Intelligent Agents for Information Presentation: Dynamic Description - - PDF document

SLIDE 1

See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/2781252

Intelligent Agents for Information Presentation: Dynamic Description of Knowledge Base Objects

Article · September 1998

Source: CiteSeer

CITATIONS READS

23

6 authors, including: Some of the authors of this publication are also working on these related projects: We Feel View project EBMSummariser View project Karin Maria Verspoor University of Melbourne

283 PUBLICATIONS 4,225 CITATIONS

SEE PROFILE

Robert Dale Language Technology Group

240 PUBLICATIONS 5,836 CITATIONS

SEE PROFILE

Maria Milosavljevic The University of Edinburgh

40 PUBLICATIONS 388 CITATIONS

SEE PROFILE

Cécile Paris The Commonwealth Scientific and Industrial Research Organisation

340 PUBLICATIONS 4,751 CITATIONS

SEE PROFILE

All content following this page was uploaded by Karin Maria Verspoor on 09 November 2014.

The user has requested enhancement of the downloaded file.

SLIDE 2 In telligen t Agen ts for Information Presen tation Dynamic Description

• f

Kno wledge Base Ob jects Cornelia M V ersp

• r
• Rob

ert Dale

• Stephen

J Green

• Maria

Milosa vljevic

• Cecile

P arisy and Sandra Williams

• Microsoft

Researc h Institute Macquarie Univ ersit y Sydney NSW

• Australia

fkversprdalesj gre en ma ri am sw il lia mg m ri mq ed u au T el

• F

ax

• yCSIR

O Mathematical and Information Sciences Lo c k ed Bag

• North

Ryde NSW

• Australia

CecilePariscm is cs iro a u T el

• F

ax

• Keyw
• rds

natural language generation user tailoring m ultilingual information presen tation Abstract Users

• f

the W

• rld

Wide W eb ha v e needs and in terests whic h can help to determine what

• f

the v ast quan tities

• f

information a v ailable migh t b e relev an t to them In telligen t agen ts migh t b e used to select con ten t for a particular user Ho w ev er it is also imp

• rtan

t to consider how that con ten t is pro vided to a user W e suggest that this information pr esentation m ust also tak e in to consideration the needs

• f

a user and discuss a set

• f

agen ts whic h utilizes natural language generation tec hniques to presen t information in an appropriate w a y

• In

this pap er w e describ e t w

• systems

w e ha v e built whic h dynamically generate descrip tions

• f

kno wledge base en tities and consider the extension

• f

the tec hniques used there for m ultilingual information presen tation W e describ e the notion

• f

a phr asal lexic

• n

as a basis for dynamic

• b

ject description and prop

• se

a mo del for dynamic m ultilingual description whic h builds

• n

that notion

SLIDE 3

• In

tro duction In the con text

• f

the W

• rld

Wide W eb in telligen t agen ts are

• ften

discussed as a means

• f

nding information that is relev an t to a particular users needs at a particular p

• in

t in time Those needs ho w ev er also inuence the most eectiv e presen tation

• f

the relev an t information as migh t the needs

• f

the information pro vider It is therefore imp

• rtan

t to consider w a ys

• f

in tro ducing usertailoring in to information presen tation In this pap er w e explore the use

• f

a set

• f

in telligen t agen ts for information presen tation

• n

the W

• rld

Wide W eb These agen ts use natural language pro cessing NLP tec hniques to generate appropriate descriptions W e argue that a k ey b enet

• f

this tec hnology is the p

• ten

tial for data reuse under v arying presen tation constrain ts and examine the particular case

• f

m ultilingual presen tation to suggest

• ne

w a y forw ard for this tec hnology

• This

w

• rk

builds

• n

existing w

• rk

in tailored kno wledge base description

• the

PEBAI I system Milosa vljevic T ullo c h and Dale

• the

ILEX system Knott et al

• and

join t w

• rk

Dale et al in press

• and

describ es ho w kno wledge base en tities can b e dynamically describ ed W e also discuss the ease with whic h w e w ere able to p

• rt

the PEBAI I system to a new domain in the PO WER system due to the use

• f

a phr asal lexic

• n

and a reusable agen t based system arc hitecture and explore ho w these features

• f
• ur

approac h migh t also b e applied in a m ultilingual generation system

• T

ailoring metho dologies Information presen tation should b e tailored to the needs

• f

a user and the needs

• f

the informa tion pro vider F

• r

instance new comers to a domain will w an t to see information at a dieren t lev el

• f

detail than exp erts In the con text

• f

a virtual m useum a m useum curator ma y ha v e a particular message he hop es to con v ey to visitors and the visitors ma y ha v e an agenda in visiting a particular exhibition Dieren t p eople ma y ha v e dieren t p ersp ectiv es

• n

individual

• b

jects John migh t w an t to kno w the history

• f

an

• b

ject while Jane is more in terested in its comp

• sition

Suc h tailoring can b e ac hiev ed in v arious dieren t w a ys W e consider a few curren t metho d

• logies

b elo w

• Adaptiv

e Hyp ertext Brusilo vsky

• reviews

adaptiv e h yp ermedia systems These systems construct user mo dels whic h driv e the adaptation

• f

a h yp ermedia page This adaptation migh t in v

• lv

e for example limiting the bro wsing space a user has access to

• r

prioritizing andor annotating h yp erlinks to guide a user through the a v ailable information These systems w

• rk

with static textual do cumen ts but v ary the access whic h a user has to those do cumen ts A daptive navigation supp

• rt

systems supp

• rt

v ariations in whic h links are a v ailable to a user at a giv en p

• in

t

• r

in the

• rder

in whic h the links are presen ted and adaptive pr esentation systems include

• r

suppress certain bits

• f

text dep ending

• n

the needs

• f

a user Although the use

• f

suc h approac hes is an imp

• rtan

t step to w ards the tailoring

• f

information to a user they do not go as far as they migh t F

• r

example they do not consider issues

• f

textual coherence

• what

is the eect

• f

limiting access to texts

• r

c hanging the

• rder
• r

their presen tation F urthermore they require all the texts p

• ten

tially a v ailable to a user to b e pre written whic h demands a large amoun t

• f

w

• rk

when man y lev els

• f

v ariabilit y are in tro duced and complete rewriting

• f

texts when c hanges

• r

additions are made in the information to b e con v ey ed This is an issue for m ultilingual systems in particular since the same information m ust b e expressed in sev eral dieren t w a ys languages

SLIDE 4

• Natural

Language Generation Natural language generation NLG systems aim to pro duce natural language text from an un deryling represen tation

• f

kno wledge The basic comp

• nen

ts

• f

an y natural language generation system can b e iden tied as follo ws

• Con

ten t selection The selection

• f

information to b e included in a text

• T

ext structuring The

• rganisation

and

• rdering
• f

the selected information

• Surface

realisation The mapping

• f

the information to b e con v ey ed to w ellformed sen tences and paragraphs

• n

the basis

• f

a grammar and lexicon for the target languages These systems in v

• lv

e goaldriv en planning in whic h the form ulation

• f

a text m ust satisfy a comm unicativ e goal Con ten t selection for instance will dier for dieren t goals as dieren t information will b e relev an t for eac h goal Suc h goals migh t b e instan tiated from sc hemas McKeo wn

• r

a mo del

• f

a users in terestsbac kground can pro vide comm unicativ e goals whic h driv e the generation

• f

a text designed sp ecically for that user Eac h

• f

the ab

• v

e comp

• nen

ts is indep enden t and em b

• dies

kno wledge sp ecic to the task the comp

• nen

t m ust ac hiev e They can therefore b e construed as agen ts whic h reason

• n

the basis

• f

a curren t comm unicativ e goal ab

• ut

what information to express

• r

ho w to express that information Eac h agen t is endo w ed with a detailed mo del

• f

its task and with reasoning pro cedures for ac hieving that task relativ e to the sp ecic con text F

• r

example the text struc turing agen t migh t utilize a nite state automata for determining what

• rder

to express selected information c hosen b y the con ten t selection agen t in The agen ts act in sequence to pro duce the desired

• utput

texts The system arc hitecture

• f

the PEBAI I and PO WER systems to b e describ ed in Section

• is

found in Figure

• and

sho ws the sequence in whic h agen ts act and the resources whic h they utilize It do es not include an explicit con ten t selection agen t b ecause in this system a command giv en b y a user serv es to delimit the p

• rtion
• f

the kno wledge base whic h will b e expressed as a result

• f

reasoning b y the text planning and surface realisation agen ts This means that information

• nly

needs to b e enco ded

• nce

and texts can b e pro duced dynamically from that information in v arious

• utput

forms A generated text migh t for example v ary at the surface realisation lev el in v

• cabulary
• r

structure for dieren t age groups

• r

in languages for dieren t linguistic groups

• r

at the con ten t selection lev el through presen tation

• f

dieren t bits

• f

information

• Dynamic

Hyp ertext Dale et al in press discusses the application

• f

NLG tec hniques to the h yp ermedia con text The result is a system whic h dynamic al ly creates a h yp ertext net w

• rk

and h yp ertext no des do cumen ts in resp

• nse

to information ab

• ut

a user

• r

particular comm unicativ e goals The use

• f

language tec hnology in the h yp ertext con text allo ws more v ariation in what information is presen ted to a user and in the w a y that information is presen ted The system decides whether a particular string

• f

text should b e mark ed with a h yp ertext link

• n

the basis

• f

whether there is more to sa y ab

• ut

a concept and what the system kno ws This approac h is an amalgamation

• f

the previous t w

• approac

hes whic h allo ws b

• th

the incorp

• ration
• f

adaptiv e h yp erlinks and the dev elopmen t

• f

a user mo del whic h includes in formation

• n

what the user has previously explored in the system Dynamic h yp ertext systems therefore ha v e the adv an tages

• f

b

• th

previous approac hes in terms

• f

exibilit y in information presen tation and giv e the system con trol

• v

er the coherence

• f

that presen tation These sys tems can b e view ed as creating a con v ersation b et w een the user and the system in whic h the system resp

• nds

adaptiv ely to the highlev el discourse planning the user p erforms b y follo wing h yp erlinks

SLIDE 5 Discourse Goals H H H H j

• Kno

wledge Base

• T

ext Planning Agen t

• User

Mo del

• Q

Q Q s

• Plan

Library P P P P i

• Discourse

History A A A A A K

• Discourse

Plan

• Surface

Realisation Agen t

• Phrasal

Lexicon

• HTML

T ext

• W
• rld

Wide W eb View er

• User
• HTML

Command

• CGI

Script

• Figure
• The

Pebai i System Arc hitecture

• What

are the b enets

• f

this tec hnology It is imp

• rtan

t to iden tify the v alue

• f

tailoring tec hnologies W e highligh t a few b enets here

• Customized

information deliv ery Whether p erusing a virtual m useum doing

• nline

shopping c hec king

• ut

the da ys news

• r

searc hing for information

• n

a particular topic a user normally has a set

• f

goals andor in terests whic h impact

• n

what information he migh t consider relev an t Adaptiv e tec hnologies allo w those goals and in terests to act as a lter

• n

what information is actually presen ted to that user

• Reusabilit

y

• f

kno wledge sources The use

• f

adaptiv e tec hnologies enables the same information to b e presen ted in the man y dieren t con texts where that information migh t b e relev an t This a v

• ids

replication

• f

information as it need

• nly

b e represen ted in

• ne

lo cation y et it can b e accessed whenev er it b ecomes relev an t and inserted dynamically

• Av
• iding

information redundancy Information that has already b een presen ted to a user can b e suppressed th us a v

• iding

redundancy in the

• utput

The system can k eep trac k

• f

what information has already b een presen ted to a user p

• ten

tially including in teractions with the user

• f

previous

• ccasions

and can either lea v e it

• ut

en tirely

• r

refer bac k to it in some w a y eg via a h yp erlink Natural language pro cessing tec hniques in particular ha v e certain adv an tages for the dev el

• pmen

t

• f

adaptiv e systems b ecause the c

• ntent
• f

h yp ertexts can b e c hanged as w ell as the links b et w een them F urthermore since an NLG system actually creates texts it has m uc h more con trol

• v

er the coherence

• f

the text

• utput

than a system whic h

• nly

includes

• r

hides

SLIDE 6 paragraphs

• r

another text unit and has no conception

• f

the impact

• f

those decisions

• n
• v

erall coherence for the end user The PEBAI I system Milosa vljevic

• for

example can mak e explicit comparisons b e t w een animals to clarify a particular description but whic h comparisons are made dep end

• n

what information the user has already seen Dynamic generation

• f

the comparison a v

• ids

the need for predenition

• f

ev ery p

• ssible

comparison that migh t b e relev an t to an y user

• the

comparison is c

• nstructe

d b y the system

• n

the basis

• f

stored information rather than hand created b y a h uman and simply retriev ed b y the system The construction

• f

information has certain b enets discussed b y Dale et al in press and summarized here

• Reduced

text construction costs

• V

ariation according to purp

• se
• V

ariation according to user c haracteristics

• V

ariation according to resource b

• undedness
• The

state

• f

the art NLP

• Ob

ject Description T

• date

there ha v e b een sev eral systems implemen ted whic h mak e use

• f

natural language pro cessing tec hniques to incorp

• rate

dynamism in

• b

ject description T ext in these systems is customised in terms

• f

con ten t andor presen tation The PEBAI I system Milosa vljevic T ullo c h and Dale

• in

teractiv ely generates h yp er text descriptions and comparisons

• f

animals The generated text v aries according to the exp er tise lev el

• f

the system user the animals whic h the user can b e assumed to kno w ab

• ut

and the animals for whic h the user has already seen descriptions within the system The

• utput
• f

this system is constructed

• n

the basis

• f

a taxonomic kno wledge base in conjunction with a phrasal lexicon see Section

• and

sho ws ho w texts can b e tailored to the needs and exp eriences

• f

individual users An example

• f

PEBAI Is kno wledge base app ears in Figure

• The

ILEX system Knott et al

• Hitzeman

et al

• aims

to sim ulate the in teraction b et w een a m useum tour guide and a visitor b y dynamically generating h yp ertext pages whic h v ary

• n

the basis

• f

what information a visitor has already seen the discourse history

• and

the goals whic h the system is attempting to ac hiev e in the course

• f

the particular in teraction ILEX utilizes a com bination

• f

abstract represen tations whic h are mapp ed via templates to a sen tence

• r

part

• f

a text and prewritten texts whic h are dynamically inserted at appropriate p

• in

ts in the in teraction In this w a y

• the

qualit y

• f

the

• utput

text is more similar to that whic h a h uman exp ert w

• uld

pro duce than that

• f

most NLG systems y et the

• utput

text is not completely static T ransfer

• f

this metho dology to the m ultilingual con text w

• uld

ho w ev er require all

• f

the prewritten texts to b e translated b y h umans to

• ther

languages and so it is unclear what w

• uld

b e gained b y using this particular tec hnology

• W

e ha v e implemen ted a new system called PO WER PEBAbased On tology With Enhanced Realizations whic h builds

• n

the PEBAI I system Milosa vljevic

• p
• rting

it from the animal domain to the domain

• f

computers whic h migh t b e

• n

exhibition in a m useum Figure

• sho

ws a fragmen t

• f

the kno wledge base for the PO WER system Note that the structure

• f

the en tities precisely parallels that

• f

the PEBAI I system in Figure

• The

c hanges whic h w ere made to the underlying PEBAI I system in v

• lv

ed

• nly

the follo wing

• Construction
• f

kno wledge base en tries for eac h

• f

the

• b

jects in the new domain and inclusion

• f

these

• b

jects in to an

• n

tology

• Enco

ding

• f

the linguistic realizations

• f

the seman tic concepts in tro duced in to the kno wl edge base for the new

• b

jects ie sp ecication

• f

the phrases whic h corresp

• nd

to concepts suc h as generalcomputation see Section

SLIDE 7 hasprop Echidna linnaeanclassification Family distinguishingcharact eri stic Echidna Monotreme bodycovering sharpspines hasprop Echidna geography foundAustralia hasprop Echidna sociallivingstatus livesbyitself hasprop Echidna diet eatsantstermitesearthwo rms

• hasprop

Echidna length quantity lowerlimit unit cm number

• upperlimit

unit cm number

• hasprop

Echidna weight quantity lowerlimit unit kg number

• upperlimit

unit kg number

• Figure
• A

p

• rtion
• f

the PEBAI I kno wledge base

• Addition
• f

some domainsp ecic information ab

• ut

the prop erties in the computer domain whic h should b e describ ed None

• f

the agen ts in the PEBAI I generation system needed to b e mo died The descrip tion and comparison sc hemas surface realisation tec hniques and con ten t selection mec hanisms transferred directly to the new domain In Figures

• w

e can see the kind

• f

descriptions generated b y PO WER from the new kno wledge base en tries with the supp

• rt
• f

the PEBAI I underlying arc hitecture Figure

• sho

ws a description

• f

the Analytical Engine including a comparison to a p

• ten

tial confuser the Dierence Engine also in v en ted b y Charles Babbage whic h is an analogue computer Figure

• sho

ws a description

• f

a General Purp

• se

Digital Computer The generation at this p

• in

t tak es accoun t

• f

the fact that the user w as just visiting the Analytical Engine no de b y noting this in the text and constructing the list

• f

subt yp es

• f

the General Purp

• se

Digital Computer r elative to the no de just visited b y sa ying Apart from the Analytical Engine whic h y

• u

just sa w the General Purp

• se

Digital Computer has the follo wing subt yp es Figure

• sho

ws a comparison

• f

t w

• b

jects generated

• n

the basis

• f

the systems kno wledge

• f

prop erties

• f

those t w

• b

jects In principle a comparison can b e generated for an y com bination

• f
• b

jects for whic h there is information in the system F

• r

a system with man y

• b

jects the use

• f

NLG tec hniques means that eac h comparison do es not ha v e to b e preconstructed b y hand but can b e reasoned ab

• ut

dynamically b y the system and thereb y a v

• ids

a resourcein tensiv e man ual task What w e ha v e learned from the ease with whic h w e w ere able to p

• rt

PEBAI I to a new domain in PO WER is that the agen ts within a mo dular NLG system can b e reused in new domains to the exten t that the structure

• f

the descriptions required in the new domains parallel that

• f

the

• riginal

domain

SLIDE 8 hasprop AnalyticalEngine inventor cbabbage hasprop AnalyticalEngine dateofinvention quantity exact unit year number

• hasprop

AnalyticalEngine purpose generalcomputation hasprop AnalyticalEngine height quantity upperlimit unit cm number

• hasprop

AnalyticalEngine width quantity upperlimit unit cm number

• hasprop

AnalyticalEngine length quantity upperlimit unit cm number

• hasprop

AnalyticalEngine dateofacquisition quantity exact unit year number

• hasprop

AnalyticalEngine sourceofacquisition dswade Figure

• A

p

• rtion
• f

the PO WER kno wledge base Figure

• A

description

• f

the Analytical Engine generated b y PO WER

SLIDE 9 Figure

• A

description

• f

the General Purp

• se

Digital Computer generated b y PO WER whic h tak es in to accoun t that the Analytical Engine no de w as just visited b y the user Figure

• A

comparison

• f

analogue and digital computers generated b y PO WER

SLIDE 10

• A

prop

• sal

for Multilingual Generation

• The

comp

• nen

ts

• f

a m ultilingual dynamic generation system The kind

• f

m ultilingual generation system w e prop

• se

to implemen t will mak e use

• f

curren t NLG tec hniques as in P aris and V ander Linden

• Giv

en an underlying language neutral kno wledge source text can b e pro duced in sev eral languages in parallel b y pro viding the systems with the appropriate linguistic resources In general text for the same discourse goal but in dieren t languages ma y v ary in b

• th

in their discourse structures and in their surface realisations v

• cabulary

and syn tax As a result the arc hitecture for a m ultilingual generation system should allo w for this v ariation as in P aris and V ander Linden

• In

some domains ho w ev er it is p

• ssible

as a go

• d

appro ximation to k eep the disc

• urse

structur e constan t and v ary

• nly

the surfac e r e alisations This is what w e in tend to do in the prop

• sed

system Th us the

• nly

agen t within the NLG arc hitecture whic h will need to c hange is the surface realisation agen t It will need to b e expanded to include a linguistic mo del for eac h

• f

the target languages The grammars

• f

languages dier as do their syn tax and seman tics for individual w

• rds

F urthermore the mapping b et w een concepts and w

• rds

can v ary b et w een languages Clearly realisations in dieren t languages from the same underlying information can dier dramatically in lexical c hoice and syn tactic structure This imp

• ses

certain requiremen ts

• n

the underlying represen tation utilised b y eac h

• f

the agen ts in the NLG system and

• n

the surface realisation agen t

• Language

neutral represen tation The information in the system domain m ust b e represen ted in a languageneutral w a y that is the concepts m ust b e captured with minimal bias to their expression in

• ne

language

• r

another

• V

ariabilit y An y adaptations whic h the system mak es to the information presen ted should b e reected at the lev el

• f

the underlying represen tation so that the v ariabilit y is consisten t across the target languages This includes v ariations in v

• cabulary

c hoice to the exten t that eac h language mak es relev an t and consisten t distinctions b et w een v

• cab

ulary t yp es eg tec hnical vs general terminology

• Structured

mapping The underlying represen tation

• f

information m ust supp

• rt

the realisation

• f

that represen tation in v arious target languages A mapping to particular surface realisations from the represen tation m ust b e dened within the surface realisation agen t for eac h language taking in to consideration the w

• rds

to whic h particular concepts corresp

• nd

and the grammatical structure

• f

eac h target language

• Language

indep endence V ariations in the realisations for dieren t languages whether in surface structure lexical c hoice

• r

the breakdo wn

• f

information in to individual sen tences m ust b e encapsulated in individual language mo dules within the surface realisation agen t The denitions

• f

these pro cesses for eac h target language should b e isolated from the denitions for eac h

• ther

target language These requiremen ts are largely met b y the existing structure w e nd in PEBAI I and PO WER since the kno wledge base in these systems is enco ded in a represen tation language whic h is then mapp ed via the phrasal lexicon see b elo w to surface realizations These systems are therefore a go

• d

starting p

• in

t for the exploration

• f

m ultilingual information presen tation Our h yp

• thesis

based

• n
• ur

exp erience in p

• rting

PEBAI I to PO WER is that through addi tion

• f

languagesp ecic phrasal lexica in to the existing PEBAI IPO WER structure and with the incorp

• ration
• f

sen tence planning pro cedures for target languages

• ther

than English the systems should extend to the m ultilingual con text and should allo w the same range

• f
• utput

v ariation as the

• riginal

systems Th us w e should b e able to r euse the underlying data in the

• riginal

systems for m ultilingual generation There are ho w ev er a few issues deriving from the use

• f

a phrasal lexicon to whic h w e no w turn

SLIDE 11

• Phrasal

Lexica A tec hnique in NLG for handling the mapping from the kno wledge base to linguistic realisation is the use

• f

a phr asal lexic

• n

in whic h linguistic c h unks can b e recorded at a lev el higher than individual w

• rds

The surface realisation agen t then con trols the com bination

• f

these c h unks to form sen tences A lexical en try in this approac h migh t consist

• f

a full phrase whic h corresp

• nds

to a partic ular concept represen ted in the kno wledge base Milosa vljevic T ullo c h and Dale

• argue

that since the kno wledge represen tation in a particular domain

• ften

uses complex elemen ts the linguistic elemen ts to whic h they corresp

• nd

should b e equally complex So for example they ha v e a lexical item whose

• rthograph

y is is a carniv

• re

and eats an ts termites and earth w

• rms

whic h corresp

• nds

to the concept eatsantstermitesearthwo rms in the kno wledge base rather than building up the structure

• f

the v erb phrase from the individual w

• rds

is a carniv

• re

and eats an ts etc They state The use

• f

phrasal lexical items

• f

this kind has t w

• sp

ecic adv an tages Reuse and Eciency If w e rep eatedly realise a seman tic elemen t in the same w a y

• it

is b etter to remem b er this and a v

• id

rebuilding the surface form eac h time Suc h phrasal units m ust exist in the lexicon regardless due to the existence

• f

idioms and

• ther

noncomp

• sitional

linguistic material V ersp

• r
• The

extension

• f

the tec hnique

• f

utilizing phrasal lexica to the m ultilingual con text means dening the lexica

• f

the target languages in terms

• f

phrases and dening grammars whic h con trol the appropriate com bination

• f

these phrases for sp ecic languages This will demand a fair amoun t

• f

linguistic sophistication

• f

the lexicon in

• rder

to accoun t for the realisation

• f

p erson n um b er and gender agreemen t in the languages whic h require it but the grammars utilised b y the surface realisation agen t can b e considerably less complex than grammars whic h start at the lev el

• f

individual w

• rds

and m ust build up all syn tactic structure An imp

• rtan

t issue whic h

• ur

researc h will address is the appropriate lev el

• f

seman tic gran ularit y for the kno wledge represen tation and corresp

• ndingly

for the phrasal lexica for the target languages The breakdo wn

• f

information in to represen tational elemen ts m ust meet t w

• comp

eting desiderata

• Co

v erage Represen tational elemen ts m ust b e negrained enough to capture the full range

• f

concepts and predicates whic h are relev an t to the domain and are lik ely to b e realized in subtle v ariations in the target texts

• Reusabilit

y Represen tational elemen ts m ust b e broad enough to reect generalisations whic h can b e made ab

• ut

related concepts This gran ularit y is furthermore aected b y the in teraction

• f

the kno wledge represen tation and the phrasal lexica since the represen tational elemen ts m ust b e c hosen in suc h a w a y that they consisten tly corresp

• nd

to a particular realisation in a particular language The appropriate lev el

• f

gran ularit y for the represen tation

• f

an individual concept ma y therefore dier for dieren t languages As a result the represen tation and the corresp

• nding

phrasal elemen ts ma y need to b e adjusted for suc h concepts as additional target languages are included Represen tational gran ularit y in the m ultilingual con text will lik ely dier from that in a monolingual con text and precisely in what w a ys is an area

• f

in v estigation w e are in terested in pursuing

• Benets
• f

natural language generation tec hniques Mac hine translation systems can b e view ed as

• ne

approac h to pro viding m ultilingual presen tation

• f

information Ho w ev er for the purp

• ses
• f

dynamic systems whic h incorp

• rate

user tailoring and generate texts

• n

the basis

• f

underlying data this approac h is inappropriate The

SLIDE 12 extension

• f

these dynamic systems to the m ultilingual con text can more eectiv ely mak e use

• f

natural language generation tec hniques a v

• iding

some

• f

the diculties inheren t in mac hine translation systems and taking adv an tage

• f

the underlying structured data Hartley and P aris

• In

particular mac hine translation systems w

• rk

with unrestricted natural language as input while an NLG system creates

• utput

texts

• n

the basis

• f

a constrained predened represen tation language This means that the NLG system do es not ha v e to address issues

• f

co v erage

• r

robustness All

• f

its inputs can b e exp ected to t within a particular framew

• rk

and it will b e designed to generate a particular set

• f

p

• ssible

sen tences from the inputs Within a w elldesigned NLG system no represen tation should b e accessed

• r

constructed b y the con ten t selection

• r

text planning agen ts whic h cannot b e mapp ed to an

• utput

form b y the surface realisation agen t while in a mac hine translation system it is virtually imp

• ssible

to allo w for the full range

• f

v ariation and creativit y whic h ma y exist in natural language inputs Mac hine translation systems need to b e able to b

• th

do natural language in terpretation to establish a represen tation

• f

the information em b

• died

b y the source text and natural language generation to pro duce the target texts NLG systems eliminate the need for an in terpretation step and can fo cus

• n

the generation

• f

high qualit y

• utput

F urthermore the am biguit y

• f

natural language is less

• f

a problem in the generation con text as the input to the NLG system can b e assumed to b e unam biguous due to the design

• f

the represen tation language A certain piece

• f

information in the kno wledge base should p ermit

• nly
• ne

in terpretation The same is clearly not true

• f

a piece

• f

input information that is expressed in natural language An accurate mac hine translation system in a giv en domain dep ends

• n

w ellstructured and easily in terpretable do cumen ts in a giv en source language and an y c hanges that need to b e made to the information to b e con v ey ed m ust b e made in that source language in a w a y that will not break the translation system In con trast an NLG system w

• rks

with a previously structured source whic h can b e c hanged as needed within the constrain ts established b y the represen tational system F

• r

instance the TECHDOC generator Rosner and Stede

• Stede
• has

b een used to p erform m ultilingual generation in English F renc h and German The particular domain in this case w ere main tenance instructions from automobile man uals The aim w as to b e able to generate these instructions in all three languages from the same underlying kno wledge source By ha ving suc h a system man ufacturers w

• uld

no longer need to man ually write instructions ha v e them translated and then reiterate when more c hanges w ere necessary

• All

that w

• uld

b e necessary is main taining the kno wledge base and regenerating the man uals when necessary

• The
• b

vious real b enet

• f

NLG for m ultilingual information presen tation is that it allo ws v ariation in the language

• f

the

• utput

text without requiring prewritten texts in eac h

• utput

language Harnessing the systematicit y

• f

languages for the purp

• se
• f

automatically mapping from a seman tic represen tation to a linguistic realisation means that ev ery text and ev ery p

• ten

tial v ariation

• f

that text for individual users and con texts do es not need to b e created in adv ance b y a h uman text writer but rather can b e generated in the form required b y the NLG system This approac h do es ho w ev er demand the construction

• f

an underlying kno wledge repre sen tation and dev elopmen t

• f

phrasal lexicons at an appropriate lev el

• f

gran ularit y for mapping to individual target languages This w

• rk

can b e arduous and timeconsuming clearly there is a need for the dev elopmen t

• f

tec hniques for automatic acquisition

• f

kno wledge bases and phrasal lexicons for the NLG task Some

• f

these tec hniques are under in v estigation but still are the greatest h urdle for the use

• f

NLG tec hniques Once a kno wledge base has b een constructed ho w ev er it should b e reusable in an y generation system including but clearly not limited to a m ultilingual system for whic h mappings to linguistic realizations for kno wledge base elemen ts ha v e b een dened

SLIDE 13

• Conclusions

A system comp

• sed
• f

in telligen t agen ts em b

• dying

tec hniques from natural language generation can pro vide dynamic exibilit y in information presen tation to accommo date the diering needs

• f

dieren t users while allo wing data reuse In this pap er w e describ ed the PEBAI I system and the p

• rting
• f

that system to a new domain in the PO WER system These are systems whic h dynamically generate descriptions

• f
• b

jects dened in a kno wledge base and in whic h distinct agen tiv e mo dules tak e information ab

• ut

the user and his kno wledge in to accoun t in

• rder

to construct a sp ecic asp ect

• f

those descriptions W e sa w that with a simple substitu tion

• f

domainsp ecic information the

• v

erall functionalit y

• f

the

• riginal

system w as easily main tained W e sp eculated that the same structure can b e utilized for m ultilingual information presen tation requiring

• nly

an extension

• f

the surface realisation agen t to include grammatical mo dels

• f

eac h target language p ending in v estigations in to the implications for kno wledge base structure

• f

the use

• f

a m ultilingual phrasal lexicon References Brusilo vsky

• P
• Metho

ds and tec hniques

• f

adaptiv e h yp ermedia User Mo deling and User A dapte d Inter action

• Sp

ecial issue

• n

adaptiv e h yp ertext and h yp ermedia Dale R J Ob erlander M Milosa vljevic and A Knott in press In tegrating natural lan guage generation and h yp ertext to pro duce dynamic do cumen ts Inter acting with Com puters

• Hartley
• A

and C P aris

• Multilingual

do cumen t pro duction From supp

• rt

for trans lating to supp

• rt

for authoring Machine T r anslation

• Hitzeman

J C Mellish and J Ob erlander

• Dynamic

generation

• f

m useum w eb pages The in telligen t lab elling explorer Presen ted at the In ternational Conference

• n

Museums and the W eb Los Angeles Marc h

• Knott

A C Mellish J Ob erlander and M ODonnell

• Sources
• f

exibilit y in dy namic h yp ertext generation In Pr

• c

e e dings

• f

the th International Workshop

• n

Natur al L anguage Gener ation Herstmonceux Sussex UK pp

• McKeo

wn K R

• T

ext Gener ation Cam bridge UK Cam bridge Univ ersit y Press Milosa vljevic M

• Augmen

ting the users kno wledge via comparison In Pr

• c

e e dings

• f

the th International Confer enc e

• n

User Mo del ling Sardinia Milosa vljevic M A T ullo c h and R Dale

• T

ext generation in a dynamic h yp ertext en vironmen t In Pr

• c

e e dings

• f

the th A ustr alasian Computer Scienc e Confer enc e Mel b

• urne

Australia P aris C and K V ander Linden

• July

DRAFTER An in teractiv e supp

• rt

to

• l

for writing m ultilingual instructions IEEE Computer

• Rosner

D and M Stede

• Generating

m ultilingual do cumen ts from a kno wledge base The TECHDOC pro ject In Pr

• c

e e dings

• f

the International Confer enc e

• n

Computational Linguistics Ky

• to

COLING Stede M

• L

exic al Semantics and Know le dge R epr esentation in Multilingual Sentenc e Gener ation Ph D thesis Univ ersit y

• f

T

• ron

to V ersp

• r

C M

• Contextual

lyDep endent L exic al Semantics Ph D thesis Cen tre for Cognitiv e Science Univ ersit y

• f

Edin burgh

View publication stats View publication stats