Adjective based inference 
Marilisa Amoia
INRIA/Universit´e de Nancy 1 &
University of the Saarland
Saarbr¨ucken Germany
amoia@coli.uni-sb.de
Claire Gardent
CNRS/Loria
Campus Scientifique BP 239
54506 Vandoeuvre-les-Nancy, France
claire.gardent@loria.fr
Abstract
In this paper, we propose a fine grained
classification of english adjectives geared
at modeling the distinct inference patterns
licensed by each adjective class. We show
how it can be implemented in description
logic and illustrate the predictions made
by a series of examples. The proposal has
been implemented using Description logic
as a semantic representation language and
the prediction verified using the DL theo-
rem prover RACER.
Topics: Textual Entailment, Adjectival Semantics
1 Introduction
Understanding a text is one of the ultimate goals
of computational linguistics. To achieve this goal,
systems need to be developed which can construct
a meaning representation for any given text and
which furthermore, can reason about the meaning
of a text. As is convincingly argued in (Ido Dagan
and Magnini, 2005), one of the major inference
task involved in that reasoning is the entailment
recognition task:
Does text T1 entail text T2?
Indeed entailment recognition can be used to
determine whether a text fragment answers a
question (e.g., in question answering application),
whether a query is entailed by a relevant document
(in information retrieval), whether a text fragment
entails a specific information nugget (in informa-
tion extraction), etc.
Because the Pascal RTE challenge focuses on
real text, the participating systems must be robust
that is, they must be able to handle unconstrained
 We thank la R´egion Lorraine, INRIA and the University
of Sarrebruecken for partially funding the research presented
in this paper.
input. Most systems therefore are based on sta-
tistical methods (e.g., stochastic parsing and lex-
ical distance or word overlap for semantic simi-
larity) and few provide for a principled integra-
tion of lexical and compositional semantics. On
the other hand, one of the participant teams has
shown that roughly 50% of the RTE cases could
be handled correctly by a system that would ade-
quately cover semantic entailments that are either
syntax based (e.g., active/passive) or lexical se-
mantics based (e.g., bicycle/bike). Given that the
overall system accuracies hovered between 50 and
60 percent with a baseline of 50 %1, this suggests
that a better integration of syntax, compositional
and lexical semantics might improve entailment
recognition accuracy.
In this paper, we consider the case of adjectives
and, building on approaches like those described
in (Raskin and Nirenburg, 1995; Peters and Pe-
ters, 2000), we propose a classification of adjec-
tives which can account for the entailment patterns
that are supported by the interaction of their lexi-
cal and of their compositional semantics. We start
by defining a classification schema for adjectives
based on their syntactic and semantic properties.
We then associate with each class a set of axioms
schemas which translate the knowledge about lex-
ical relations (i.e. antonymy) the adjectives of the
class are involved in by extracting this information
from WordNet (Miller, 1998) and a set of seman-
tic construction rules and we show that these cor-
rectly predicts the observed entailment patterns.
For instance, the approach will account for the fol-
lowing (non)-entailment cases:
(1) a. John frightened the child
j= The child is afraid
150% of the cases were true entailment and 50% were
false ones, hence tossing a coin would get half of the cases
right.
20 KRAQ06
b. Peter claims that John is a murderer
j= John is an alledged murderer
6j= John is a murderer
c. This is a fake bicycle
j= This is a false bike
j= This is not a real bike
6j= This is a bike
d. John is not awake
j= John sleeps
6j= John does not sleep
The approach is implemented using Description
Logic as a semantic representation language and
tested on a hand-built semantic test suite of ap-
proximately 1 000 items. In the latter part of the
paper we discuss this testsuite and the philosophy
behind it.
2 A fine grained classification for
adjectives
As mentioned above, we propose a classification
of adjectives based on their lexical, their model
theoretic and their morpho-derivational properties.
To facilitate the link with compositional semantics
(the construction of a meaning representation for
sentences containing adjectives), we also take into
account syntactic properties such as the predica-
tive/attributive or the static/dynamic distinction.
We now detail each of these properties. The over-
all categorisation system is given in Figure 1.
2.1 Model theoretic properties
The main criteria for classification are given by
(Kamp, 1975; Kamp and Partee, 1995) seman-
tic classification of adjectives which is based on
whether it is possible to infer from the Adj+N
combination the Adj or the N denotation.
Intersective adjectives (e.g., red) licence the
following inference inference patterns:
A + N j= A
A + N j= N
For instance, if X is a red car then X is a car and
X is red
Subsective adjectives (e.g., big) licence the
following inference pattern:
A + N j= N
For instance, if X is a big mouse, then X is a mouse
but it is not necessarily true X is big
Privative adjectives licence the inference pattern:
A + N j= :N
For instance, if X is a fake gun then X is not a gun
Plain non-subsective adjectives (e.g., alledged)
do not licence any inference
For instance, if X is an alleged murderer then it is
unknown whether X is a murderer or not
2.2 Lexical semantics
From the lexical semantics literature, we take
one additional classification criterion namely
antonymy. As described in (Cruse, 1986), this
term covers different kinds of opposite polarity re-
lations between adjectives namly, binary opposi-
tion, contraries and multiple oppositions.
Binary oppositions covers pairs such as wet/dry
which license the following inference pattern:
A1  :A2 ^ :A1  A2
So that in particular:
wet  :dry ^ :wet  dry
Contraries are pairs such as long/short where the
implication is unidirectional:
A1 j= :A2 ^ :A1 6j= A2
A2 j= :A1 ^ :A2 6j= A1
and in particular:
long j= :short^:long 6j= short
short j= :long ^:short 6j= long
Multiple oppositions involve a finite set of adjec-
tives (e.g., linguistic/economic/mathematical/... )
which are pairwise mutually exclusive. For a set
of opposed adjectives A1 : : : An, the following ax-
ioms schemas will be licensed:
8i; j s:t: 1  i; j  and i 6= j
Ai j= :Aj and :Ai 6j= Aj
2.2.1 Derivational morphology
We also take into account related forms that is,
whether there exists a verb (Va) or a noun that is
semantically related to the adjectives being con-
sidered. Moreover, for nominalizations we distin-
guish whether the morphologically related noun is
an event noun (Ne), a noun denoting a theta role
of the related verb (N ) or a non-event noun (Na).
As we shall see, this permits capturing entail-
ment relations between sentences containing mor-
phoderivational variants such as for instance :
21 KRAQ06
(2) a. John is asleep (Adj ! Va)
j= John sleeps
b. John is absent (Adj ! N )
j= John is the absentee
c. John is deeply asleep (Adj ! Ne)
j= John’s sleep is deep
2.2.2 Syntactic properties
To better support the syntax/semantic interface,
we refine the adjectives classes distinguishable on
the basis of the above criteria with the following
syntactic ones taken from (Quirk et al., 1985).
Attributiveness/Predicativeness. English adjec-
tives can be divided in adjectives which can be
used only predicatively (such as alone), adjectives
which can be used only attributively (such as me-
chanical in mechanical enginner) and adjectives
which can be used in both constructions such as
red.
Modifiability by very. We distinguish between
adjectives such as nice which can be modified by
very (i.e. very nice) and adjectives such as alleged
which cannot (*very alleged).
Gradability. We distinguish between adjectives
such as big which express gradable properties and
have comparative and superlative forms (bigger,
biggest) and adjectives such as rectangular which
don’t (*more rectangular).
Staticity/Dynamicity. Dynamic adjectives can be
used in imperative constructions and in the pro-
gressive form (Be reasonable, He is being reason-
able), static adjectives cannot (*Be short, He is be-
ing short).
3 Semantic Classes and textual
entailment recognition
In order to build our classification, we have anal-
ysed a set of about 300 english adjectives each
of which was manually mapped to the WordNet
synset correspondent to the more frequent mean-
ing of the adjective. In some case, when an ad-
jective presents polysemic forms which belong to
different semantic classes more than one form has
been considered. For example, for the adjective
civil we consider two senses/forms civil1 (syn-
onym of polite, as in civil man) and civil2 (as in
civil engineer) which belong to different semantic
classes, the first being intersective and the second
subsective. As Figure 1 shows, the proposed clas-
sification includes 15 adjective classes, each with
distinct syntactic and semantic properties.
To account for these differences, we define for
each class a set of axiom schemas capturing the
model theoretic, lexical semantics and morpho-
derivational properties of that class. Lexical se-
mantics and morpho-derivational information are
derived from WordNet. For example, the axioms
describing antonymy are obtained by extracting
from WordNet the antonyms of a particular adjec-
tive and then by considering the direction of the
entailment relevant for the class the adjective be-
longs to:
asleep  wake vs. polite <rude
Morpho-derivational information are derived from
WordNet by extracting the derivationally related
forms for the given adjective and then iterating the
extraction on nouns and verbs in order to obtain
information about their antonyms and hyponyms.
For scalar adjective like tall, WordNet contains
also a relation is a value of which offers a
pointer to the noun concept the adjective is a value
of. Moreover, WordNet links the noun concept to
a list of attributes which describe the scalar prop-
erty it represents. For example, the adjective tall
is a value of fstature,heightg and attributes
offstature,heightg are tall and short.
Based on some basic syntactic patterns, we then
show that these axioms predict the observed tex-
tual entailment patterns for that class.
Before we illustrate this approach by means of
some example, we first show how we capture log-
ical entailment between NL semantic representa-
tions in a description logic setting.
3.1 Using description logic to check
entailment between NL sentences
As argued in (Gardent and Jacquey, 2003), de-
scription logic (DL) is an intuitive framework
within which to perform lexical reasoning: it is
efficient (basic versions of description logics are
decidable), it is tailored to reason about complex
taxonomies (taxonomies of descriptions) and it
is equipped with powerful, freely available auto-
mated provers (such as RACER, (Volker Haarslev,
2001)). For these reasons, we are here exploring a
DL encoding of the entailment recognition task for
the set of examples we are considering.The partic-
ular language we assume has the following syntax.
C; D ! Aj>j?j:A j C u D j C t D j 8R:C j 9R:C
The semantics of this language is given below with
 the domain of interpretation and I the interpre-
tation function which assigns to every atomic con-
22 KRAQ06
Adjective Class Predicative/Attributive Modifiable by very Gradability static/dynamic Antonymy Related forms Semantic class
Class 1: afloat predicative-only - - static multi-opposition Va, Ne, N intersective
Class 2: asleep predicative-only + - static binary-opposition Va, Ne, N intersective
Class 3: polite both + + dynamic contraries Na intersective
Class 4: dry both + + static binary-opposition Va, Ne, N intersective
Class 5: open both - - dynamic binary-opposition Va, Ne, N intersective
Class 6: male both - - static multi-opposition Na, Ne, intersective
Class 7: authentic both + - static binary-opposition Ne intersective
Class 8: big both + + static contraries Ne subsective
Class 9: good both + + dynamic contraries Ne subsective
Class 10: cultural attributive-only - - static multi-opposition Na subsective
Class 11: recent attributive-only + - static multi-opposition Ne subsective
Class 12: fake both - - static binary-opposition Va,Ne privative
Class 13: former attributive-only - - static multi-opposition privative
Class 14: questionable both + - static contraries Va, Ne plain non-subsective
Class 15: alleged attributive-only - - static contraries Va plain non-subsective
Figure 1: Classes of Adjectives
cept A, a set AI   and to every atomic role R
a binary relation RI     .
>I =  
?I = ;
(:A)I =  nAI
(C u D)I = CI \ DI
(C t D)I = CI [ DI
(8R:C)I = fa 2  j 8b(a; b) 2 RI ! b 2 CIg
(9R:C)I = fa 2  j 9b 2 CI ^ (a;b) 2 RIng
Now one basic problem with using DL to check
entailment between NL expressions, is that DL
formulae are “directional” in that they refer to a
given set of individuals. For instance the sentence
The boat is floating might be represented by either
of the two formulae given in 3 but these two for-
mulae do not stand in an entailment relation (since
they refer to different kind of objects namely float-
ing event of a boat in 3a and boats that float in 3b).
(3) a. float u9theme.boat
b. boat u9theme 1.float
To remedy this shortcoming, we introduce the
notion of a rotation. Given a DL formula which
only contains conjunction (disjunction is trans-
lated in DL as different formulas)
 = ui=1;n Eventi uj=1;m 9Rj.Typej
a rotation of this formula is defined as:
1.  
2. 8j 2 f1; :::; mg :
Typej u 9R 1j .(ui=1;nEventi u1<k<j;j<k<m
9Rk.Typek)
so that the formula:
Event1u Event2 u :::u Eventn u9R1.Type1 u9R2.Type2 ...
u9Rn.Typen
corresponds to the following n Rotations each of
which describe the same situation from the point
of view of a particular type
0. Event u9R1.Type1 u9R2.Type2 ... u9Rn.Typen
 Event
1. Type1 u9R 11 .(Event u9R2.Type2 ... u9Rn.Typen)
 Type1
2. Type2 u9R 12 .(Event u9R1.Type1 ... u9Rn.Typen)
 Type2
...
n. Typen u9R 1n .(Event u9R1.Type1 ... u9Rn 1.Typen 1)
 Typen
So for example, the sentence Mary knows that
John is the inventor of the radio will be repre-
sented as a predicate logic formula
9x1mary(x1) ^9x2john(x2) ^9x3radio(x3) ^9e1know(e1) ^
9agent(e1; x1)^9topic(e1; e2)^9e2invent(e2)^agent(e2; x2)^
patient(e2; x3)
the denotation of this PL formula corresponds to
the set of individuals fx1; x2; x3g[fe1; e2g. The
corresponding DL representation will be the un-
derspecified representation
know u9 agent.mary u9 topic.( invent u9agent.john u9 pa-
tient.radio)
the denotation of which corresponds to the set
fe1g and all its rotations which permit to access
the other sets of individuals asserted in the sen-
tence. Thus for example, the set fx1g which
describes the individual Mary can be accessed
through the following rotation:
Rotation1: mary u9 agent 1.(know u9 topic.( invent
u9agent.john u9 patient.radio))
Finally, we say that an arbitrary for-
mula/representation  1 implies the formula
 2 iff it is possible to find a rotation Rotationi of
 1 the denotation of which describes a subset of
the denotation of  2:
Definition
 1 j=  2 iff 9i:Rotationi( 1) v  2 (1)
23 KRAQ06
3.2 Example class axioms and derivations
We now illustrate our approach by looking at two
classes in more detail namely, class 1 and class 8.
3.2.1 Class 1
Syntactically, Class 1 contains adjectives like
adrift,afloat,aground which can only be used pred-
icatively, are non gradable and cannot be modified
by very. Semantically, they behave like intersec-
tive adjectives which enter in multiple opposition
relations with other adjectives. They are further-
more morphologically derived from verbs and can
be nominalized. To reflect these semantic proper-
ties we use the following axioms.
Model theoretic semantics. Adjectives of class
1 are intersective adjective. They will thus li-
cence the correponding inference patterns namely:
A + N j= A (2)
A + N j= N (3)
Lexical semantics. Adjectives of class 1 enter in
multiple opposition relations. Hence For instance:
afloat j= : aground ^: afloat 6j= aground
aground j= : afloat ^: aground 6j= afloat
sunken j= : afloat ^: afloat 6j= sunken
afloat j= : sunken ^: sunken 6j= afloat
Morpho-derivational semantics. Adjectives in
Class 1 can be related to both nouns and verbs.
Thus, for example the adjective afloat in WordNet
is related to the noun floating which is related to
the verb float, by assuming that the semantics as-
signed to the verb float is float(e), theme(e,a), the
adjective afloat is assigned the following seman-
tics:
afloat  9 Theme 1.float
This is encoded in the following axiom schemas:
MDR 1. Adj1 <: Adj2 If Adj1 = Anto(Adj2)
e.g., afloat <: sunken
MDR 2. Adj1  9 Theme 1.V1 If Adj1 is related to V1
e.g.,afloat  9 Theme 1.float
MDR 3. V1 <: V2 If V1 = Anto(V2)
e.g., float <: sink
MDR 4. N1  V1 If Adj1 is related to an evt denoting N1
e.g., floating  float
MDR 5. N1 <: N2 If N1 is an antonym of N2
e.g., floating <: sinking
MDR 6. N11  9 Theme 1.V1 If Adj1 is related to a
noun N11 denoting the theme role of the verb V1
e.g., floater  9 Theme 1.float
We make the following assumptions about the
syntax/semantic interface that is, about the seman-
tic representations associated with given sentence
patterns.
SCR 1. NP toBe Adj
ADJ u NP
SCR 2. NP toBe clearly Adj
ADJ u NP
SCR 3. Ni[+event] of NP is clear
V i u9theme:NP
SCR 4. Nii[-event] is clear
9theme 1:V i
SCR 5. NP toBe V[+ing].
V u9Theme:NP
Given the above axiom schemas and semantic
constructions rules, the following inference pat-
terns can be handled:
1. ADJ1 + N j= N
Ex. This boat is afloat. j= This is a boat.
2. ADJ1 + N j= ADJ1
Ex. This boat is afloat. j= This is afloat.
3. ADJ1 + N 6j= : N
Ex. The boat is afloat. 6j= This not a boat.
4. ADJ1 + N j= : ADJ2 u N
Ex. The boat is afloat. j= The boat is not sunken.
5. : ADJ1 + N 6j= ADJ2 u N
Ex. The boat is not afloat. 6j= The boat is sunken.
6. ADJ1 + N j= N u9theme 1:V 1
Ex. The boat is afloat. j= The boat is the floater.
7. ADJ1 + N j= V1 u9theme.N
Ex. The boat is afloat. j= The boat is floating.
8. ADJ1 + N j= N1 u9theme.N
Ex. This boat is clearly afloat. j= The floating of the
boat is clear.
9. ADJ1 + N j= N u9theme 1.N1
Ex. This boat is clearly afloat. j= The floating of the
boat is clear (or the boat is the floating object).
10. : (ADJ1 + N) j= : (V1 u9theme.N) 6j= : N
Ex. This is not a floating boat. 6j= This is not a boat.
11. : (ADJ1 + N) 6j= : Adj1
Ex. This is not a floating boat. 6j= This is not afloat.
12. : (ADJ1 + N) 6j= : V1
Ex. This is not a floating boat. 6j= This is not floating.
13. : (ADJ1 + N) 6j= : N1
Ex. This is not a floating boat. 6j= This is not a floating.
14. : (ADJ1 + N) 6j= :9 theme 1.V1
Ex. This is not a floating boat. 6j= This is not the floater.
15. : (ADJ1 + N) 6j= :9 theme.N
Ex. This is not a floating boat. 6j= This is not a floating.
24 KRAQ06
In the inference patterns 10 to 15, the negation
of the adjective-noun compound : (ADJ1 + N) is
syntactically blocked, as the adjectives in this class
are used predicative only, however the equivalent
representation V1 u9theme.N can be used to mo-
tivate the inferences.
The following show in more detail how the first
three of the above (non) entailments are recog-
nised.
(4) a. The boat is afloat.
b. j= The boat is floating.
4a  Boat u Afloat (by SCR 1) A
4b  Float u9Theme:Boat (by SCR 5) B
Afloat  9Theme 1:Float (by MDR 2) C
1  Boat u9Theme 1:Float (from A and C) D
D j= B (By Defn 1) E
(5) a. The boat is afloat.
b. j= The boat is the floater.
5a  Boat u Afloat (by SCR 1) A
5b  Boat u9Theme 1:float (by SCR 4) B
Afloat  9Theme 1:Float (by MDR 2) C
A j= B (from B und C) D
(6) a. The boat is afloat.
b. j= The boat is not sinking.
6a  Boat u Afloat (by SCR 1) A
6b  : sink u9Theme:boat (by SCR 5) B
Afloat  9Theme 1:Float (by MDR 2) C
Boat u9Theme 1:Float (from A and C) D
float u9Theme:boat (By Defn 1) E
E j= B (by MDR 1) F
3.2.2 Class 8.
Class 8 contains adjectives like
big,fast,tall,deep which can be used attribu-
tively and predicatively, are gradable, can be
modified by very. Semantically, they are classified
as subsective adjectives and their antonyms are
contraries. They are morphologically related
to nouns which describe the particular property
denoted by the adjectives and to nouns of which
they are attributes.
Model theoretic semantics. Adjectives of
class 8 are subsective adjective. They will thus li-
cence the correponding inference patterns namely:
A + N 6j= A (4)
A + N j= N (5)
Lexical semantics. The Adjectives of class 8 en-
ter in contrary opposition relations. Hence, the fol-
lowing axioms schemas will be licensed:
Ai j= :Anto(Ai) and :Ai 6j= Anto(Ai)
(6)
For instance:
long j= : small ^: long 6j= small
deep j= : shallow ^: deep 6j= shallow
Morpho-derivational semantics. Adjectives in
Class 8 can be related to nouns but not to
verbs. Moreover, such adjectives are mapped
in WordNet to noun concepts through two dif-
ferent links: derivationallyrelated to
and is a value of. For example, the adjec-
tive tall in WordNet is derivationally related to the
noun tallness and is a value of the concept noun
height. The adjectives in this class describe grad-
able properties so that their semantics corresponds
to:
has-property(Related Noun u9has-measure.Top)
in which the role has-measure account for the
value of the scalar property described by the adjec-
tive, which remain underspecified (Top) if the ad-
jective is used without a reference to the value of
measure. When the value of the measure is speci-
fied, for example by combining the adjective with
a noun, as for example in This is a tall man, then
the noun is assigned as a value of the measure role:
man u9has-property.(tallness
u9has-measure.man)
which translate This is tall as a man.
This is encoded in the following axiom
schemas:
MDR 1. Adj1 <: Adj2 If Adj1 = Anto(Adj2)
Ex. tall <: short
MDR 2. Adj1 <9 has property.(N1 u9has measure.Top)
If Adj1 is related to a noun N1 denoting the property
described by Adj1
Ex. tall < 9 has property.(tallness
u9has measure.Top)
MDR 3. N1 <: N2 If N1=Anto(N2)
Ex. tallness <: shortness
MDR 4. N1  N’ u9has value.Adj1
If Adj1 is an attribute of the noun N’
Ex. tallness  height u9has value.tall
MDR 5. N2  N’ u9has value.Adj2
If Adj2 is an attribute of the noun N’
Ex. shortness  height u9has value.short
MDR 6. N1 < N’ If N1 is an hyponym of N’
Ex. tallness < height
25 KRAQ06
MDR 7. N2 < N’ If N2 is an hyponym of N’
Ex. shortness < height
MDR 8. Adj11 < Adj1 If Adj1 is a
scalar attribute with value less then Adj11 (hyponymy
is not defined for adjectives)
Ex. giant < tall
For the moment, we don’t account for the se-
mantics of comparatives forms of adjectives but
we will do that in the feature, by also introducing a
representation for scales as described in (Kennedy,
2005).
We make the following assumptions about the
semantic representations associated with basic
sentence patterns.
SCR 1. NP toBe Adj
NP u9 has property.(N1 u9has measure.NP)
SCR 2. That toBe Det Adj NP
NP u9 has property.(N1 u9has measure.NP)
SCR 3. NP toBe clearly Adj
NP u9 has property.(N1 u9has measure.NP)
SCR 4. N1 of NP is clear
NP u9 has property.(N1 u9has measure.NP)
SCR 5. The Adj N’ of NP
NP u9 has property.(N’ u9 has value.Adj
u9has measure.NP )
SCR 6. NP1 toBe Adj as a N
NP1 u N u9has property.(N’ u9 value.Adj u9
has measure.N)
SCR 7. NP1 toBe NP2[+measure] Adj
NP1 u9has property.(N’ u9 value.Adj u9
has measure.NP2)
SCR 8. NP1 toBe NP2[+measure] Adj N
NP1 u N u9has property.(N’ u9has value.Adj u9
has measure.NP2)
Given the above axioms, the following exam-
ples can be handled:
(7) (a) John is a 1.50 meter tall man.
j= (b) John is 1.50 meter tall.
7a  John u Man u9has property.(height A
uhas value.tall uhas measure(1.50 meter) )
(by SCR 8)
7b j= John u9has property.(height uhas value.tall B
uhas measure(1.50 meter) )
(by SCR 7 and from A)
A j= B C
(8) (a) John is a 1.50 meter tall man. 6j= (b) John
is a tall man.
8a  John u Man u9has property.(height A
uhas value.tall uhas measure(1.50 meter) )
(by SCR 8)
8b j= John u Man u9has property.(height u B
has value.tall uhas measure(man) )
(by SCR1 and from A)
A 6j= B C
4 Implementation
For each of the 15 classes, we have specified a set
of axioms schemas, some basic semantic construc-
tion rules and a set of inference patterns which
could be deduced to follow from both of these.
The axioms schemas were implemented in De-
scription Logic using RACER and for each infer-
ence pattern identified, the corresponding Descrip-
tion Logic query was checked to verify that the
proposed axioms and semantic construction rules
did indeed correctly predict the deduced inference
patterns.
5 Further work and evaluation
The main contribution of this work is a detailed
analysis of the interactions between derivational
morphology, lexical and compositional semantics
and of their impact on the entailment patterns li-
censed by sentences containing adjective or their
related nouns/verbs.
To turn this analysis into a computational sys-
tem, its components need to be integrated into a
semantic analyser and the behaviour of that anal-
yser tested against a collection of data. We are
currently working on developing such an anal-
yser within a symbolic grammar framework. We
have also started to develop an evaluation test
suite geared towards entailment recognition be-
tween sentence pairs containing adjectives. At the
moment, the test suite contains about 1 000 infer-
ence pairs. Each item in the TestSuite (see fig. 2)
is annotated with a judgement about the truth of
the entailment between the pair of sentences, with
the type of inference involved and with the speci-
fication of adjective involved. Moreover, each ad-
jective is annotated with the WordNet sense corre-
sponding to the given class.
The idea behind this test suite is similar to
that underlying the creation of the TSNLP (Test
suite for natural language processing) (see (Oepen
and Netter, 1995)) or the Eurotra testsuites (see
(Arnold and des Tombe, 1987)) namely, to pro-
vide a benchmark against which to evaluate and
compare existing semantic analyzers. Thus this
26 KRAQ06
<pair id="1" value="TRUE" class="[CLASS1]" inference="Adj/Verb">
<t>The boat is <sn n="1"> afloat </sn>.</t>
<h>The boat is floating.</h>
</pair>
<pair id="2" value="FALSE" class="[CLASS6]" inference="Antonymy">
<t>This is not a <sn n="1"> rectangular </sn> table.</t>
<h>This is a <sn n="1"> round </sn> table </h>
</pair>
<pair id="3" value="TRUE" class="[CLASS8]" inference="Adj/Noun">
<t>The line is 2 meter <sn n="1"> long </sn>.</t>
<h>The length of the line is 2 meter.</h>
</pair>
<pair id="4" value="FALSE" class "[subs/intersective]" inference="Attr/Pred">
<t>The treasurer is <sn n="2"> present </sn>.</t>
<h>This is the <sn n="1"> present </sn> treasurer.</h>
</pair>
Figure 2: TestSuite
test suite illustrates the semantic and syntactic be-
haviour of adjectives and their related verbs/nouns
with respect to textual entailment. One could
imagine other test suites illustrating the seman-
tic behaviour of verbs, of quantifiers, of discourse
connectives, etc. Just as the TSNLP still proves
useful in supporting the development of new sym-
bolic parsers/grammars, hand built test suites of
artificial examples might prove useful in improv-
ing the accuracy of semantic analyser wrt textual
entailment. Indeed the Pascal RTE challenge has
shown that existing systems fares rather poortly at
the textual entailment task. Providing a set of hand
crafted semantic test suites might help in remedy-
ing this shortcoming.
Beside implementing and evaluating the anal-
ysis of adjectives presented in this paper, we are
also working on refining this analysis by combin-
ing it with a detailed analysis of noun semantics so
as to handle (non) entailments such as:
(9)
Lyon is the gastronomical capital of France
6j= Lyon is the capital of France

References
D.J. Arnold and Luis des Tombe. 1987. Basic Theory
and methodology in Eurotra. Cambridge University
Press.
DA. Cruse. 1986. Lexical Semantics. Cambridge Uni-
versity Press.
Claire Gardent and Evelyne Jacquey. 2003. Lexical
reasoning. In Proceedings of the ICON’03 (Inter-
national Conference on Natural Language Process-
ing), Mysore, India.
Oren Glickman Ido Dagan and Bernardo Magnini.
2005. The PASCAL Recognising Textual Entailment
Challenge.
Hans Kamp and Barbara Partee. 1995. Prototype the-
ory and compositionality. Cognition, (57):129–191.
Hans Kamp. 1975. Two theories about adjectives. In
Edward L. Keenan (ed.), Formal Semantics of Nat-
ural Language, pages 123–155. Cambridge Univer-
sity Press.
Christofer Kennedy. 2005. Vagueness and grammar:
The semantics of relative and absolute gradable ad-
jectives. Ms., pages 129–191, June.
K. J. Miller. 1998. Modifiers in wordnet. In
C. Fellbaum (ed.), WordNet An Electronic Lexical
Database. Cambridge, MA, The MIT Press.
Stephan Oepen and Klaus Netter. 1995. TSNLP -
test suites for natural language processing. Gronin-
gen, The Netherlands. Conference on Linguistic
Databases.
I. Peters and W. Peters. 2000. The Treatment of Adjec-
tives in SIMPLE: Theoretical Observations. Athens.
R. Quirk, S. Greenbaum, G. Leech, and J. Svartvik.
1985. A Comprehensive Grammar of the English
Language. Longman.
V. Raskin and S. Nirenburg. 1995. Lexical Semantics
of Adjectives, a micro-theory of adjectival meaning.
MCCS Report.
Ralf M¨oller Volker Haarslev. 2001. Description of the
racer system and its applications. In Proceedings
International Workshop on Description Logics (DL-
2001, Stanford, USA.
