M)A 14 491 A LNGUISTIC OVERVIEW OF THE NIGEL TEXT GENERAI ON /GRAMMAR(U) UNIVERSITY Of SOUTHERN CALIFORNIA MARINA GEL

REY INFORMAT ION SCIENCES INST W C MANN OCT 83

N~(, A I F0 ISI/RS83-9F49620-79-C-0R F/657 N

ME'ND

111.0 hE.

1265 MIA33.6

MICROCOPY RESOLUTION TEST CHARTNATIONAL. BURAU OF STMANDRS -I3 Atl l

0

IS! Repfn: softsISIRS-83-9

William C. Mann

A inuiic Oerie

of the NigelText Generation Grammar

;L*

__ I ELECT.f rx public releaso and 9Cd ifidd~trb71 toi unlzmit4 A

INFORMATIONSCIENCEST IIwan

Il 1 07 oa[

UnckalfedSECURTY CLASSIFICATION Of THIS PAGE9 (Whes Do* ee.

REPOR DOCUMENTATIN PAGE S 071CTMI. IMPORT HUMmE CATV ACS~ME N ALO IMUIN

ISI/.83 4nd3' Sutte 5. TYPE OF REPORNT A PERIOD COVERED

A Linguistic Overview of the Nigel Text Generation Grammar Reeac ReorS. PERFORMING ORIM RE1O1T11U11ET

7. AuTmOR(e) 11. CONTRACT OR.GRANT 111111091()

William C. Mann F92-9C08

9. PERFORMING ORGANIZATION NAME AND ADDRESS 19S.

USC/Informtion Sciences Institute TAMK

4676 Admiralty WayMarina del Rey, CA 9029 _____________

TO. CONTROI.LING OFFICE NAME AND ADDRESS It. REPORT DATEAir Force Office of Scientific Research October 1003Building 410, Boiling Air Force Base Is. MUMERp OF PAEWashington, DC 203M 16

14. MOtIITORING AGENCY NAME & ADDRESS(i lf Wuitee kete CemkIifte 0M) III, SECURITY CLASS. (ofm IV m)

is. fAIC&U1ATO/bNRhG

16. DISTRIBUTION STATEMEN4T (of thle RopAMJ

This document Is approved for public release; distribution Is unlimited.

17. ISTRIeUTION STATEMENT (of the abee enewed ho Week S. it 4111laeuM M Re")

15. SUPPLEMENTARY NOTES

This repotIs a ! t of a paper that appears In the proceedng of The Tenth LACUS Forum, held iQuebec In August, 1963. The Tenth LACUS ForumIs publshed by Mornbean Pres.

19. K(EY WORMS (Cewif.. " .t~we, sif. Of .meemt md 0=14 IV' WOOe NMim

artificial Intelligence, choice experts, computer generation of IMx, Inquiry semantics, bwee2; representation methods, natursl 'inguage, Nigel, Penman, semantics of English, systemi gramar,

text gene14ration

KABSTRACT -OWn "~war" aftNaeg~p I~I

(OWM)

ODp in I'ON 3 Vfr ow it lo as ae. eum~irrsWINu 41IM-41oo J

01ON1&Q 00I I*rom

f

USMCiaSt CLMICATM OF THIS PAfin. DW D

20. ABSTRACT (continued)

Rcent text generation research resemables recent research In synthesis of vaccines. The research Isdesigned to construct entitles which previously arose naturally. This constructivo approach createspractical and theoretical benefits.

Our text generation research has produced a large systemic English grammar, which is embedded ina computer program. This grammar, which Is called Nie, generates sentgnces It is controbd by asemantic stratum which has been added to the basc ystemic framework.

This paper describes the program, which also Is called Ngel. It identifies M enMtatlons of varixuprecedents In the osysemic framwork, and it Indicates the current status of the program. The paperhas a dual focus. FlIt, on Nigel's processes, It describes the methods Nigel uses to control text tolulflll a purpose by using Its new semantic stratum. Second, concerning Nigel's interactions with Itsenvironment, It shows reasons why Nigel Is easily embedded In a larger experimental program.Although the paper does not focus on Nigel's syntactic scope, that Its scope is non.trMl is indicatedby thie fact that all of the sentence and clause structures of this abstract m within that actl cscope.

0 • ,

ISIber no~i

WilliamIC.SMann

laws-i 11JUL6511

WilliamO C.Oan

ISBN"%~~~~o the* Nfigel 1 1 1 fSSIIV*Tex Geeato Grammar t t 4I

ISI,staff and publbhs In proilal omiancnf. proceedngs For

Docun WbudontwClwformallan Science Inute46Th Admfr* WtyMarin del Roy. CA O29IUSA

-7.

Contents1. -progress In Immunology: Synthetic Vaclinas I2. Penman: Constructive Rlesearch In ingpuistlos I3. NWI: Penman's Gramm ar 2

&I Leien 2W. Rseallation 33. Choice: Systu. and Gen 4U 4Choosinip 4

3.5 biul" 53.6 Environment 5

4. The Inquiry Stratum as a Semantics 05. State of Development SReferences

low

t LZ

1. Progress In Immunology: Synthetic VaccinesIn February of this year, SclenMf American dJ r bd if breakthrough In medical science, the

Mibrtory synthesis of vaccines against flu mid other vius diese Lamr 01. Ever since vaccineswera ocs vd rsac on vaccines mnd Immunology hes focused on natral substances and theireffects. Now It hNo become wpossible t synthesize vaccines to Wih many common diseases.

In synthesizing vaccines, scientists had to supplement the estableied methodologies. Inparticular, they had to develop methods of vaccine const ruct ion to supplement existing methods ofvaccine kientfcation and description.

When the work on synthesis began, ther was a substnti amount of evidence to suggesttho thetsk would be overwhelmingly complex, and, In a practical genes Impoasble. The ScIentifcAmerican article dri bes how they did It. They worked with mondels at the molecular level, a finerlevel of detail than that of most of the preceding- work. One of the key stage in the process camswhen they discovered how to caus syntheized vaccine elements to wrpm the many abstractatiutes by which va ccie had previously bee described (without refeenc to fth molecularlevel.)

(We found that ...synfthel peptide can mkric the distinctions revealed by serologicstudies; in designing synthet vaccines, one wil be able to take advantage of serologicevidence."

The key evidence of success, of course, Is that the vaccines work. Out of voluminous anddetaile reasoning about molecules their shapes, and their Interactions, come chemicais thatactually preen heInections that they wre supposed to preventL The succms of the vaccinesvalidates the work at the molecular leMel and also validates the particular t1heories at higher level onwhich 4wa based. When a synthesized vaccine is Ineffective, It point to Iiadquacies In the

So, work aimed at creat synthetic vaccines has also created a powerful new too foraclenific inqiry. Its worth can be fully justfied on t scientific beneft aone, or on hts practicallbe-nRimsiaone.

2. Penman: Con structive Research In LinguisticsThe work derbed in ti paper is another variety of aclendt work In a constructive

methodology. in this cas, the Items bein constructed are tt rather than vaccines anid the test ofelfectivenees involve reading the ts athe than resisting a dIse-ae Despite the differences, thework, on syntei vaccine helps us to understand curt work on MWx g"nerstn.

in tat g-nesaomen the local tish Is to creat a Wkxt In fluet natu rna*e (often Engish)i respoa ta some particular need. The M&c Is to do so using onl the exlctknowledges ofbI guag speifedby oes ais*tothiesW ratherthanusinglthe stle ofa person. Pesarc of96 adr has been gead In in scattered helulon foramr a decade, general using a. onompua asthe Ot of mynthIl mob&da the repositoy of the most detailed leve of theory [Davey 71k Mann

Tkbe rupsi b L~kiftdPW 90 lsumi MftI 'Mua I Tfw Tees LAOW AmMui MeW In baebe* Ift SUAI%ML rs. T1WOI ft nb pulmll y"uMkMOMI

1.1 zz

2 A LINGUISTIC OVERVIEW OF THE NIGL TEXT CONRATION GRAMMAR

81, Mann & Moore 0% MlCeOwn 6,Mom & MatthilseON 8Ua). ComputerS, here INI 0he role Ofmolecular genetics for th vaccine work: they are a relatively now technology, enabing the work in apractical sense but not In any way funidamental to IL

Text generation reeearch uses a constructive methodology, but It Is vitally dependent on prior

requires extensivereoclainoMecnrbtnthoisJutafrthVcietebdvalidation of the constructive theory Is the observable effectiveness of the synthesized results.

Several years ago, we began work on a new text generation system, named Penman, designedto write texts a few paragraphs long. A previous round of research had led to a computer progrmwhich was able to write a limited range of two-paragraph texrts, but which had several seriousshortcomings, especially In the narrow ri y of its grammar.

We therefore wante to Include In Penman a significant, linguistlcally Justified grmmr. Now,several years later, we have such a grammar, named Nigel after Haltiday's learner [Hailday 753. Thispaper pa1s1ses over the parts of Penman devoted to Invention, to tex planning and to retrospectiveimprovement of the tex, and concentrates entirely on Nige, the grarmar.

3. Nigel: Penman's GrammarThe grammatical framework of Penman Is " tile ytmirmework, begun by Michael Hiliday

In the late 1950s. 1 it draws on a wide range of prior work, Including [Halliday M6 Hudson 78, Halliday& Martin 81, Fawcett 80, Berry 75, Berry 77, Hallday & Homan 761 and others. In order to reconcile thevarious fragments of grammar and to augment them, all of the prior work has been altered In someway, sometimes by a simple notational shift, and sometimes by thorough re-representation.

I will describe Nigel In a senis of stages, In effect working backward through the generationprocess from the level of lexical Items to the level of conditions In Nigel's environment which affectthe particular text produced. The whole discussion will be about Niger role In generating singlsentences, because Penman Is desined to plan text down to the sentence levvel, including therelations that each sentence will express, and the to have Nigel execute each sentence planIndependently.

3.1 Lexicon

NigW' lexicon Is deliberately oversImplf~d because we felt that the 11ultin technicalproblems were elsewhere. It Is a lexi of Independent lexical ftms witimu morphology. Thelexicon in wall elaborated for leiAl featur theo are over 100 SdisIc femide features, enid toelexicon haes Items representing over 800 distinct coraitioa f tsmdcal features. However, thmefigure ae not pertcufaly signrfcant, since the lexkwcon not beew efenevef developed orteed

17m wo an Noi woWM ethew Wea poWUI W o t 'alve pmlScp dsn Owlolm hft5mnM ml ftx ip- OWa mmd WWA pemdOsrbise a1he oiled mmd aewINKV0

fN

UNNU POIMOS GRAMMAR 3

3.2 Realization

Nigel bulds syntactic structures by a set of activities usually called oalization in thesystemic framework They are distinct from act*vt~e ~hc specify the charateristics of a syntacticunit, formally termed grammatical features. Each syntactic unit is fira developed sea set ofgrammatical features, which realization converts to a eswitsctio strufture All of the control over whati built Is exercised during the creation of the feature aet there Is no opftiol or syntactic varkbiltIn realiation.

In Nige, al~l rliatonisr a aosnlM sh Each grammatical feature may have oaeor more realize g on statent associated wit toeah conulngof a rebatoupeor ida number of operands. Each realization statement makes some change or introduces somerestriction on the structure being produced.

There are three groups of realization operators those that build structure (in tern. ofgrammatical functions), those that constrain order, and those that aseoclats features withgrammatical functions.

1.- The realization operators which build structure wre Insert, Conflate, and Expand. Byrepeated use of the atructure-building functions, the grammar Is a"l to construct satofunction bundles, also called fundles. None of these operators are new to thesystemic frameworkf.

2. Realization operators which constrain order are Partition, Order, OrdorAtivront, andOrderAtEnd~ Partition constrains one function (hence one funft) to be raied to theleft of another,-but does not constrain them bo be adjcent Order constrains Just gePartition does, and In addition constrains Vse two to be realized ajacntl. OrderAFrontconstrains a function to be realized as the letmost among Vs daughters of it mothe,and OrderAtEnd symmetrically as the rightmost. Of thes, only Partition Is new to VsaSYSOemic bhmewo&

3. Realization operators that associate features with functions ase Presee~, whichassociates a grammatical feature with a funtion (and hence with Nts fundls Classfy,which asciates a lexical feature with a function; Out~lesaffy, whc ascistes alexical feature with a function In a preventive war, and Lexify. which forme a partlalexical item to be used to realize a function. Of these, OutClassify, and LA*it we new,taking up roles previously filled by Classify. OutClassify restricts Vse realization of afunction (and hence fundis) to be a lexical item which dM- no bear Vs named feature.This is useful for controlling RIin exception categories (eg., reflexives) in a localze.manageable way. Lexify allows the grammar to force selection of a particular itemwithout having a special lexical feature for that purpose. it is Preselec which maoe togrmmar recursiv, since Preselec requires choosing a particular grammatical feture ina low-rranked pass through Vse grammar.

In adition to these realization operators, there is a set of Default Funt Ion Order Llets.Them we Nets of function-. which will be ordered in particular ways by Nigel, provide thVfunctions on Vs Eaits occur In Vse structure and that Vse realization opertor have not alreadyordered t1hose futions. A large prportion of Vse constraint at order Is performe thro S us of

Pt toie 1deso~ n of Ive wes of orer0 la toe ieti No mov A Iwon mubstiarame tr knimp @Wo PCOgMMiN go oi sdoue apeeWe of NWga has oo Aid w S

4 A LNGUISTIC OVEIVEW OF ThE NIGEL TEXT GENERATION GRAMMAR

systemic ordering is In fact a fairly complex matter. The (as yet unpublished) ordering algorithms ofNigel constitute a definite and testable proposal for the meanings of the realization operatoWr fororeing.

3.3 Choice: Systems and Gates

Nigel has systems of alternatives, called systems as In the systemic tradition (to theconfusion of the computational tradition.) The alternatives are grammatical features. Each systemalso has an entry condition, a logical expression of grammatical features. The entry condition mustbe satisfied in order to enter the system, I.e., to have the set of alternatives avalable for choice. Whena system has been entered, one of the alternative features must be chosen.

In addition to the systems there are Gates. A gate can be thought of as an entry conditionwhich activals a particular grammatical feature, without choice. These grammatical features areused Just as thos chosen In systems. Gates are most often used to provide a feature to be realized,In response to a collection of features.2

3.4 Choosing

The systemic literature has many discussions of the oppositions of language, the directalternations represented In systems. There is much less discussion of which alternative is mostsuitable In particular cases. Of course for text generation, making good Individual choices is anessential activity, and so there must be some representation of how choices In the grammar are to bemade.3

In order to specify explicitiy how choices are made, a new definitional stratum has been addedto systemic notation. For each system, a formally defined proem called a chooser or choiceexpert Is crested. Each such process consists of stem potentially of several kinds. The principalkinds of steps we Information gathering, discrimination between kinds of conditions, and choice.When a system is entered, the corresponding chooser process Is executed, yellding a choice amongthe system's alernatives.

By defining choosers in this way, we can make explicit what particular choices depend upon,and we can examine whether particular natural examples conform to t conditions of choice whichhave ben defined.

The sct of defining choose oft reveal reguarides (or irregularitie) which thegrammar does not represent S ral choosers my depnd in ft same way on the amedeterminative condiltion. Or a notion such as mark dnms may turn out to represent very different

2Thgr s.j anv fe nP rduns wte imen m n es fras^ md no n nm eoa pn~ dh a iMnM

howilosrn - mb TegM f M-Al dr oft nimsr o ote. wMl a no e eft igt k am*r

aft ld It -W f eWqlm ie an ft term grmo rit m m m ta S m imueon - of wn"FWMfte INn f - WWl semle b kt e Is a rolf IU mdnn dilb bow. Or MI udAeq

.RL

-----------

N10=- PUOWAW$eaAM 5

cxondtIn hIts various grammatical systems. Defn choosers typically lead to local efinmen Ofthe grammar, along with sughed iflonfOrw the p rtl~bom used.

3.5 Inquiry

ft would be possible to allow chousers to have some sort of unrestricted access to theknowleodge which surrounds them, but t would be unsatisfactory a thO" said unmanageable aapractical text geneation. resource WW4s~ Nige has a very W mple highly Ubastrce method forfthossr* to gain acem to the information the need. Choosers gain inkformaton to guid their work

only by Ieasuing Inquiries stated in a simple Inquiry lanp"*g.

The boundary of t grammar sparates; two nearly Independent symbol systems. Outsid ofthe boundary. In the eawlronmont, there ls knoldge of what needs to be said. includin bo&lgeneral knowledge sand the text plan. Inside the boundary are grammatca features, grammatic.function symbols, chooser definions, steom and ge defiitions, end realiatio statements. AlN cthes are beodthe reach of the environment and cannot be desateod for manipulation by ItConversely, the symbo system outside of the environment le not directly availa9ble to t grnmwWhen the grammar needs a symbol to use In som loter inqury, such ma designation of the soentca process so that it can Inquire whether the agent ls multiple, it sa for a tempxory symbol for tubpurpose. These symbols are disarded once the unit has been built, and the separation of symbolsystems bs thereby maintained.

(Lexica Items are exceptions to these remark about symbol syste separation. Thechoosers assume that assoclations are maintained btenthe releivant knoldge said lexcal tems,so that, for example, the grammar can elci a set of dinolationally appropriate San.il as cndidae toserve as the head of a nominal group.)

This way of defining and using an inquiry langug has kmportan practical and teoreticalborbels. It pe-rmile development of the grammar and its semnantics while avoiding two tras:

1. dollnhng the grammors semantics In terms al perIwlmr conventions of knowledgerap M eeaon;

I. defining the grammar's semantics relative to particular syntactic structures, rather than tothe functions! ino

Avs*din to fIrM of thee traps Is particularly Inportant N Nigel Is to be used In Ohe arifclillgence 'el prolct Since Nigel can be I idependsnt of particular knowledge

rapsoesllan fomlmto rpid. progressm being 'mae In knowledge repesentation will not-ef -lpfdonfn obsof

3.6 Environment

I"e Imhnet is Me W a the -'em Mobu of sysibel 0~Ms 11e1ond the gramarsbow-ky ft of tee rrnsr~h dukmrt o tt*1'kf nufLu nfraly th0y aro

1. 1he Xat Mom ,akmoms whish afed pWie to the demand far which ten Is

All~uie Ae

-. J

A LINGUISTIC OVERVIEW OF THE NIGEL TEXT GENERATION GRAMMAR

2. the Text Plan: information which Is created In response to that demand.

Nigel leans heavily on both. It presumes that the text plan contains definite intentions aboutthe ideational, interpersonal, and textual functions of the unit being generated; much of the ideationalinformation comes from the knowledge base.

4. The Inquiry Stratum as a Semantics

Although definitions differ widely, the term "semantics" is usually used to represent some sortof specification of correspondence between elements of a linguistic system and elements of anothersystem distinct from it. Taken this way, there are two senses in which the Inquiries of Nigel constitutea semantics.

First, given a grammar of a language, including choosers, the collection of inquiry operatorsused In the choosers constitutes a specification of what can be expressed in syntactic structure. Forexample, multiplicity, intention to emphasize, and time precedence are identifiably expressed inNigel's grammar of English. And we can also say, on the basis of the collection of inquiry operators,that English tense is indifferent to the contrast between moments and intervals. In this sense, thecollection of inquiry operators provides a semantics of the collection of syntactic structures.

In the second sense, given the particular choosers, systems, and realization statements of agrammar, we can construct the mapping from particular conditions, i.e., particular collections ofenvironmental responses to inquiries, to strings of symbols which they yield. This is a semantics of

the grammar of particular utterances.

Note that, in both casee, a semantics of the grammar is specified, rather than a semantics ofthe language as a whole. Lexical aspects are specified in only a very rudimentary way, and thesemantics above the level of the largest grammatical unit is likewise only slightly constrained. T.eselimitations can be regarded as advantages, because they provide a principled factoring of a verycomplex field of inquiry.

5. State of Development

The generation mechanisms of Nigel have been programmed and tested. Choosers for abouttwo thirds of its 200-odd systems have been defined. Whenever a new cluster of choosers is defined,there is an Inevitable reexamination of the systems of that region, and of their justification. As aresult, Nigel as a whole is evolving toward an increasingly hom.-eneous grammar in a fairly

consistent definitional style.

When there are choosers for all of Nige's systems, many new tests will become posble.They wIN Involve generating units on demand, attempting to Imitate natural examples, andcharacterizing syntactic units by the demands for which they were produced. Such tests, while vitaland Informative, we local to the grammar. They cannot show how adequate the grammar Is a andement of a text generator.

--

STATE OF DEVELOPMENT 7

We look forward to eventually mating Nigel with a programmed text planner, and laer withprogrammed search processes (for invention) and text improvement processes well. Only then

can Penman be tested as a synthetic vaccine Is tested--by judging its operational results. Seeing ft

products of other text generators, we expect that Penman will eventually generate vry high qualitytext, and that the process of defining the generator will be filled with exciting and Informativeresearch.

4

4AddftonaI donrmelon an pe uulr smpec of the work can be found in miled report &id pusb rcamc Panmenadeftmnn a], Owar dulnl (Mann , Ns Fou ( M & Matlimusmn e3b. Inqury m aPge rmN.exta ndd simiims d N a o : (MmS M uami aS .

i. n

A LINGUISTIC OVFRVIEW OF THE NIGEL TEXT O&ERATION ORAILMAR

References(Berry 75] Berry, M., Introduction to Systemic Linguistics: Structures and Systems, B. T. BEkord,

Ltd., London, 1975.[Berry 77] Berry, M., Introduction to Systemic Linguistics. Levels and Links, B. T. Botsford, Ltd.,

London, 1977.

[Davey 79] Davey, A., Discourse Production, Edinburgh University Press, Edinburgh, 1979.

[Fawcett 80] Fawcett, R. P., Exeter Linguistic Studies. Volume 3: Cognitive Linguistics and SocialInteraction, Julius Groos Vedag Heidelberg and Exeter University, 1980.

(Halliday 75] Halliday, M. A. K., Learning How to Mean, Edward Arnold, 1975.

[Halliday 78] Halliday, M. A. K., System and Function In Language, Oxford University Press, London,1976.

(Halliday & Hasan 76] Halliday, M. A. K., and R. Hasan, Cohesion in English, Longman, London, 1976.English Language Series, Title No. 9.

[Halliday & Martin 81] Halliday, M.A.K., and J. R. Martin (ads.), Readings In Systemic Linguistics,Batsford, London, 1981.

(Hudson 76] Hudson, R. A., Arguments for a Non-Transformational Grammar, University of ChicagoPress, Chicago, 1976.

[Lerner 83] Lerner, Richard A., "Synthetic Vaccines," Scientific American 248, (2), February 1983.

[Mann 81] Mann, W. C., "Two discourse generators," in The Nineteenth Annual Meeting of theAssociation for Computational Linguistics, Sperry Univac, 1981.

[Mann 82] Mann, W. C., The Anatomy of a Systemic Choice, USC/Information Sciences Institute,Marina del Ray, CA, Technical Report RR-82-104, October 1982. To appear in DiscourseProcesses

(Mann 83a] Mann, William C., An Overview of the Penman Text Generation System, USC InformationSciences Institute, Marina del Rey, CA 90291., Technical Report RR-83.114,1983. To appear Inthe 1983 AAAI Proceedings.

(Mann 83b] Mann, William C., "Inquiry Semantics: A Functional Semantics of Natural LanguageGrammar," In Proceedings of the First Annual Conference, Association for ComputationalUnguistics, European Chapter, September 1983.

(Mann & Matthissen 83s] Mann, W. C., and C. M. I. M. Matthlessen, Nigel: A Systemic Grammar forText Generation, USC/Information Sciences Institute, RR.83-105, February 1983. The papers inthis report will also appear in a forthcoming volume of the Advances In Discourse ProcessesSeries, R. Freedle (ed.): Systemic Perspectives on Discourse: Selected Theoretical Papers fromthe 9th International Systemic Workshop to be published by Ablex.

[Mann & Matthlessen 83b] Mann, William C. and Christian M. I. M. Matithiesen, An Overview of theNigel Text Generation Grammar, USC Information Sciences Institute, Marina del Rey, CA 90291.,Technical Report RR-83-113,1983.

i I I

[Mann & Moore 80] Mann, W. C., and J. A. Moore, Computer as Author-Rsults and Prospect,USC/Information Scienc Institute, RR.79.82, 100.

(McKeown 82] McKeown, K.R., Generating Natural Language Text In Response to Questions aboutDatabase Structure, Ph.D. thesis, Universt of Pennyvanlia, 1982.

74-6 .