ON THE ACQUISITION OF CONCEPTUAL DEFINITIONS 
VIA TEXTUAL MODELLING OF MEANING PARAPHRASES 
BEATE FIRZLAFF and KARIN HAENELT 
Integrated Publication and Information Systems Institute 0PSI) of GMD 
Dolivostral~e 15, D - 6100 Darmstadt, Germany 
firzlaff @darmstadt.grad.de, haenelt@darmstadt.gmd.de 
phone ++49/(0)6151/869-811, fax -818 
1 INTRODUCTION 
Conceptual knowledge bases are core components 
of curlgnt natural language semantic information pro- 
cessing systems. Their coment, structure and size cru- 
cially determines the applicability of these systems, 
and still their development and maintenance is a seri- 
ous bottleneck. Taking into account the enormous 
amount of entries required for practical purposes, 
manual modelling procedures are hardly practicable. 
Much effort is therefore devoted to achieving a con- 
trolled automatic handling of these tasks. 
In tiffs paper a proposal is made to use a meaning 
dictionary as a source of conceptual knowledge, and 
to apply text understanding methods for processing 
this information. These methods include the automa- 
tic acquisition of conceptual definitions from mea- 
ning paraphrases, their transformation into a suitable 
representation for automatic processing, and the au- 
tomatic update of conceptual definitions by means of 
text analysis. 
A meaning dictionary explains the use of natural 
language expressions with respect to tile access they 
provide to world knowledge. In principle, the kzlow- 
ledge accessed and thus represented in a meaning dic- 
tionary includes many of the different kinds of know- 
ledge which are required for semantic information 
processing (i.e. "how nature works, practices, institu- 
tions, facts of nature, regularities, ways of doing" 
\[Searle 80:227\]). 
Current research work on extracting inlbrmation 
from machine-readable dictionaries (cf. e.g. \[Bris- 
coe/Copestake/Boguraev 90\], \[Walker/ Zampolli/ 
Calzolari 88\], \[Alshawi 87\], \[Nakamura/Nagao 88\]) 
has mainly concentrated on semantic relationships of 
the vocabulary and has proposed methods for semi- 
automatic or partial extraction of this information. 
These methods basically rely on syntactic structures 
of meaning paraphrases or database-like coded lea- 
tures and values (see also \[Briscoe/Copestake/Bogu- 
raev 90\]). The task of acquiring conceptual informa- 
tion, however, laces the problem that this information 
in traditional meaning dictionaries is itself conveyed 
in terms of the means natural language provides for 
communicating conceptual information, namely na- 
tural language texts (meaning paraphrases, defini- 
tions, even longer explanatory texts or abbreviated 
textual lorms). Thus understanding conceptual defi- 
nitions of meaning dictionaries requires textual 
mechanisms to be recognized and handled in order to 
analyse the information provided and to construct a 
semantic representation. 
In this paper a new method of textually modelling 
meaning paraplwases is proposed for the acquisition 
of a formal representation of tfie conceptual informa- 
tion conveyed in the meaning paraphrase. It is de- 
scribed how these paraphrases are subjected to a 
manual text analysis process in accordance with the 
KONTEXT model \[Haenelt/K0nyves-T6th 91 \]. The 
principles shown are general principles of text analy- 
sis and are also applied in the KONTEXT system to 
textual analysis of corpora (provided a basic set of 
dictionary entries is available). The conceptual mo- 
delling is based on the model of Semantic Emphasis 
\[Kunze 91\], which so far has been worked out for 
verbs. In order to be able to cover the meaning para- 
phrase as a whole, and thus to provide a basis for auto- 
matic bootstrapping, this approach has to be exteuded 
to modelling other word classes. A proposal for tiffs is 
made in this paper. 
First of all this paper introduces the approach to 
textually modelling meaning paraphrases, then an 
example is given. The concept and text structure re- 
lated information in the system dictionary which is 
used as a basis of acquisition as well as for represen- 
ting the newly acquired entries is presented in detail. 
The application of the dictionary to the analysis of a 
meaning paraphrase is explained. 
Act~.s DE COL1NG-92, NANTES, 23-28 AO~r 1992 l 2 0 9 PROC. Ol: COLING-92, NANTES, AUO. 23-28, 1992 
2 TEXTUAL MODELLING OF MEANING 
PARAPHRASES 
2.1 APPROACH 
The KONTEXT model \[Haenelt/KOnyves-T6th 
91 \] assumes that natural language texts do not com- 
municate complex concepts immediately and at once, 
but rather provide sequential instructions concerning 
how to construct them incrementally. The instruc- 
tions are established by the natural language expres- 
sions, which, used in texts, not onty have the function 
of denoting concepts, but rather can be seen as opera- 
tors which contribute to the construction of thematic 
structures and reference objects. Thematic structure 
and reference objects then guide the access to eon~ 
eepts and the incremental construction of new con- 
cept definitions communicated by the text. The KeN- 
TEXT model structures the information conveyed in 
a text and the information describing its contextual 
organization into five layers of text representation: 
sentence structure, thematic structure, view (on back- 
ground knowledge), and background knowledge. 
In order to construct a conceptual definition in ac- 
cordance with a meaning paraphrase, the paraphrase 
must be analyzed step by step under textual aspects. 
Each natural language expression contributes in- 
formation towards the construction of at least one of 
the four upper layers of text representation. The iu- 
formation of rite expressions of the meaning para- 
phrase is analyzed and a five-layereA multi-state text 
representation is generated. 'ltfis generation includes 
operations which incrementally construct (new) con- 
cept(s) and establish the relationship between natural 
language expressions, reference objects ,and concepts 
via the thematic structure. 
2.2 EXAMPLE 
As an example the fi~llowing meaning paraphrase 
of "blank" is used: "A blank piece of paper has no 
writing or other mark.~ on it." \[COBUILD 87\]. A tex- 
tuM representation of this paraphrase is shown in fi- 
gure 1. The figure illustrates how the conceptual defi- 
nition of "blank" is constructed step by step. It shows 
a text representation with Iigur layers and five states. 
The 'sentence structure' contains the lexemes, their 
dependency structure (and further syntactic features 
not shown here). The information of the 'sentence 
structure' basically lollows the PLAIN-grammar 
\[Hellwig 80, 88\]. The 'thematic structure' traces the 
discourse development. It shows contexts (cL boxes 
1-5 in figure 1; the numbers refer to file discourse 
states of their creation), and reti~rences to contexts 
(represented by the lines between the boxes) which 
correspond to discourse state transitions. The 'refe- 
rence structure" contains the reference objects, their 
Selltence 
Structure 
Thematic 
Structure 
Reference 
Structure 
View 
((blank) piece ) 
of paper 
E+ 
lutn ((no)writing 
E+ 
} 
(on (it) ) 
E- 
rl 
r2- r2 r2 - 
r3 r3--- ...... ' r3 
\[11 \[21 \[31 \[41 \[51 
131" , 
Figure 1." Text~l modelling of t~e meaning paraphrase." "A blank piece of paper has no writing on it" (COBUILD 87) 
AC'YES DE COLING-92, NANq'ES, 23-28 ^OF;f 1992 l 2 1 0 PREC. OF COLING-92, NANfES, AU(L 23-28, 1992 
development and relationship. The layer of 'view' 
shows views on the 'background knowledge' which 
have been established by state transitions. The 'back- 
ground knowledge' (which is shown in figures 2 and 
3) contains conceptual descriptions of prototypical 
situations in accordance with the model of Semantic 
Emphasis \[Kunze 91\]. These descriptions are ex- 
plained in more detail in section 2.3.1. 
The states trace the incremental construction of the fi- 
lial view (\[5\]) of the meaning paraphrase and the 
(re-)definition of the relationship between natural 
language expressions and reference objects. So, 
whilst the dellniendnm "blank" had no conceptual 
definition in the initial state (1), in the final state (5) a 
conceptual definition has been constructed by analy- 
zing the meaning paraphrase under textual aspects. 
2.3 THE KONTEXT DICTIONARY 
The contribution natural language expressions 
make towards the layers of the text representation is 
described as a lexicalized text grammar. The informao 
tion of all the five layers is modelled in feature struc- 
tures which are processed with the CFS-system 
\[K6nyves-T6th 91\] \[B6ttcher 91\]. Due to the forma- 
lism chosen the basic operation is unification. 
2.3.1 CONCEP'I~dAL INFORMATION 
The conceptual modelling of the KONTEXT dic- 
tionary is based on the model of Semantic Emphasis 
\[Kunze 91\]. The core components of this model are 
'basic semantic forms' which describe prototypical 
situations commonly referred to by tile verbs of a verb 
field. A 'basic semantic form' is a proposition con- 
sisting of predicates and (propositional or elementa- 
ry) arguments. So, e.g. the propositional description 
ofa 'write'-situation is: CAUSE ( ACT (x), El" (BEC 
(EXIST(w,t)), BEC (PLACE-ON (q,w)))). This de- 
scription can be paraphrased in the tollowing way: An 
action of 'x' causes a 't' to become to exist as 'w' and 
this 'w' to become placed somewhere such that a rela- 
tion 'ON' holds true between 'q' and 'w'. The scheme 
of 'write' is based on the general scheme of a pro- 
duce-situation. 
Situation descriptions provide the basic schemes 
of the layer of background knowledge of dictionary 
entries. They are used for the construction of the layer 
of 'view' during text analysis. They can immediately 
be applied to the modelling of verbs. For modelling 
further word classes the following extensions are 
made: Similarly to Conceptual Dependency ap- 
proaches (e.g. \]Schank 75\]) in file KONTEXT ap- 
proach 'basic semantic forms' are the backbone of 
conceptual modelling. Verbs and deverbative abstract 
nouns are modelled as referring to prototypical situa- 
tions following the descriptions of the model of Se- 
mantic Emphasis. Prepositions are modelled as deno- 
ting parts of those situations which have actants 
which in the surface form can be realized with the 
preposition in question. Noons are modelled as deno- 
ting participants of situations, and adjectives as deno- 
ting situations focusing on a particular participant. 
The defining situations of nouns and adjectives are 
determined in accordance with the meaning para- 
phrases of \[COBUILD 87\] following the textual 
conditions of applying the situations concerned. CO- 
BUILD definitions are also used in order to describe 
field external distinctions of verbs. 
The formal components of the conceptual repre- 
sentation are the feature structure representation of 
'basic semantic forms', instantiation rules and role 
derivation rules following tile approach of \[Kunze 
911. 
The propositional description of the 'write'-situation 
is modelled in feature structures in tile following way: 
\[ CAUSE@<> 
prep: \[ al :\[ A(Tr@<> 
prop:al :ref: 'x' \] 
a2:\[ET@<> 
prop:\[ al:\[ BEC@<> 
prop:al:\[ EXIST@<> 
prop:\[ al:ref: 'w' 
a2:ref: 't' \]\]\] 
a2: \[ BEC@ <> 
prop:al:\[ PLACE-ON@<> 
prop:\[ al:ref: 'q' 
a2:ref:'w' \]\]1\]\]\]1 
Figure 2: Propositional description of a 'write'-situation 
Predicate names are written in capital letters. @<> 
indicates reference to a type, where <> enclose path 
descriptions. The arguments of predicates are values 
of the feature 'prop'. Elementary arguments corres- 
pond to possible reference objects. 
Depending on the predicates used, instantiation 
rules are applied in order to get a more refined de- 
scription of the situation. E.g. a proposition BEC(A) 
instantiates a monotonic path such that initially 
NOT(A) holds true, mad finally (A) holds true. 
NOT(A) is also called the 'presupposition' of the situ- 
ation, and (A) is (one of) its 'assertion(s)' \[Jung/ 
Kilstner 90\]. The predicate ET has two propositional 
arguments and requires its arguments to be instan- 
tiated at the same time. Thus, either the presupposi- 
tion or the assertion of both arguments is selected. 
For elementary argnments case relations are 
derived in accordance with a purely formal definition 
scheme (Kunze's approach (91) provides a theoreti- 
cal description and formalization of Fillmore's (68) 
deep cases).The resulting case relations of e.g. the 
ACT~ DE COLING-92, NArerES, 23-28 AOt)r 1992 I 2 l l Pane. OF COLING-92, NANTEs, AUG. 23-28. 1992 
'write'-situation are: 'agens of ACT', 'to-object of 
EXIST', 'goal of PLACE-ON', 'goal of EXIST', and 
'to-object of PLACE-ON'. 
A perspective on the situation is determined by em- 
phasis and the selection of actual case relations from 
the derived potential roles. By means of emphasis 
\[Kunze 911 certain partial propositions are moved 
into the foreground. 
2.3.2 TEXT STRUCTURE INFORMATION 
Based on the model of Semantic Emphasis in- 
formation concerning the reference structure and the 
thematic structure can be derived systematically from 
conceptual information. 
Potential reference objects are provided for all pos- 
sible participants of the situation. The roles they play 
in the situation are expressed as case relations. 
Thematic information includes information such as 
which participants can be realized emphatically (not 
emphatically) in which surface form. So, e.g. file pri- 
mary effect of the action of writing (in the prototypi- 
cal case) is that a writing exists, while it is less impor- 
tant where it appears. Therefore at least one of the ar- 
guments of the EXIST-predicate has to be realized 
emphatically (i.e. not by a prepositional phrase) and 
those of the PLACE-ON-predicate are either realized 
without emphasis (by a prepositional phrase) or they 
are not realized at all. This information then provides 
the basis for mechanisms of referring to contexts. E.g. 
only those reference objects which have been realized 
emphatically can later be referred to by an anaphori- 
cal pronoun. 
2.4 APPLICATION OF THE KONTEXT DIC- 
TIONARY TO MODELLING THE MEA- 
NING PARAPHRASE 
The words of the meaning paraphrase ("A blank 
piece of paper has no writing on it") contribute to the 
constitution of the text representation in the following 
way: 
"Has" does not make a contribution to the con- 
ceptual definition. It only "provides a verb for the 
structure" \[COBUILD 87:6671 and indicates the ab- 
stractum "writing". 
"Blankpiece ofpaper" introduces a participant of 
a situation. Its surface form and its syntactic function 
contribute to determining which participant of which 
situation it can be. The defining situation of "paper" 
(cf. meaning paraphrase of "paper") must be com- 
patible with the newly composed situation it becomes 
a participant of (e.g. the situation of this meaning 
paraphrase). Otherwise a further text is requirod in or- 
der to solve this conflict. 
"Writing" as a noun denotes a participant of a si- 
marion, and as a deverbative abstractum at the same 
time introduces the situation of which it is a partici- 
pant. This is the 'write'-situation described in section 
2.3.1. 
"No" does not make a conceptual contribution of 
its own. It, however, has the effect that tile presuppo- 
sition of the 'write'-situation must be selected. 
"On" as a preposition denotes parts of those situa- 
tions which have aclants which in the surface form 
can be realized with this preposition. In this case it is a 
complement of "writing" and serves to introduce the 
actant of "writing" which is determined by the first 
argument of PLACE-ON. As a preposition it 
introduces a non-emphatic realization of the actant. 
"It" as a complement of "on" is identified as the 
'goal of PLACE-ON' of the 'write'-situation. As a 
pronoun it refers to an antecedent which had been 
realized emphatically. In this case this is the subject of 
the sentence. Thus, a reference identity of anaphora 
and antecedent is constructed, and then both expres- 
sions refer to the reference object which denotes the 
'goal of PLACE-ON' of the ' write'-situation. 
In terms of feature structures the result can be ex- 
pressed in the following way: 
\[ sent: 
praed: \[ has 
subje: \[ paper 
re\['." =<refer 1> \] 
trans:\[ writing 
ref: <refer 2> 
quant: no 
caspp: \[ on 
caspn: \[ anaphor' 
ref: =<refer t> \]1\]\] 
refer: 
\[ 1: \[ 5:view:=<view 1> 
10: \[ view:=<view 5> 
role: =<view 5 roles goal place-on> \]\] 
2: 8: \[ view:=<view 5> 
role: =<view 5 voles goal exist> 1\] 
view: 
\[t: l\] 
5: \[ inst: \[ sit-l: I name: negation 
al: \[name: exist-as 
al: ref: =<view 5 roles 
goal exist ref> 
a2: ref: =<view 5 roles 
to-obj exist ref> \]\] 
sit-2: \[ name: negation 
at: \]name: placed-on 
al: ref: =<view 5 roles 
goal place-on ref> 
a2: ref: =<view 5 role"s 
to-obj place-on ref>111 
role.s: \[ goal: \[ exist: \[ ref: =<refer 2> 
emph: plus \] 
AC'~rES DE COLING-92, NAICrES, 23-28 AO~I 1992 1 2 l 2 Pate. OF COLING-92, NANTES, AU~3.23-28, 1992 
place-ml: \[ ref: =<refer 1> 
emph: minus \]\] 
to-obj: \[ exist:\[ ref: \[ \] 
emph: plus \] 
place-on: \[ ref: \[ \] 
emph: minus \]\]\]\] 
background: prop: write@<prop> \] 
Figure3: Feature structure representation of the result 
The value of the feature 'inst' denotes the part of the 
situation description that corresponds to "no writing 
on it". It has been derived by applying instantiation 
rules to the propositional part of the conceptual de- 
scription and by selecting the presupposition. Rela- 
tionships to other layers of text representation are de- 
noted as patti specifications (<>). 
3 EXPLANATION OF THE RESULTING 
CONCEPTUAL DEFINITION 
The value of the feature 'inst' can be regarded as 
the conceptual definition of "blank". The meaning 
paraphrase, however, does not really explain 
"blank". It rather explains what it means lor some- 
thing to be blank. So, the conceptual definition ac- 
quired can be regarded as a rule for composing the ad- 
jective's meaning and a noun's meaning, if one takes 
meaning paraphrases as rule descriptions. The rule 
covers conventionalized readings of "blank" (this 
supports an observation described in \[Boguraev/Pus- 
tejovsky 90\]). Used in texts, however, these readings 
undergo further changes (cf. \[Haenelt 92\]). So, e.g. a 
word can stepwise be related to a complex concept in 
a text and then be used as a term. 
4 CONCLUSION 
In this paper a method has been described for 
transforming the conceptual information of a traditio- 
nal meaning dictionary by means of textual model- 
ling into a suitable representation for automatic se- 
mantic text processing. The concept definitions ac- 
quired can be used for content and text structure ba~d 
retrieval (i.e. access to concepts or relevant text pas- 
sages) or for the update of the dictionary (cf. \[Firzlaff/ 
Haenelt 92\]). The method proposed is also used for 
automatic text analysis, and although it requires a ba- 
sic set of dictionary entries to be transformed by 
manual analysis, its goal is to allow for an automatic 
continuation of this transformation. So, if the basic 
vocabulary of meaning paraphrases has been ac- 
quired it should be possible to analyze the majority of 
meaning paraphrases automatically. The conceptual 
definitions acquired from meaning paraphrases then 
can serve as a basis for analyzing further descriptive 
texts such as e.g. encyclopedias. In this sense the me- 
thod described in this paper is meant to be a contribu- 
tion towatvls the automatic bootstrapping of know- 
ledge base construction. 
5 REFERENCES 
\[Alshawi 87\] Alshawi, Hiyan: txrocessing Dictionary Defi- 
nitions with Phrasal Pattern Hierarchies. In: Computatio- 
nal Linguistics, vol. 13, 1987. 
\[Boguraev/Pustejovsky 90\] Boguraev, Bran; Pustejovsky, 
James: Lexical Ambiguity and rite Role of Knowledge 
Representation in Lexicon Design. In: Proc. COLING 
1990. 
\[B6ttcher 91\] BOUcber, Martin: The CFS System User 
Manual. Technical report IPSI 1990. 
\[Briscoe/CopestakegBoguraev 90\] Bdscoe, Ted; Cope- 
stake, Ann; Boguraev, Bran: Enjoy the Paper: Lexical 
Semantics via Lexicology. In: Proc. COLING 1990). 
\[COBUILD 871 Sinclair, John (ed. in chief): Collins CO- 
BUILD English Language Dictionary. London, Statt~ 
gart: 1987. 
\[Fillmore 68\] Fillmore, Charles: The Case for Case. In: 
Bach, E.; tlarms, R.q: (eds.): Universals in Linguistic 
Theory. New York, 1968, pp. 1-90 
\[Firzlaff/Haenelt 92\] Firzlaff, Beate; Haenelt, Karin: Ap- 
plying Text Linguistic Principles to Modelling Meaning 
Paraphrases. In: Proc. EURALEX 1992. 
\[Haenelt 92\] Haenelt, Karin: Towards a Quality Improve- 
ment of Machine Translation: Modelling Discourse 
Structure and Including Diseourse Development in the 
Determination of Translation Equivalents. In: Proc. 4th 
Int. Conf. on Theoretical and Methodological Issues in 
Machine Translation. 1992. 
\[Haenelt/KOnyves-T6th 91\] Haenelt, Karin; K0nyves- 
T6th, Michael: The Textual Development of Non-Ste- 
reotypic Concepts. In: Proc. 5th conference of European- 
ACL, Berlin 1991. pp. 263-268 
\[Hellwig 80\] Hellwig, Peter: Bausteine des Deutschen. 
Gemmnistisches Seminar, Universitat Heidelberg, 1980. 
\[Hellwig 88\] Hellwig, Peter: ChartParsing according to the 
Slot and Filler Principle. In: Proc. COLING 1988. 
\[Jung/K0stner 90\] Jung, Uwe; Kilstner, Herbert: Seman- 
tische Mechanismen der Negation. Berlin: Akademie- 
Verlag, 1990. 
\[K6nyves-T6th 91\] K6nyves-T6th, Michael: Incremental 
Evaluation of Disjunctive FeatureTerms. Arbeitspapiere 
der GMD 1991. (technical report) 
\[Kunze 91 \] Kunze. Jfirgen: Kasnsrelationen und Seman- 
tische Emphase. Berlin: Akademie-Verlag, 1991. 
lNakamura/Nagao 88\] Nakamura, Jun-ichi; Nagan, Mako- 
to: Extraction of Semantic lnforalation ti'om an Ordinary 
English Dictionary and its Evaluation. In: Proc. CO- 
LING 1988, pp. 459-464 
\[Schank 75\] Schank, Roger C. (ed.): Conceptual Informa- 
tion Processing. Amsterdam: North Holland, 1975. 
\[Searle 80\] Searle, John R.: The Background of Meaning. 
In: Kiefer, E; Bierwisch, M. (eds.): Speech Act Theory 
and Pragmatics. Dordrecht / Boston / London: D.Reidel 
Publishing Company, 1980. 
\[Walker/Zampolli/Calzolari 88\] Walker, Don; Zampolli, 
Antonio; Calzolati, Nicoletta (eds.): Automating the 
Lexicon: Research and Practice in a Multilingnal Envi- 
mnruent. Cambridge University Press, Cambridge 1988. 
ACq'ES DECOLING-92, NAh'n':S, 23-28 AObq" 1992 1 2 1 3 l'ao(:. Ol: COL1NG-92, NANTES, AUG. 23-28, 1992 
