TOWARDS LINGUISTIC KNOWLEDGE DISCOVERY ASSISTANTS: 
APPLICATION TO LEARNING LEXICAL PROPERTIES OF CIIINESE CIIARACTERS 
Gcorges FAFIOTI'E & Fran~:ois TCItEOU 
GETA, IMAG (Universit6 Grenoble 1 & CNRS) ~ BP 53, F-38041 GRENOBIJ2 C&lex 9, 15ance 
gea)rges.fafiotte@imag.ti' / francois.tcheou@imag.fr 
ABSTRACT 
It is highly desirable that users of systems which 
include NLP-based components, ranging from grmnmar- 
checkers to MT systems, can access the underlying 
Linguistic Knowledge Base in a natural and gratifying 
way. Our research aims at developing such Linguistic 
Discovery Assistants by merging hyperdocumcnts, Data 
Base Management Systems and interpretive adaptive 
interfaces. 
We have followed a stepwise approach to the idea in 
the context of the discovery and learning of lcxical 
properties of Chinese eh,'uacters, by developing several 
prototypes. We see this system as a facet of a broader base 
including dictionary knowledge. 
Keywords 
Cooperative Discovery Assistant, I,inguistic Knowledge 
Observatory, Lexical Properties Discovery, Computer 
Aided Learning, Kanji, I lanzi, Chinese Characters 
1. MOTIVATIONS : C OMPUrER A IDED 
DISCOVERY OR LEARNING OF LINGUISTIC 
KNOWLEDGE 
1.1 A need for making linguistic knowledge accessible 
to the user, in Personal MT 
A current trend in Personal Machine Translation tends 
towards opening to the nser the linguistic data that the 
system is operating upon \[1\]. Such 'discoverable' 
environments should allow some free, self-plmmed, or 
coached investigation to users, and provide these in a 
suitable expl~matory fonn with a l,'u'ge part of the linguistic 
material emba)dicd in the personal lingwarc: lexical data 
bases, syntactic patterns or syntactic rules modules, 
semantic conlrastive aspects, etc. 
Our work is oriented towards a particular aspect of 
such 'open lingware': the grasp, evcnlu:dly the learning, 
by monolinguai writers or editors of a document who are 
working in a language they know imperfectly, of the 
lexical proi',crties of the language to be used. 
1.2 A new resource different from dictionaries 
In the context of lexical properties we may at first 
consider dictionaries to be a straight rcsponsc to such a 
demand. They usually require of the user some 
premodelled view of the very organic'alton of the lexical 
data, a pragmatic know-how of their legibility, or real 
mastery in order for the searcher user to perform a sensible 
pruning of the available information. This is particularly 
true with the lexieal properties of the languages we are 
concerned with in this project, Chinese and Japanese. 
Users may experience thc complexity of the process 
when, starting from uncertain or incomplete chunks of 
recollected knowledge, they wish to investigate a word to 
be ,'tscertained, a num~ce to be expressed, an ideogram to 
be remembered. Such situations clearly demand a 
reshaping of dictionaries as interactive knowledge bases, 
and the proposal of components and coopemlive interfaces 
wlfich could offer altenmte access schemes to lexic~d data 
b:t~es, or views of them \[11\]. 
Some integrated systems for I)ialogue Based Machine 
'l'rm~slation intend to provide the author with the means for 
interactive consulting of linguistic facts or rules, for 
instance in the context of lexical or syntactical 
disambiguation or indirect pre-editing of contextual 
semantic features, specific to the text to be composed. The 
I.IDIA architecture \[2\] and the NAI)IA model \[9\] 
certainly illustrate this approach. 
1.3 Discovery Assistants, Cooperative Observatories 
The development of interactive environments for 
monolingnal writers leads to modelling new functions for 
documentation, self-documentation, self-learning and 
management of indivitlualized personal knowledge bases, 
to bc pooled into opcn encyclopaedic 'discovery 
environments', specific components for NLP systems. 
Such technologies as hyperdocuments, multimedia and 
voiced data bases, adaptive interfaces, and the benefits of 
Computer Aided l.earning techniques may merge to offer 
solutions in the reahn of such 'cooperative obserwttorics' 
of linguistic knowledge. 
Our project has a stepwise approach to the idea, in the 
e:tse of the lexical properties of the Chinese ideograms. 
2. PROJECT OUTLINE AND PROTOTYPIN(; 
SCIIEME 
2.1 Comlmter Aided Chinese Character l,earning 
The work we report here stems from the initial 
m()delling of an AAOCC systcm (for 'Apprcntissage 
Assist6 par Ordinatcur des Caract&es Chinois'), intended 
to provide motivated users with an adaptive enviromncnt 
for the atltonomous discovery or rcvicw of character 
properties Idl, with a deliberate restriction to a hanzi / 
kanji subset of characters. 
llan-zi simply means Chinese (~aractel, and kan-ji (which 
alliterates the Jboner, and means the same in Japatwse) t'~'tw to 
a small subset of characters that written Japanese almost 
entirely borlmved from hanzL and fivm the combination of which 
Japanese words are derived. We shall call 'hanzi / kanji' the 
intersection set, that is the hanzi which also are Japanese kataji. 
The conceptu:d model of the lcxical data base 
schematizes different views and levels of investigation of 
the material. 
A first alternative lor the user is to explore a language- 
independent view (Fig. 1) of the characters (intrinsic 
moq~ho-semantic properties of the hanzi / kanji, shared 
by written Chinese, Japanese and Korean), contrasting 
with language-related views (Fig. 2, Fig. 3). 
These are enriched with groups of other clmracter 
properties (phc)nctics, morphological similarities, 
contextual semantics...), all su'ongly relevant to one of 
tim three languages of nse --presently, only with the 
core of the Chinese instanciation. 
287 
Towards linguistic knowledge discovery assistants... G. Ft!\[~otte & F. Tcheou 
,.r~h~lOOUI 41 
fl 
SUN 
somont,c~ \]- son 
- day 
ll~x~l, t2 * 1 \[Mnemos ~\]L~ Thls 
Is intended to be ~ pictograph el the SUN. 
Please remember that any round form Is depicted 
I~ d ¢~ st.!ire Then It IS easy to she the cirela end the 
blfl smile, which characterize our slmplest drawln~s of 
the SUN. 
AS I prlralttv% I~en~ ell~. r 5UI4 or DAY ~r m tong~ vNQti~l 
I1~ the rMvt h. This 14tier melrd nO ~rlves frenl ~ft Md eMrleter , ~all~ ~1~ 
Ilk* 'pay,dh' ilrJ ~rlt k i¢ thrust the ~4rn~, 
StructurI Personal mnemonics T~ d~ #rille.lll~I 
In IM ~;Ird¢;tfler we |~k or pier0 *l I~ sun, tSsra 
~m ~ I~ Is i lp,ll (I ilkolp~) thl~ termini Pa ifl the ¢U1, When V~I ¢1o~,0 *ur ¢~s 
Fig 1: Language-imlependent fundamental itformation Jbr the 
character SUN (o r DA Y) 
fl 
Homographs 
~Isll Fi Rmll 
SUN 
Phunvlirv j ~. 
rl 
..... Semislltlcs I. do u 
~- 2, sun 
thlmology :' 
plNuceathe~t~, erl~i~lllu @,lN*lctrelewtlh ~ddla 
tll4c¢ld~r, ther, ii ~ ",~ll~i 
Ikl sl rulm In ~he ~d ddhe iym~ll~ tim anotylj s~ t~' (:e uf Ihe .~IL Is 
~p LI4144 to eP~pfl~4:ls, 
~ 'Palrsnnal mnemanlr~ \[.Q 
Fig 2," Language-depemtent complementary information 
(Etymology)for the character SUN (or DAY) 
~L_~ 
BRIOHT / 
IIUllll I I( b mln~ 
Seminllca j 1. bright 
"2. t omolToW 
P, 
Main Homophones : min~ 
,.in, ....... ~ ,. y: d,~k 
t.ltlo 
mt.~ : l0 screar, ~:.. mk.y : ctes~ the 
fl'or enllrial~) u~J~',= = to dl~ 
/ \ 
g~ minD: tosng .... + ruing: life 
Fig 3: Ixmguage-dependent complementary information 
(Chinese ttomophones) for the character BRIGItT 
In addition, the user may differenliate between two 
levels of information: a main view condensing 
fundamental properties (Fig. I) and a second 
complementary view with more advanced lexical 
ch,'u'acteristics (Fig. 2, Fig. 3). 
We willingly refer to J.W. lleisig's work \[7\], which 
emphasizes the role of a corpus of predefined mnemonic 
labels attached, one to one, to kanji and possibly to some 
subcharacter morphological components. Such mnemonic 
m,-u-king will supplement the etymology, while enhancing 
the user's ~dmaginative memory>> and strongly rehtying 
visual memory. We also strongly invite the user to 
personalize his knowledge ba~, through adding his own 
mnemonics or imaginative productions. 
2.2 Basic functionalities 
The lexical base initially m(xlelled provides a good 
coverage of character properties \[3\] : 
using one view of the character base, the user may 
explore language-free morphology (pictogram, stroke 
number, overall structural vignette, semantic radical, 
confusing sinfilarities...) and universal semantics, 
on the other hand --m~d on the other view-- an author 
can discover language-relevant morphological 
properties (phonetic component ill its structural 
valency, homographs, positional wtriants for 
compound characters, use in composition...), 
language-tied phonetics (written and voiced pinyin and 
tone, homophones...), h'ulguage-related semautics. 
A very detailed set of structural vignettes (simihu" in 
spirit to Ilalpern's patterns \[6\]) is proposed with a digital 
coding, and validated over some 1500 kanji in the lleisig 
progression. They may deeply improve le,'u'ner recall of 
the over~dl structure of a character. We have lor instance, 
with file grey tint giving the position of the semantic key: 
bright rising sun 
tea tiny 
Fig 4: l(ramples of structural vignettes 
Such patterns may also be invoked by the user for the 
study of compound ohmactors to be derived from a given 
kanji, when recalling some particul,'u" structural model. 
2.3 Multiple prototyping development schenle 
The first prototype implemented the conceptmd m<u.lcl 
of the property base as a highly structured hypertext, using 
llyperC,'u'd on Macintosh. We considered this platform a 
g(~ trade-off between a vahmble interaction management 
system, and a temperate framework for expressing the 
object-oriented and reusable view of the base. 
We then adopted a twofold development scheme, with 
both incremental prototyping ill llyperCard, for surface 
multimodal data and for tile interlace fimctional layer of 
the application, and parallel prototyping on different 
development platlk)nrts, to explore different structuring 
and searching methods for tile character base. 
Three tracks were experimented to express character 
search modalities in more relincd interactive ways: 
first a Data Base Management System approach, 
flmmgh Iwo varianls: (1) ollject-oricnled modelling of 
tile base (in tOOl'S on a Xerox workstatioq), (2) a 
rehltional scheme on a standard medium (an Oracle 
encapsulation in l lypercard on the Macintosh); 
(3) various sketches of a knowledge base were 
prototyped in F'rolog with a simplified user interface. 
3. A D ATA BASE ORIENTED MOI)EL, WITII 
FLI~XIBLE EXPORAT1ON 
The prototype may here administer different users 
(learners, didacticians) with data protection, manage a 
standard static ch,'u'acter properly base, maintain session 
journMs and user profiles, tracing their work in the base. 
3.1 Multlcriterial search of Chlnese Characters 
There are two main access schemes to characters. 
Direct designation is based on a simple selection of the 
character icons Oil character boards. The surface 
organization of file base in character series or lessons 
matches here structural and pedagogical motivations 
initially expressed by a didactician (Fig. 5). 
288 
7bwards linguistic knowledge discovery assistants... G. ~tfiotte & F. Tcheou 
Multicriterial search (MCS) allows the user, starting 
from a partial description of the character, from 
tentatively diserimiuaut properties lie may recall, to 
refine or focus his request. Partially erroneous 
demands should be nmuaged adequately by the system 
(while suggesting dcfault or alternate detenniuatious 
for dubious prol~)sals, or suppressing irrelevant ones). 
The elements remembered or ew~k~ by the learner are 
put forward in a criteria array orgimized in 3 lexical 
property subclitsses: l~teutial characteristics about thc 
sought character itself, about its semantic key, and 
about its phonogr,'unme if it exists (the phonetic marker 
component, very often structurally present). See Fig. 6. 
The grid presents main discrimiuaut criteria (left) and 
secondary characteristics (right). In the MCS of 
complex characters, we put some emphasis on using 
structural vignettes and positional morphology aspects. 
Lolllll~rl 2/.a ( ~,4 r:h~raetera or- prtmttlv~=t ) 
Fig 5: l)irect character selection 
(series of characters with the semantic key MAN) 
Meenhlg: 
Plnyln: 
Radlcal ? Yes NO 
Edlt a ~tructoro vlonotto: 
~ILtm_cm t_l~ _ K,g_ g 
ttpnntnn : 
Plnylo : 
Position In the chnrt~cter : 
E~tlr~.~. ~ - 
i'leanlng : 
PIngln : 
Position Ill tile charocter : 
Indlflororlt 
- -\[Tone : Stroke Nur0hor : 
Select e ~lrtlcturo on the II~t : 
Nelghbour of form : 
IKo!jwotd~ In Ilnemonlcs : \[ 
..... lKeyword~ in P ...... l Notos: 
~0 ko~ In the II~t ! _~ 
Fig 6: ll~e mtdticriteHal search grid 
Results of the sc~wch are available as character icons. 
'lhe user may collect some of them ti~r later study. Direct 
observation of a chmracter, straightaway, is possible too. 
3.2 Ewdutive surface struCtoration for the base, and 
flexible session planning facility 
In its surface design, the standard main clmractcr base 
is originally segmented ill lessons, or collections, 
according to lteisig's view of a pedagogical progression 
for a methodical discovery of kanji. 
On the I)BMS driven prototypes, surface 
recontiguralion of the base is made possible, using in turn 
Ixlth multicriterial search or direct character collecting. 
It may allow dMacticians to express different views of 
the intrinsic property base, to restructure oh;tractor 
lessons for pedagogic;fl reasous, to elaborate alternate 
progression schemes (involving for instance use 
frequency, series with a comnu)n semantic key, series 
with a shared phonogrmnme, ctc). They arc here 
enabled to propose new palettes of 'predcfined lessons' 
with alternate discovery paradigms or mnem(mic 
systems related to various linguistic and cultural vicws, 
on which learners may express preferences its wcll. 
The learner himself may build and maintain one or 
several personal (sub)bases, or collections: series of 
characters that he selects using coherent criteria, which 
he plans to explore in future sessions, collections 
reflecting a personal thematic organisatitm of the 
discovery, simple reservoirs of characters built by free 
picking or rational collecting, through digressive or 
systematic navigation in the bitse. The leluner might 
here express, discover, refine, some personal discovery 
'customs' according to a cognitive style. 
Both types of users are thus allowed (with appropriate 
access rights to such restructuriug resources) to enrich file 
collection of existing views on the property basE, to edit 
and to reshape predcfiued lessons or collections into 
'personal lessons'. 
Along with instant feedback and regular reviewing of 
tile actual work, the systenl here has some inccmivcs to 
more intention-driven, sell'guided learner activity, through 
sllort-tern~ session plmming and long-range curricnhmt 
self-organization. Case by case spontaneous consultatiou 
of lexical information is of course still advised. 
3.3 Observing user activity 
In our view such a function is essential on the way It1 
inteq)reting and rot×lolling user activity. On one of the 
prototypes, the tracing resources providcxl a first basis 
to el~d~)rate llexiblc, analytic and synthetic feed-back 
or witness functions for the user-discoverer interface, 
to build up an infommlion pool about user I)chaviour 
mid discovery smltegies. 
While extracting data from file sessions base, we may 
sketch synthetic session journals, synthetic views on each 
character (or character property) for a user or- I~lr all users. 
3.4 Parallel prototyping 
The DBMS view and the MCS (MultiCritcrial Search) 
were prototyped on diflerent development plat forms. 
a) an Object Oriented prototyping (LOOPS) 
A COCOA rclcase of the AAOCC project was 
rewritten iu a homogeneous Object-Oriented fi'mne h)r a 
co~pus of al×)ut 100 characters. Results and pcrlonnances 
were quite encouraging, though on this small scale model. 
b) a classical relational framework (Oracle) 
Wc also adopttal, on another prototype (CACAO-4), an 
integrative scheme merging two functionally specialized 
environments: IlypcrCard fi)r multimodal interactive 
Ii'out-eud resources, and ml (hacle kernel lor mmlagiug the 
bases and the user queries. We here aimed at exploring 
implementation paradigms for htrgcr scale character bases. 
The system first configurates entities in an Oracle 
property base, while extracting relevant data from the 
llyperCard hypcrdocument fields. At query time, 
arguments sent from the interface layer will generate SQL 
requests. System response is displayed back to the user, 
who then collects characters for litter use, or directly picks 
up n~ded multimedia data on the properties sought. 
~lhough on a still small scale prototype, the relatioual 
DBMS scheme cased data security, coherence qualities, as 
well as some quautitative devclopmeut aspects. 
2~9 
Towards linguistic knowledge discovery assistants... G. Fafiotte & F. Tcheou 
4. EXPERT ASSISTANCE TO CItARACTER 
IDENTIFICATION 
4.1 Expert System oriented schemes 
We prototyped a similar 'scale model' corpus into 
smidl knowledge bases (facts, structural aud other property 
rules), providing for deductive and explanatory fnuctious. 
Data-driven and goal-driven schemes were experimented 
in a small Expert Assistant for Character Identification. 
We try to initiate more interactive multicritcrial 
searches, while coupling very discriminant semantic 
characteristics (the meanings of the expected character, of 
its semantic or phonetic keys), less selective indexiug 
(stroke number, pinyin...), with a tentative iconic 
specification of structural properties. 
When efficient criteria ,are missing, we tbiuk such a 
visual structural recall to be helpful, with or without strong 
spatial positioning and applied to subcomponents with 
semantic or phonetic key functions, through a progressive 
opening or refinement of structural vignettes. 
4.2 Cooperative search 
Later the user should be able to express preferences 
regarding the prompting profile or the search strategies 
adopted by the Discovery Assistant (I)A) in cases of 
underspecified or possibly erroneous queries. The DA 
could group results or hints, in order to prevent over- 
stepped talkative dialogues. System explanation, if 
activated, could justify or illustrate side-hints with details. 
In the Annex example, the user tries to recall a 
ch~waeter ,~.,,, a morphological tree of which is shown 
below (the semantic keys of the subcomponents being 
squared) ---but he actually knows very little of all this. 
~ z~a~utb 
Fig 7: A morphological tree for the character ILLUMINATI~ 
A first scenario (see the Annex) exemplifies a 
cooperative dialogue with a beginner user whose 
spontaneous search strategy will favour visual structural 
characteristics (with no particular attention here to any 
phonetic unit in the churacter). The I)A repeatedly asks for 
any semantic recall of another compouent (a most 
diseriminant criterion), but the user decliues, lie is 
therefore asked and helped to visually refine the overall 
finally I:~. structure, sucessively , ~ then , 
In a second scenario, the user soon proposes (at 
dialogue step 2) a component kmfe ( ~ ),which he figures 
to be present --and which really belongs to the 
investigated character. With the current kauji base, adding 
this very discriminant element directly produces the 
unique final solution: to illuminate. 
In a third scenario, the user erroneously recalls (at step 
2) the possible presence of the radical strengh, ( ~ ), 
instead of the very similar knife radical. The system will 
first exhibit ,an empty respond, which is correct here. But 
knowing about this misleading similarity (a 'faux ,'mils' 
property), it then suggests chm~giug strength into knife. If 
the user acknowledges the proposal, the proper character 
to illuminate is reached right away. 
5. A PROSPECTIVE VIEW 
5.1 Functional development methodology 
In the context of our prototyping effort, we would like 
ideally to design the application with a three-fold 
functional architecture: 
a highly interactive interface layer developed on an 
appropriate authoriug environment, for tile surface 
multim(xlal representation of the lcxical lo:owlege, 
,an object oriented DBMS to express the core of the 
structural knowledge, to implement efficieutly heavy 
data searching and to structure and update the user 
history profiles mid personid bases, 
a declarative or deductive progriunming environment 
or expert system generator, in order to express both the 
strategy m(xlels of a coached or error-compensatory 
multicritcrial character search, and first elements 
towards typed behavioural profiles and users' 
discovery strategies. 
This could possit)ly lead to 'client-server' m'chitectures 
with distributed logical resources in the way of 'white- 
board' schemes \[8\]. lit seems that the iutcropcrability 
expected from multiplatform developmeut cnviroucmcnts 
will further such luuctionally distributed design. 
5.2 Cooperative adaptive accompanying interface 
To summarize, we intend to develop the first draft of 
the exploration assistant, towards 
free surface resUucturing by the leander of his personal 
kuowledge base, according to a thematic or 
methodologic:d view he lkdlows, 
personal management---plauniug, monitoring and 
reviewing-- of sessions or iuquiry sequences, 
synthetic or analytical follow-up of the discovery, 
working on a metaphor of the subbase bciug explored, 
with qualitative indicators on the actual navigation, 
productiou (according to user preferences) of session 
journals, profile status, global cun'iculum surveys, 
issuing of persoualized written, magnetic or audio 
documents, for remediation and iu-depth work. 
5.3 Towards integrated polymorphic or multiple-view 
Lexical Knowledge Bases 
Our system can be viewed as a facet of a broader 
'environment for an encyclop:edic discovery' with other 
mcxles of activity: sell-review, semi-tutored lessons, where 
character thematic 'collections' would drive the discovery. 
It would be desirable to be able to find, througb 
different views, in one and the same knowledge base, 
all the information that the I Ialperu Japanese-English 
dictionary \[7\] offers, with the words built from 
clmraeters, Japanese pronunciation, a sound thesaun~s, 
the data of a large 'Chiuesc-usual language' dictionary, 
character etymology \[ 10\], classical, usual calligrapby, 
a language-independent view of the hanzi / kanji, 
angmented with a progressive and comprehensive 
proposal of mnemonics in lteisig's style \[7\], but 
culturally related to tile user's native or usual language, 
the resources and mod:dities modelled in the AAOCC 
prototypes, for accompanied hyperdocumental 
navigation, expert character identification, for the 
creation and updating of pe~onal subbases or thematic 
character collections, among other features to appear. 
290 
7bwards linguistic knowledge discovery assistants... G. 1,'~ffi'otte & 1': Tcheott 
CONCLUSION 
We advocate the development tff system components 
lot helping authors to access the underlying liuguistic 
knowledge, mnong others in l:'ei~onal or DBMT systems. 
Such Discovery Assistants (DA) slmuld certainly be 
highly cooperative, namely show sensible interactivity 
(within multimodal hyperdocuments and object I)BMS 
frameworks), provide some ways to tentatively adapt to 
users' nmemouic and cognitive customs, and preleralfly 
first be user-tunable: i.e. they could offer means for the 
users themselves to refine and express their prcli:rences in 
terms of search strategies (Slmnlaneous, self-phmned or 
coached), their planning intentions for a working 
sequence, as well ~ts means for an efficient follow-up of 
their activity. DAs should in our view rather first enhance 
Imth user's natural intelligence towards more refleclive 
interactitm m(ules, and user's self-gui&mce aptitudes. 
Iu the framework of a lexical PrOlverty base of hanzi / 
kanji, we have developed, as very first steps, .'t multiple 
prototyping of such functions, while exphMng object 
orienled, relalional, aud dcduclive (rule-driven) schemes. 
We expect progress from patient observation and 
modelling or user activity, and from the availalfility of 
multiplatform sol'tw~tre development tools, merging 
different classes of functionals, heading lowards 
Ixflymoqflfic or multiple-view knowledge bases. 
RlgFERENCES 
\[1\] Boltet Ch. (19~)) 7bwards Personal MT." on some aspects 
of the LII)IA projecr Prc, ceedings of Coling-.90, 08/90, vol. 3/3, 
pp30-35. 
\[2\] l|oltel Ch. & Blanchon 11. (1993) Dialogue based MTJor 
monolingual authors attd the LII)IA project. Proceedings (~f 
NLPRS'93, Fukuoka, Dec 6~7 1993, wfl. 1/1, pp2tJ8-222. 
\[3\] li'aflotte G. (1990) A Self-IJmrning System Jot" Chinese 
Characters. Proceedings of COLING 90, IIelsinki, Aug 20~25 
le~)0, II. Karlglvn ed., ACL, vol. 3/3, pp351-354 . 
\[4\] Fafiotte G. (1990) Appretttissage assist(par ordinateur 
des earact~res ehinois. Proceedings of 10~mes Journdes 
Internationales "l.es systt~mes experts et leurs applications", 
Avignon, May 28-Jun 1 1990, liC2, vol. 8/8, pp61-70. 
I5\] (;luse 1). (1988) Conqtuter-Based hltelligent Tutoring for 
I:oreign lztnguages. Proceedings of Asia-Pacific Conference on 
Computers in Education, Shangtmi China, Oct 88. 
\[6\] ilMperla J. (19911) New Japanese-l,2t~gli,rh (~taracter 
Dictionary. Kenkyusha, Tokyo, 1992 p. 
\[71 tleislg J. W. (1977) Remembering the Kanji I ~ A 
complete course on Japanese characters. Japan Publications 
Trading Co, Tokyo, 2 vol., 495 p. 
181 Sellgm'an M. & lhfitet Ch. (1994) 77w 'Whiteboard" 
At('hilet'lare: a way to itlteqrale helerogetteolts comltotletllS Of 
NLI' systems. Proceedings of Coling 94, Kyoto, Aug 5-9 1994. 
\[91 ,";6rzsset (;. & Blanc 1~;. (1993) Utte approche par 
accq~tion pour les bases lexicah's multilingues. Pro¢. of 1'.-'I'A-- 
TAt) '93, Montrdal, Sep 3it-Oct 19!)3, Ihfiversit6 de Monh'dal. 
\[10\] Wleger l,. (1972) Caract~)res chinois. Etymologie. 
Graphies. l.exique, Kuangchi Press, Taichung. 
\[11\] Yokol T, (1993) Very LatT,,e-Scale Knowledge Bases 
Embodying Intelligence Space. Proceedings of KB&KS'93, 
Tokyo, Dec 1-2 1993, JIPDI;;C, ppl 1-20. 
ANNEX An example of cooperative search, through stepwk~e structural vignette relinement (First ~e~mrio) 
"lhe u~r interface was scl~matized here ft~' a more compact linear reading. I leading numlv, trs indicate dialognc stele;. DA strolls for 
Di,~overy AssistanL 'User:' ann~xmces a seleetimlor an entry, 'I)A-->' a I)A i~'ompting al~t 'DA.,.' a DA reslmnse. Bold typing 
sl~',ws user's selections or entries. ,";K and PK reslx~ctively s land for Semantic Key and Phtmelic Key, two possible fu nctkms I'tx" .,.xmm 
colnllO\[le, l|ts. ~ix lnaill lx)sitions ftn" a COIIIIXHlelIt inside a character are: Sou ill. N~th. West, East, Central and l!xterna\]. Grey tint in Ihe 
patterns indicates the actual i~sition ofa.fire b;etnantic Key, 
0. DA--> Any component known ? If not, do you have structure proposals ? Any other criterion ? 
0. Uscr: YeS No Yes Nil Yes Nil 
1.1)A--> Component...Memfing ? l:unction ? Position ? Vignette list ? Any other crilerion ? 
1. User: fire SK PK South Yes No Yes Nil 
1. DA... Ok, I have several l~rOlmsals, with different vignettes ... for the global structure 
2. DA--> Any oilier COmlmnent known ? If not, 1 have the following vignette pr!\])osals ...for the global structure : 
2. User: Yes NO 1 L-~ 2 3 4 
2. DA... Ok, 1 prol×~se to bttrtt .,.for "2 , and.., l have several prol)osals ... for \[ El 
3. Uscr: Yes N 0 
ES\] 
3. DA--> Any olher colnponezli known '? If nol, I have relined slrttclmes ...for 1 r---I 
\[ILl 3. User: Yes Nil 1.1 USSI 1.2 IS~ 1.3 c==r-~ 
\[qLl 3. DA... Ok, I have 2 solutions ...lot 1.2 EEl. 
\[3H 
4. DA--> Any other eomlXment known ? If not, I have relined slruelures ...for 1.2 ff~l : HH H~ 
4. User: Yes Nil 1.2.1 ~ 1.2.2 ~_~ 
4. I)A... Ok, I prolmse .~ to illuminate ..,for 1.2.2 ~. 
Renmlrks 
As a Semantic Key, rite may have alterm~tive piclt~grlmis such its # l iuM #2. With rile current kanji blt~, clx)osing vignetle 3 (#3) at 
step 2 wotlkl lead to #4, and vignette 4 (#5) t~ #6. At step 3, clumsing 1.l (t17) its a final refined structure would yield ~gd, #9 and #I0, 
while chcx~sing 1.3 (# 11 ) wouH bring in # 12. At step 4, ~,lecling film strtbcture 1.2.1 (#13) would lead to #14. 
#1 #2 #3 tt4 ashes #5 #6 charcoal #7 18 inflammation #9 disaster #tO to char #l I #112 spot #13 #14 ardent 
291 
