symbols, and we use V 
to denote the set of all nbn-null strings consisting of symbols 
* 
from V; using e to denote the null string, we also define V to be 
A rewrite system RW is a pair (V, R) where V is a vocabulary 
and R is a finite set of rules (productions) of the form u + v, 
+ * 
where u E V and v E V ; u is known as the left hand slde of the 
production (Ihs .) and v is its right hand side (rhs .) . 
+ 
A word x E V is said to directly derive or generate -- in R 
* 
another word y E V (denoted by x =>R Y) iff there exist words 
U,V,W,Z such that x = wuz, y = wvz and u + v belongs to R. 
Let =+>R 
be the transitive closure of =>R, and =*> its transitive reflexive 
R 
closure. A sequence of words w1,w2,...,~, such that 
w1 =>R w2 => . . . => w is said to be a (free) .R-derivation (or 
R R n 
simply a derivation) of w from w 
n 1' 
* 
Given a rewrite system RW = (V,R) and a subset AX of V 
the language generated by R from axiom set AX with free derivations 
is defined to be the set ~(AX,RW) = {wlueAx, u=*> w). 
R 
Given the rewrite system RW = (V,R) , define the domjnance 
relation < on V x V by: d<b iff xby + udv is one of the 
productions in R (for some strings x,y,u,v) or if there exists 
some c in V such that d<c and c<b. Then RW is defined to be 
acyclic (abbreviated a . iff the relation < is anti-symmetric 
and anti-reflexive. 
If u + v is a production in a fewrite system, it will be 
called a null rule if v is the null string e, and it will be 
called context-free if 11.11, the length of u, is 1. 
A rewrite grammar G is a quadruple (N,T,S,P) where N and T 
are the sets of nonterminals and terminals respectively, 
distinguished nonterminal and f = (NUT) P) is a rewrite system. 
In this case, 
if =*? 
w then this is called a G-derivation, or 
a derivation in G, and the language generated by,G,denoted by 
* 
L(G), is defined to be the set (tl~ =*> t in G, t6T . We assume 
the reader is familiar with the terminology of type 0 (recursively 
enumerable or RE), type 1 (context sensitive) , type 2 (context 
free) and type 3 (regular) languages, and corresponding families 
of grammars and automata. A type 2 grammar will be called linear 
if all its productio~l~ are of the form A + aBb, where A,B E N, 
a,b E T u {el, and will be called selfembedding if for some AEN 
there is a G-derivation A =*> uAv where u and v are not null. 
New languages can be obtained from old ones through such 
set operations as union, intersection and concatenation. 
One can also define- mappings over strings and then extend 
them to sets of strings in the obvious way. One such mapping 
is the substitution s which associates with every symbol b of 
some alphabet T, a set of words ~(b) over another alphabet Tt; 
&fining s (xy) = s(x)s (y) and s(e) = e, a substitution can be 
extended to strings. If the sets s(b),are regular, finite or 
e-free then s is said to be regular, finite or e-free respectively; 
if s(b) contains a single word then s is called a homomorphism, 
and the braces for sets are dropped. 
A homomorphism h can also 
be e-free, or it can be length-preserving, if lh(b) 1 = 1 for all 
symbols b. 
If L is a family of languages then we use H ( L ) and 
0 
H ( L) to represent the families of languages obtained from 
elements of L through e-f ree homomorphisms and ~omomorphisms 
respectively. 
One final operation on strlngs is reversal defined by 
&v(b) = b if lbl ( 2 and Rev(xy) = Rev(y)Rev(x). 
Ohe can also use automata to perform mappings between 
0 
strings. The a-transducer M = [K,T ,T ,k F,) is an extension 
12 
of the finite automaton, where T1 and T2 are the input and output 
'dr * 
alphabets, and T is a finite subset of K x T1 x T2 x K (the transi- 
* * 
tion set) . The relation I- is defined on K x TI x T2 by the rule 
(k,uv,z) I- (kf,v,zx) if (k,u,v,kT) E T. The output of M for 
input-word w is one of the strings in the set 
0 * 
{zl (k ,be) I- (k,e,z), keF). An a-transducer is said to be 
e-output free if for an$t (r,u,v,s) in T, the string v cannot be null. 
A collection of languages A is said to be closed under 
the operation o if a(L) E A whenever L E A. 
A (full) trio is a 
family of languages containing at least one nun-empty set, closed 
under e- free homomorphism (arbitrary homomorl?hism) , inverse 
homomorphism, and intersection with regular languages. 
Flnally, omitting detailed justification (see C 33) , the 
following formal definition captures the essential aspects of the 
notion of stratificational grammar, as presented by Sampson C 131: 
Definition An n-stratal rewrite grammar (n-RSTRAT) is a 5- tuple 
RST = (n,TCT,RLZ,V V ), where VC and VE are the set of "content 
C' E 
units1! and tlexpression unitst1 respectively, TCT = (G19G2, . . ,Gn) 
is a vector of n rewrite grammars, and RLZ = (RO,R1, ..., R ) is a 
n 
vector of ~+1 acyclic rewrite systems. The transduction perfo~med 
+ 
by such a grammar will be defined by T-RSTRAT(RST) = { (u,v) lw0=u~vC, 
* * 
W 
n+l 
=vcVE, there exist w c L(G.) such that w = >w~+~ 
j 
via R - 
J j j 
derivations for j = 0,1,. .., nl. Its language is described by 
In this formal model, the grammar is thought of as 
transducing "meaningu into vtsound" in the following manner : 
starting with a string of "content units" (expressing the meaning 
of *an utterance), the realization rewrite rules are repeatedly 
applied until a string of "expression units" is obtained. The 
realization derivation is constrained by the requirement that for 
each tactics there exists an intermediate stage in the realization 
derivation which conforms to the tactics specifications (i .e. 
belongs tb the language generated by the tactics) . 
The above 
formalism is based mainly on Lamb s version of strati Eicationa~ 
linguistics ; an alternate approacki, closer in spirit to Gleason's 
model, is presented in C31. 
3. 
Generative power of acyclic rewrite systems 
To begin with, we remark that the formal definition of 
s tratif icational grammars in C 13 1 allows in the reali zation sys tern 
rewrite rules with null left-hand sides (i .em rules of the form 
e + u). 
Unfortunately, such rules could be applied to some string 
an arbitrary number of times. 
In our stratificational model, 
this would result in any string having an infinite number of 
realizates . Furthermore, rules of the form e + u can also be 
used to establish context-free dependencies in strings generated 
even from singleton axiom sets. For example, if R = ({c,d),{e+cd}) 
* 
then (el R ) = {w~wE(c,~} , w has the same number of "c" and 
"dl! symbols}, which is known to be a non-regular centext-free 
wage. The phenomena described above do not appear to have 
linguistic equivalents, and run counter to the s tratif icational 
philosophy which envisages only flnitely many realizations for 
any structure . 
As it turns out, in practice rules of the form e + u 
are only required to introduce in the realization derivation 
syntactically determined elements, such as "do" in questions. 
Such inserti~ns need however be performed only once, at the end 
of every realization derivation between two tactics. Therefore 
they can be accomplished through normal acyclic rules if each 
e + u in R is replaced by rules v + uw and v -+ wu for all v + w in 
R. For this reason, we will continue to use the definition of 
rewrite systems which only allows productions with pon-null left- 
hand sides. 
We next investigate the effect of a.r. on simple types of 
axiom sets. 
Theorem 3.1 Let AX be a regular set over alphabet T and let E be 
some alphabet disjoint from T. If RW = (V,R) is an a.r. then 
* 
~(AX,RW) n E is also a regular set. 
Proof Let G = (N,T,S,P) be a type 3 grammar generating AXy and 
without loss of generality assume that N 
n VR is empty. Further- 
G 
more, normalize R so that all its rules are of the form a+bc, a+e 
or bc+d. This can be accomplished in a 3-step process: first, 
* 
replace rules of the form u + abv (a,beV, u,vd ) by rules u + aa, 
a -+ bv where a is a new symbol; repeat this until all rules have 
rhs. no longer than two symbols. Next, replace rules of the form 
- - 
abu -+ v by ab + a, au -+ v, until all lhs. of rules are at most two 
symbols. Finally, eliminate rules of the form a + b by adding to R 
a rule y + zbzl whenever y + zazl is in R. 
Our goal is to produce a type 3 grammar such that R- 
derivations are Mprecomputedu in its productions, For example, 
if the grammar G originally had productions X+aY and Y+bZ, while 
R contained the rule ab+d, then the final grammar would contain 
production X+dZ. 
Fbr this purpose, consider the following iterative construo- 
tion: 
INITIALIZATION: Eet GI be G; let T9 = T u VR. 
CONSTRUCTION 1: For every integer i, given grammar Gi = (Ni,Ti,SG,Pi), 
construct from l't a type 3 grammar Gi+l 
- 
- T' 9SG ,Pi+l) as 
follows: 
for every aaTi, let P(i.a) be the set of all productions 
in Gi which have the symbol ??a1' on the rhs.; 
to begin with, let Pi+l contain P., 1 and Ni+l contain Ni; 
IF b+cd is a production in Ry THEN for every A+bB in P(i,b) , 
ADD ta N,+l a nonterminal [A;B;b+cd], and ADD to P i+l 
productions A+c[A;B;b+cd~-and [A;B;b+cd]+dB; 
IF b+e is ih R, THEN for every A+bB in P-(i ,b) ADD production 
A+B tp Pi+l; 
IF bc+d is in R, THEN for every pair of productions A+bB 
in P(i,b), and C+cD in P(i,c), ADD to Pi+l the new production 
A+dD if B=>*c in Gi; 
END ; 
Suppose that we were able to establish that 
From the construction it is easy to see that P is always a sub- 
i 
set of Pi+l (and hence L(G.) c L(Gi+l)), and if 
1 
G =G for some index m 
m m+ 1 
(i.e. no new productions are added to G in Construction I), then 
m 
G would be equal to G for every j > m. 
I m 
But, if such an m exists then L(L(G);R) = U L(G.) = L(G ) 
i=l 
1 m 
and G is the type.3 grammar we are looking for. Therefore, it 
m 
remains to establish equalities (1) and (2). 
To,prove (I), we first define a new type of derivation 
("singl&, left-right pass") relation "=o>~ " as follows: 
=Q='~ 
v iff there exists integer n such that for j = l,...,n 
x + y. is a rule in R and z is some string with the property that 
J J j 
u = z x z' ..'x z and v = z y z ...y z 
0 1 1' n n 0 11 n n 
(if n = 0, then u = v). 
We then claim that 
L(Gi+$ = Iw(3 veL(G.) such that v =o>~ W) 
1 
This, equaLi ty can be demonstrated by straightforward inductions 
on, res-a, SA~ ~hengt-r sd d~&v;n.ric~ 3% e, 
i+l ' 
and the 
integer n appearing in the definition of In both case% 
the Important points are that if A => bB in G (A,B€Ni, bcT.) 
i 1 
n 
then either A => bB in Y~+~ (by step 2 in Construction 1) or A => uB 
if b + u is in R (by steps 3 or 4); and if A => bB =*> bc => bc~ 
in G 
then A => dD in Gi+l in case bc -+ d is in R (step 5). 
1 
We are now in a position to prove (1). First, suppose that w 
belongs to Z(L(G),R) and w was obtained from u E L(G) = L(G~) in 
an R-derivation with n steps: u = u1 => 
- - 
R U2 ->R 
... -> R u n => W. 
If we note that for any strings x,y x =>R y implies x =o> 
y 
R 
thenby (3) we have for i = 1,n thatu. c L(G.1. But then 
1 
00 
1 
w = un must belong to L (G ) , and hence to U L (G. ) . Conversely, 
00 
n 
i-1 
1 
if to 
" L(G.) then there must exist an index m such that 
1-1 a 
w c L(G ) Using (3) it is then trivial to prove by induction on m 
m 
that there exists v E L(G) such that v = vl =o>~ v2 =o> .. *=o> 
R R Vm 
= W 
for some'v E L(G.) (1 = 1,m. But in that case w E &(L(G),R) 
i 1 
because by definition x =o>~ y implies that x =*> y for any 
R 
strings x,y. This concludes the proof of identity (1). 
To prove (Z), one night try to demonstrate that the 
construction halts after some precomputable number of steps. 
This 
approach unfortunately runs into the following problem: 
the 
addition of a new production to G in step 4, allows new pairs of 
i 
variables B ' and C ' to be connected by a derivation B '=>*c 9 this 
may aPlow new production A'+dD1 to be added to Gi+14 in step 5, 
which in turn may eventually allow step 4 to add a new rule to 
for some 
The above compels us to look for an alternative prod of 
(2): exhibiting a grammar GO such that every Gi is a subgrammar 
of GO. This would mean that the increasing sequence of grammars 
G1,GZ, - is bounded above, and hence converges to one of its 
elements. 
To construct GO, remember that by definition of R there 
is an anti-symmetric relation < on V 
R' 
Using this, we assign to 
every symbol in V and every production in R a unique index number, 
according to the following algorithm: 
INDEXING ALGORITHM: 
1. I(b) := 0 for every beV such that there is no daV and d>b; 
2, FOR i=O to IvI DO WHILE not all symbols have an index; 
IF I(b)=i 6 b+cd is in R, THEN I(c) := I(d) := i+l and 
I(b+cd) := i+l; 
IF I(b)=i 6 I(c)si 6 bc+d is in R THEN I(d) := i+l and 
I(bc+d) := i+l; 
PF I(b) = i 6 I(c)ai 6 cb+d is in R THEN I(d) := i+l and 
I(bc+d) := i+l; 
IF I(b) = i and b+e is in R THEN I(b+e) := i*l; 
END 
END 
By the acyclicity of R, the above algorithm produces a unique 
value for every symbol and production. Suppose the highest index 
value assigned is n. Then GO will be c~nstructed from G by repeated 
modification in n passes through the following: 
CONSTRUCTION 2: Let GO = (N',T' ,s,P') be G initially; 
FOR i=l toqn DO 
* in the i-th pass, add to GO all possible productions 
representing derivations by index i 
rules */ 
I 
1. For every symbol Id' in VR such that I(d) = i, 
let P(d) be the set of all productions currently 
0 
In G , with 'd' on the rhs.; 
2. IF b-tdc (or b+e) 1s in R and has index i (iff 
I (b) i) , THEN alter GO in exact.1~ the same way as in 
steps 3 (or 4) of CONSTRUCTION 1; (except that 
PO and NO are used instead of Pi+l and N 
) . 
i+l 
3. IF bc+d is in R and has index i, THEN for every pair 
of productions A-tbB i.1 P(b) and C+cD in P(c), ADD 
to PO the new production A+dD (whether - or not 
Note that in the i-th pass the only prbductions added to GO 
have on the rhs: a terminal symbol of index strictly greater than 
i. 
Therefore, in successive passes through the loop after the i-th 
one, P(d) remains unchanged for all symbols I1d" with L(d) i. 
Furthermore, the output GO remains unchanged if passed through 
CONSTRUCTION 2 a second time. 
Secondly, note that if some grammar K remains unchanged by 
CONSTRUCTION 2 then it does so through CONSTRUCTION 1 as well, 
because every rule in R is eventually considered in CONSTRUCTION 2, 
and in each case at least those productions which would have been 
added by CONSTRUCTION 1 are added by CONSTRUCTION 2. 
0 
Therefore, since G is a subgrammar of G , Gi will be so for 
every i greater than 1, and the proof is completed. 0 
Increasing the range of sets from which we choose the axiam 
sets, we obtain the following: 
Theorem 3.2 Let G = (N,T,S P ) be an arbitrary type 0 grammar. 
G' G 
Then there exists a linear context-free language LING+ and an a.r. 
R (which is dependent only on the set N u T), such that 
0 
Proof (Notational convention: let V = N u T, and if E{-,-,v) then 
use V to represent the set {&la~V}, and - 
- alw *aj 
if w = al..-.aj .) 
It is known (Cll) that there exist two linear context-free 
languages L1 and L2, as well as a homomorphism h, such that 
L(G) = h(L1 n L2). We have constructed (C3'1) pairs of new such 
laqguages: 
+ 
v . €VG $Rev (v 
) = v 
i-1 
for i = 1,. . . ,m-13 
1 m+i-1 
and 
- - 
a' 
- - IUN N V 
+ 
L2 = {%wla.. .w %$%w~+~ %...p ln>O, w.aV for icn, 
n 2n 1 
and Rev(w ) => w 
G n+i+l 
for i = 0, ..., n-11 
n-i 
In this case, the homomorphism h is defined as h(;) = x if xeT, null 
otherwise. Observe that L1 is dependent solely on the vocabulary 
V and it only checks whether the strings around the central 
.I %sG%$%' are mirror images of each other. But the fallowing 
rewriting system does exactly the same job, and, in addition, 
performs the homomorphism h : 
v 
Ro 
= (fS B$Fe, %Fey G+e for all XEV~, a+a for aeT) 
G 
* 
Then T 
n de (L~,R~) = h(L1 n L2) = L(G) and by observation it is 
0, 
clear that R is acyclic. 0 
This result is surprising, especially from a linguistic point 
of view, and demonstrates the power of acyclic rewrite systems. 
Since it is undesirable that linguistic mechanisms be so powerful, 
we will attempt to put bounds on them. One way to do so is to 
restrict the places where steps in derivations can occurl. 
Essentially, in a k-leftmost derivation there is a k-symbol 
wide "windoww on the derivational forms where rewriting can occur, 
and this window is only allowed to move to the right. 
Definition 3.1 Let wo => w => ... 
1 
=> Wn be an R-derivation, where 
for i = 1, ..., n productions ui+vi are used to obtain w 
= x.v.u 
i rii 
from w~-~ x.u.y.,,(x 111 i ,yi,wicv--), For any integer k, this is said 
to be a k-Jeftmost derivation if for all i = 1, . . n-1 there exist 
* 
s and t in V such that xi = t.s with 1s.l 5 k a~d 1t.I.s iti+l 
i i 11 1 1 
I 
This definition of R-derivations gives rise to the new language 
* 
%(Ax,Rw,~-Id) = Iw ~XEAX, x=>~w in k-leftmoht derivation). 
Theorem 3.3 Given an a.r. RW = (V,R), there exists an a-transducer 
OR such that $ (AX,R,k-ld) = $$AX) . 
Proof By the acyclic nature of the rules in R, any string of 
t 
length k can be rewritten into a string of length at most kd where 
d 5s the length of the longest rhs. of a rule in R, and t is the. 
nufiber of symbols in V. Therefore, if we define a Turing machine 
transducer O which simulates on its wofking tape k-leftmost R- 
R 
derivations, then it need have only a bounded, f inite-length tape. 
But this can obviously be kept in a finite memory, and hence OR 
can be made into an a-transducer, 0 
Theref~re, leftmost carl3traints on R-derivation6 lead to a 
much more rest,ricted version caf rewrite ~ystems because all trios 
(in particular LINEAR-CFL) are closed under a- transduction. 
A second method of bounding the power of a.r. is to restrict 
the .form of the productions allowed in R. 
Thporem 3.4 If RW = (V,R) is an a.r. with no null productions, 
then for every axiom set AX, ~(AX,RW) = oe(AX,RW,k-ld) for some 
Proof Suppose R has r rules involving t symbols and let c be 
the length of the longest lhs. of a production. Now observe 
* t 
that if v =>R w, then no symbol in w can have more than k=c .~ncestors 
in v, all of which must be adjacent in v (i.e. the presence of a 
symbol in w can depend only on the presence of at most k adjacent 
symbols in v). Thi's value of k can be obtained as follows: since 
the rules are acpcllc, new symbols must appear after every application 
of a production and hence every symbol in w can be the result of 
applying ae most t.productions; since each of these uses at most c 
t 
symbols as context, we get the value of c . The adjacency require- 
ment comes ffom the condition that there be no e-r-ules in R. 
* 
Consider now some R-derivation tXv =*> t: (t,?,~e~ , XEV), where 
no symbol in t is rewritten, but X is. IVe will prove by induction 
on the length of the derivation that there is an equivalent k- 
leftmdst derivation. 
Basis. If the derivation has 0 6r 1 steps then it is 
clearly k-leftmost. 
Inhction step. Break up the dfiivatiom into steps, to 
see where X is rewritten: 
where we used rule Xw A'Yz in stcp 0. 
Now find the 1as.t production in @ which produces only 
non-ancestors of Y. 
- 
(a) 
11 thsre is no such production, then by wr opening 
remarks @ must be k-leftmost, and hence @ can be made k-leftmost 
by induction. 
(b) 
Otherwise, suppose that the last such rule was a+p. 
Then 
we claim that the derivatiod in further detail is 
N 
tXv =*> tXu cry => tXu pu =*> tXh* 'pu => tY zu pu =*> t? 
0 2-@ 2@,e = z'@ 2@ 
The signifitant part of this chin is that no prdduction in @ 
affects the string pu2, and this is true hy our choice of a+p 
as the last production generating non-ancrstors of Y, hence of 
Yz, and the necessary contiguity of ancestors. But now note that 
step @ can bc ~ostponed to yield the following reordered 
derivation: 
By repeating the construction in part (b) on @ = @@ 
this time (instead of @ @@ ) we will eventually (by a 
second induction, if desired) achieve case (a), and thus 
complete the proof. # 
Note that in the above proof we had only excluded the use 
of null rules (the symbol p could not be null) so that other 
productions with left-hand sides longer than right-hand sides 
are still allod in R. 
Finally, we include for completeness the following result 
whose proof :is trivial.. . 
Propositiori 3.5 Let RW = (V,R) be an a.r . which has only 
context free rules. 
Then there exists a finite substitution sR 
such that for every axiom set AX, &(AX,RW) = sR(AX). 
# 
4. Stratificational Grammars 
We now return to the notion of stratificational grammar which 
led us originally to consider acyclic rewrite systems. To begin 
with, note that the original definition of n-RSTRAT grammar has 
*o constraint on the derivations occuring on the tactics, while in 
practice linguists appear to view the derivations as- being leftmost 
(1. e. the leftmost nonterminal is the one rewritten) . Therefore, 
throughout the following discussion we will examine the 
differences arising out of this variation. 
First, we present a recursive characterization of the 
n-RSTMT languages. For this purpose, define the language generated 
by a 0-RSTRAT grammar RST' = (0, () , (R') ,VC ,VE) as L.-RSTRAT (RST') 
* 
&($,R*) n VE. 
Then the following theorem is an obvious conse- 
quence of the definition of L-RSTRAT: 
Theorem 4.1 If RST = (n(G G ),(Ro, ..., 
n 
R-) ,VC,VE) is an 
n 
n-RSTRAT gramman, and TOP (RST) is the (n- 1) -RSTRAT grammar 
(n-l,(G1, ..., G' ),(Ro,...,R 
n- 1 n- 1 
) ,.V T ) , then L-RSTRATCRST) = 
C9 n 
Using Theorems 4.1, 3.1 and the known closure properties 
of the regular languages, it is easy to see that if all the tactics 
G1: 
..., GI of an n-RSTRAT grammar arc non-sclfcmbcdding then the 
n, 
stratiTic-ational grammar can generate only a regular language. 
On thc other hand, as soon as one or the tactics is a/llowed 
to bc of type 2 and selfembcdding, then by Thaorcms 4.1 and 3.2 
the.RSTRAT grammar can generate an arbitrary RE set. Evermore 
surprisingly, this can be accomplished using a "universal 
realization relationM, mea-n-ing that to obtain any RE set we need 
only vary the tactics, not the realization rewrite system. This 
situation is similar to that found for TG in C91, where the 
transformational component can be varied while the base grammar is 
kept fixed. 
Therefore, in this stratificational model there seems to 
be no alternative between the insufficient descriptive power of 
finite,autornata and the excessive poker of arbitrary Turing 
machines. These results hold even if the derivations on the 
tactics are cons trained to be leftmost. We must therefore search 
for further limitations on the realization process. In section 3 
we considered several possible ways of d6ing this, namely 
eliminating null or context-dependent rules, or making the 
realization derivation leftmost. In linguistic grammars there is 
a clear need for context-dependent realization ~ules, hence these 
cannot be eliminated. Although in Sampsonts model nu11 realiza- A> 
tion rules appear to be needed (more on this below), it is 
possible to envisage alternative models which avold them. By 
Theorem 3.4, the absence of null rules is equivalent to rest~ictl%~ 
the realization derivatidn to being k-leftmos t. Furthermore, 
based on current linguistic literature there appears to be no 
objectiofi to limiting the realization to being k-leftmost. 
Therefore, we will examine the generative power of n-RSTWT 
grammars under rhi-s cans traint . 
Theorem 4.2- I6 STR = (n,CG1,.. .,Rn),(RO, . . .,R n ) ,VC,VE] is an 
n-RSTRAT grammar with realisation derivations restricted to be 
k-leftmost for some, integelr k, then there exist homombrphism h 
and languages L1,. . . , L such that for i = 1,. . . ,n L(G.) is of the 
n I 
same type1 as.Li, and L-RSTRAT[STR) = h(L1 n...n L ). 
n 
Proof The proof is based on a number of results aboutjtrios, 
which we summarize here from [4]: 
(a) For i =, 1,n the families of languages of the same 
type as L(G.) are trios. 
1 
6 1 
If L is a trio then H(L) is a trio and H'(L) is a full trio. 
(c) 
If LIP-mb,Ln are trios then H(H(L~ n.. .n ) n L n ) is 
a trio and it is equal to H (L1 n. . . n Ln-l n L 1; similarly 
n 
0 
HO(H(L~ n.*-n -1 nLnI = H IL1 neben L,-l n L) n is 
a full trio. 
(d) trios are closed under intersection with regular sets and 
e-output bounded a-transduc'tions , while full trios are 
also cl-osed under arbitrary a- transduction. 
We now Trove the. theorem by induction on n. 
Basis. For n=l, by Theorems 4 .I and 3.3 there exist a, 
* 
transducers O and O such that L-RSTRAT(RST) = o~(O~(VE)~I,(G~~)~V~; 
1 0 
then our theorcm holds by notes (a) and (d) above with h being the 
identity map. 
induction step. For the .cam n+l, by Theorem~r4.1 and 3.3 
there exis ts a- transducer Onvl such that L-RSTRAT(RST) is equal to 
* 
On+ 1 
(L-RSTRAT(TOP(RST)) n L(G )) n V,. 
n+ 1 
(4) 
1 
Meaning type 3, type 2, type 1, type 0, linear language. 
But by induction, there exist homomorphism hM and languages 
L:, .,. , Ltt such that L-RSTRAT(TOP(RST)) = hTt(L!i n.. . n- Lt\3 . 
n n 
Substituting this in (4) and appllying notes (a), (b) and (c) we 
find a homomarphism h and languages LI,. .-, L n+l of the same type 
as L';, ..., Ltt and L(Gn,+l) such tkat L RSTRAT(RST) = h(Ll fi . . . nL 
) . # 
n n+ 1 
Remark that by Theorems 34.4 and 3.5 the same result holds 
in the case when the realizations do not contain null rules, and by 
examining the above proof it can be seen that the homomoxphism h 
can be restricted to being e free in this case. 
The following canverse to Theorem 4.2 can be easily 
established: 
Theorem 4.3 Given homomorphism h from T to TI., and rewrite 
grammars G . . . ,G with terminal alphabets T, then for i = 0, .. .,n 
1' n 
there exist context-free acyclic rewrite systems Ri such that 
Proof For j cn, define R. to be {a+al aaT] , by the definition of 
3 
RSTRAT-derivations this will simulate the intersection of the 
languages generated by the tactics. Finally, define R to be 
n 
{a+h(a) 1 ad), thus perf~rm~ng the horrtomorphism on the intersection. 6 
To begin with, the above theorems partially confirm Sampson's 
hitherto unproven claim (C 13 : page 111) 
that stratificational 
languages are the ~esult of intersecting the languages of the 
tactics. 
Note however two important qualifications to this claim: 
the realization derivation must be k-leftmost and a homomorphism 
must be applied to the intersection of the languages. 
Theorems 4.2 and 4.3 show that with k-leftmost realization 
derivations, the type i languages (i = 1,2) can be obtained by 
using a type i grammar on one of the tactics, and making the 
other ones non-selfembedding. If all the tactics generate 
context-free languages (as in the case when tactic-derivations 
are leftmost) then n-RSTRAT grammar can generate the hornornorphic 
intersections of the CFLs. For 1122, this is known to equal the 
RE sets if null realizations are allowed; if null realization rules 
are not allowed then for ns3 the n-RSTRAT grammars generate the 
family QUASI of sets recognized by nondeterministic Turing 
machines in real or linear time (C 21) . These observations 
demons t~ate that n-RSTRAT grammars can be appropriately modif ied . 
so that they generate various language families intermediate 
between the regular and RE sets. Unfortunately, even when the 
realization derivation is restricted to being k-leftmost' 1-RSTRAT 
grammars with context-sensitive tactics and 2-RSTRAT grammars 
with context-free tactics can generate the RE sets, unless 
null realizations aro restricted. The baslc problem with 
restricting null rules lies in the pronounced bias of this mob1 
towards the realization of terminal units from one tactics to the 
next. In practice, in order to describe linguistic phenomena it is 
necessary to hwc information about the enti-rc derivation process 
on some tactics. Sampson accomplishes this by introducing "pseudo- 
terminalsvv into strings; for example, if the application of 
production x-+y is to be noted for later use, then either rule x+py 
or xtyp would be used in the tactics to introduce p as a marker 
of the occurrence of x+y. 
The chief drawback of this approach is 
that the npseudo-terminals" such w p must eventually be delete-d, 
making nhll rules necessary. One possible solution may be to dis- 
cover some bound on the number of null rule applitations needed, 
resembling the "cpzling functiontf proposed by Peters and Ritchie 
(C91). Another solution is to consider a new form'al model which 
allows rgallzation to access u~iformly all parts of the derivations 
on tactics; this approach is considered in C31. 
Before concluding, we take a brief look at thc prablams 
raised by one addition to the basic model discussed so far, 
namely ordered yules. It has often been found useful in 
linguistic descriptions to use rules af the form "A+u if some 
c,ondition C hlds, otherwise A+vv; basically, these types of 
r~leS' av~id stating the negation of ~ondi~tion C, which may be 
cumbersome. In certain stratifi~at~onal descriptions. this has 
lapsed into the use of rules of the form "A+u if this can lead 
to a completed derivation, otherwise A+vu. This notion is 
formalized by Sampson through the assignment of 'lwcightsr' or 
"preference values" to certain rules. Thus A+u may be given 
value 1 while A+v receives value 0, and these values fire 
accbmulilt%d throughout the derivation. At the end, only those 
expression strings resulting from some content strlng are taken 
which have derivations with maximal preference values. The 
fundamental problem with this use of flordered rulesv is that 
even in context-sens-itive grammars it is in general recursively 
undecidable whether a certain derivation can be successfully 
completed or not. In faat, we show that using "ordered rules" 
we can genelate even non-recursively enumerable sets, an obviously 
undsirable situation. 
Theorem 4.4 There exists a context-sensitive grammar G with one 
"ordered ruleu which generates a non-RE language. 
Proof The proof rests on the we11 known result that there exists 
an RE language LO over some alphabet T, bhose complement - is - not -3 RE 
and that there is a type 1 grammar GO = (NO ,T~,S',P~), where 
TO = T u {b,#), such that L(G') = {uf#b 
i (w) 
 we^', i(w) is some 
integel; depending on w 1 (1 . Consider the grammar 
N 
G' = (N1,T?,Sf,Pr) where Nf *= NO u TO u {Y,Sv,Z), To' = T u T, 
Pt contains PO and additional productions as described below, 
The grammar G' behaves informally as follows: 
* 
(a) from St, we gene~ate some string wSf such that weT , 
using productions from Hf + as9 I BET-); 
(b) 
then wc apply thk ordered rule " St -+ YSO wlth weight 1, 
S' -+ ?Z with weight "0"; the plan is that the new nontermlnal Y can 
be rewritten Into a terminal, 7, if and only if Y appears in a 
L 
string $clanging to {wYwlf} {b} (i .e. iff rules of GO can be 
3 
uscd to ~ewrite SO into somc w#b , where w is thc same as the guess 
0 
made in fa)). Once some de~ivation from S is completed, it is 
clear that context-sensitive rules can bebused to check out the 
aoove condition for Y. In addition, the same rules can place 
tvbarsv over all the symbols thus checked, resulting, if successful, 
---- 
j 
in a sentence of the form w#w#b . 
(c) 
Z on the other hand simply travels across the string w 
and places ffdotsvl on top of every symbol, using rules from 
{sZ + ZS~S~T} 
---c 
The result will be that L(Gv) = (w#wtbj lwcI,O) u {~Iw~L') . 
Suppose that L(Gf) is RE, and let h be the homomorphism which 
deletes all symbols not in T, and removes the dots from the 
others. Then h(C(G1)) is also RE because the RE sets are closed 
under homomorphism; but h(L(Gt)) is the complement of LO, and 
thus not in RE by our choice of LO. Therefore by contradiction, 
L(G1) is not RE. 
A similar pr0o.f can be given fo'r stratificational grammars 
with two or more context-free tactics. These results draw 
attention to the need to redefine the notion of "ordered yule" 
in stratificational usage, and point out that care must be 
taken 
whenever f~rmalizing aspects of linguistic practice. 
In -conclusiofl, om investigation of the formal properties 
of the stratificational model proposed by Sampson revealed certain 
unintuitive properties which make it less desirable as a tool for 
natural languitge description. Thus, the use of realiz$tion rulcs 
with null lcfthafftd sides was shown to allow unbounded number of 
realizations for certain strings. More significantly, we showed 
that n-RSTRAT grammars with even one tactics allowing self- 
embedding could generate all RE sets. Since there are well known 
problems raigea by this possibility, most significant being the 
inability to decide grammaticality, we identified a lipguistically 
acceptable restriction on the realization, namely k-leftmost deriva- 
tions*, which led to improvements in some situations. Under this 
addi t3onal constraint , classes of n-RSTRAT grammars were shown. to 
variously generate the context -f ree languages, the QuasA-realtime 
languages and the context-sensitive languages. Unfortunately, even 
in this case n-RSTRAT grammars could generate non-recursive sets, 
unless null realizations were restricted, and we discussed the 
problems inhepcrrt in this approach. Finally, we examined the 
definitio~ of "ordered rulesff used in some stratificatichal 
gxammars, and formalized by Sampson, showing that it allowed the 
generatton of even non-RE sets with type, 1 tactics. 
The above formal results about the generative power of 
stratifgcational grammars hopefully answer the requests of critics 
such as Pittha ([101), and demonstrate the inaccuracy of 
Postal's classification of stratif icational grammars a$ simply 
vaniants 01 context-free phrase structure grammars- The 'results 
also indicate same of the problcm areas in this formal model for 
stratification~l linguistics. We emphasize though that the problems 
arc specific to this particular formalism, ahd should not bc taken 
a% condemnations of stratificatibnal linguistics iq general, since 
there are other stratificational models which avoid these pitfalls- 
Acknowledgements 
I would like tp thank Professor Ray Perrault for his 
much appreciated advice and help both during my graduate student 
career and aftt6r it. 
I am also grateful tn Teresa Miao for typing 
this paper and to Peter Schneider for proof reading it. 

REFERENCES 
[I] 
Barnas, R. and Gumb, R., LThe Completeness of Presupposition- 
Free Tense Logic, 
Zeitachrif t fiir mathematiache Logik und 
Grundla~en aer hathematlk, f orthcomlng. (Abstract in 
Journal or Symbolic Loalc, 42 (1 9779, 146-7. ) 

[2] 
Gum&, M. and Gumb, R;, mLopical Techniques for Pinpointinp 
Inconsistencies in the Knowledge Proceqdi-n~s of the 
41st Annual Conf erenca of the American Society for Information 
Science, 1978, forthcoming. 

133 Gumb, R. Evolvinq Theories. Flushinp, N .Y . : pavan, forthcomine;. 
(Abetracts in IJCAIS, v. 1, 567, end JSL, forthcoming.) 

4 Gumb, R. , Extended Joint Consistency Theorem for a Family 
of Pree Modal Lo~ics with Equality, forthcoming. 
ES] 

Gumb, R. # *An Extended Joint Consistency Theorem for Free 
Lopic with Equality, * Notre Dame, Journal of Formal Logic, 
fdrthcominp. (Abstia'ct in JSL, 42(1977), 146.) 

[6] 
Leblam, H. and Guml), B., Y3ouudness anti Cornnleteness- Proofs 
for Three Brands of Intuitionistlc Log3~,~~fOrthcorning in 
a Haven antholoay. 

[?I 
Weaver, 0. and Gumb, Re, VF rst;-Order Prooerties of Relat<ions 
Havin~ the Monotonic Closure PronerBy , " forthcoming. 
(Abstracts in P-roceedlnkof the 1978 ACM Computer 8c.ienca 
Conference, 19, and JSL, forthcoming.) 

Davis, R. (1977): Generalized Procedure Calling and Content-Directed 
I nvocat ion. Proceedings of the Symposium on Artificial Intel1~- 
yence dnd Programing Languages. SICART News1 etter 64, 45-54 

Davis, R Buchanan, B., Shortliffe, E. (197.1): Production Rules as a 
Representat ion for a Knwledge-Based Consul tat ion Prqram. 
Artrficial Intelligence 8, 15-45 

De Kleer, J., Doyle, J., Stcele* G.L., Sussman, G.J. 
(1977): AMORD - Expl icy 
i t Control of Reasoning. Proceedings of the Sympos~un on Artlfl- 
caal fntelllgance and Programming Languages. SIGART Newsletter 
64, f16-125 

Grice, H.P. (1975): Logic and Conversation. Cole, P., Morgan, J.L. (dsJ: 
Syntax and Semantics. Vol. 3: Speech Acts. N. Y. : Acadern~c,41-58 

Grosr, 8. (~976) : Resolv~ng Oef mi te Noun Phrases. Malker, D.E. (ed.) : Speech 
Understand~ng Research. Stanford Research Insti tu te, Chapter 9 

v. Hahn, W (1978a) : Uber legungen xum %omnun4 kativen Status und der Testbar- 
keit von natOrl~chsprachlichen Artificial-Intelligence-Systemen 
(Some Thoughts on the Communicative Status and the Testability 
of Natural Language AI-systems). HAM-RPM Report No 4, April 
1978. (a 1 so to appear in: Sprache und ~a tenverarbei tung) 

v. Hahn, W (1978b) : Probleme der Simul at  ons st he or ie und Fragepragma t I k bei 
der Simulation natiirl ichsprachl ichcr Dialoge (some Problems of 
S~mulation-Theory and the Pragmatics of Questions in Connection 
w~th the Simulation of Natural Language Dialogues). HAM-RPM 
Report No. 6, May 1978 (also to appear in: Ueckert, H., Rhentus, 
0. (eds.) : Komplexe mensch1 iche I nformat ionsverarbei tung. Bet tra- 
ge zur Tagung 'Kogni tive Psychologie' in Hamburg. Bern: ~uber) 

v. Hahn, W., Hoeppner, W., Jameson, A., Wahlster, W, (1978). HAM-RPM* 
Natural Dialogues with an Artificial Partner. Proceedings 
of the AfSB/Gf Conference an Artlficiaf Intel lrgence, Hamburg, 
122-1 31 

Herrmann, T. & Laucht, H. (1976): On Multiple Verbal Codability of Objects. 
Psychological Research, 38, 355-368 

LeFaivre, R.A. (1977): FUZZY Reference Manual. Rutgers University, Computer 
Science Oept., March 1977 

Ritchle, G.O. (1977): Computer Modelling of English Grammar. University of 
Edinburgh, Computer Science Dept., Report CST-1-77 

Scott, C.A., Clancey, A,, Davis, R., Shortliffe, E.H. (1977): Explanation 
Capabilities of Production-Based Consultation Systems. American 
Journal of Compu ta tlonal Lrnguis t ~cs, Mzcrof iche 62 

Stallman, R.M. 6 Sussman, C.J. (1977): Forward Reasoning and Dependency- 
Directed Backtracking in a Systm for Computer-Aided Circuit 
Analysis. Artif~cial InteJligence 9, 135-196 

Wahlster, W. (1978): bie Simulation vager lnferenzen auf unscharfem Wissen: 
Eine Anwendung der mehrwertigeo Programmiersprache FUZZY  h he 
Simulation of Approximate Reasoning Based on Fuzzy Knowledge: 
An Appl i cat ion of the Many Val ued Programmi ng Language FUZZY) . 
HAM-RPM Report No. 5, May 1978 (also to appear in Ueckert. ti., 
Rhenius , 0. (eds .) : Kompl cxe mensch1 Iche I nformat ~onsverarbel- 
tung. Beitrage zur Tagung 'Kognitive Psychologie' in Hamburg. 
Bern: ~uber) 

Winograd, T. (1972) : Understanding .Natural Language. N.Y. : Academic

[l] Baker, B . and R. Book (1974) . "Reversal Bounded, Mu1 ti-pushdoyn 
Machinestf, J. Computer Systems 
Sc jence - - 8, 1974, 315-332 .. 

[ 21 Book, R.V. and S .A. Greibach (1970) . "Quasi-realtime Languages", 
~ath; Systems Theory - - 4, 1970, '97-111. 

[3] Borgida, A.T. (1977) : "Formal Studies of Stratificational 
Grammars", Ph.D. Dissertation, University of Toronto, also 
Technical Report No.112. 

C41 Ginsburg, S. (1975) . Algebraic and Automata-Theoretic 
Properties of Formal Languages, North-Holland Publishing Co. 

C51 Gle-ason, H.A. Jr. (1964). "The organization of language: a 
s tratificational viewIf , ~ono'graph Series on Language and* 
Linguistics - 17, p .75-95, Georgetown Universi ty. 

C61 Lamb, S. (1966) . Outline of Stratif icatiqnal Grammar, 
Georgetown University Press, Washington. ' 

C 71 L'ockwood. D.G. 11972). Introduction to ~tratifi-tathnal 
~in~uistics, ~arcourt Brace Jovanovich,*Inc. 

C81 Peters, P .S. and &W. ~iwhie (1971). "On ~estricting the 
base component of Trans~ormatidnal  grammar^'^. Information and 
Control - 18. 

C91 Peters, P.S. and R.W. Ritchiea('1973). "On the generative 
power of Transformational ~rarnmar~" 1.nformati06 Sciepces 
49- 83. 

C-101 'P'itfha, P. (1974). "On a new form of Lamb.'$ stratificational 
grammarft, Slovo aa S.lovesnost - 35,' p ,208-'218; translated. from 
the origihc Cxchl by D .G. LockiEod. 

[lk] Postal, P. (1964). Constituent Structure: A Study of 
Contemporary Models of Syntzctic Dewscription, lndiana 
University, Bloomington, Ind. First appeared in Int. J. Amer. 
Linguistics - 30.1, part 3. 

E 121 Salomaa , A. (1973) . Formal Languages, Academic Press,New 
York. 

[13 J Sampson, G. (1970) . Stratificational grammar: 9 Defani-tion 
and an Example, Janua Linguarum, Series Minor: 88, The Hague 
Mouton. 
