I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
Augmenting WordNet-like lexical resources with distributional evidence. 
An application-oriented perspective" 
Simonetta Montemagni, Vito Pirrelli 
Istituto di Linguistica Computazionale, CNR 
Via della Faggiola 32, Pisa, ITALY 
e-mail: {simo,vito} @ilc.pi.cnr.it 
Abstract 
The paper deals with the issue of how and to what extent 
WordNet-like resources provide the necessary information 
for an assessment of semantic similarity which is useful for 
practical applications. The general point is made that 
taxonomical information should be complemented with 
distributional evidence. The claim is substantiated through 
experimental a~t8 and an illustration of a word sense 
disambiguation system (SENSE) capable of using 
contextually-relevant semantic similarity. 
1. Introduction 
Assessment of semantic similarity has proved to be be 
essential for a variety of Natural Language Processing 
(NLP) tasks, including syntactic disambiguation (either 
structural or functional), word sense disambiguation, 
selection of appropriate translation equivalent, 
assessment of lexical cohesion in texts for automatic 
summarisation, query expansion and document 
indexing in Information Retrieval. 
Typically, the semantic similarity between words 
is computed on the basis of taxonomical relationships 
such as hyperonymy. Given two word senses W a and 
W:, their similarity is captured as a function of their 
belonging to more general semantic classes. The 
approach presupposes prior availability of independent 
hierarchically-structured repositories of lexico- 
semantic information such as WordNet. An interesting 
issue here is to evaluate how nsefi~l this type of 
resource is in capm.,-mg semantic similarity at the 
desirable level of granularity, given the requirements of 
the abovelisted applications. 
As a general comment, the taxonomical approach 
to semantic similarity tends to neglect the role of 
linguistic context as a perspectivising factor affecting 
the perception of a semantic similarity between any two 
words considered. There is substantial experimental 
evidence supporting the view that human similarity 
judgements are affected by the pressure of contextual 
factors (see, among others, Goldstone et al. 1997): 
intuitively, while candle and barbecue would score 
poorly on semantic proximity if considered 
independently of their use in context, their occurence in 
expressions such as light a candle, light a barbecue 
would immediately throw in relief a (possibly weak, 
but nonetheless contextually relevant) semantic 
association between the two, established by their 
connection with the process of burning. This 
• The work reported in this paper was joindy carried 
out by the authors within the SPARKLE project (LE- 
2111). For the specific concerns of the Italian Academy 
only, S. Montemagni is responsible for sections I, 2, 
3.2, 3.3, and V. Pirrelli for 3.1, 4 and 5. 
association is relevant insofar as it plays a role in 
carving out the set of plausible objects of the verb light. 
Taxonomies are not in principle incapable of capturing 
cross-classifications like those based on relational or 
role properties such as "being a product" or "being a 
typical object of event/process". There are allowances 
in the latest Wordnet version (1.6) for defining pointers 
from each concept to = say = nouns representing its 
parts, or from nouns to verbs to represent functions 
etc., although the latter are not actually implemented 
yet. Nonetheless, it is not obvious how many of these 
cross-classificatory dimensions should be overlaid on a 
taxonomy to attain the desirable level of context= 
sensitivity required by real applications. From an 
application-oriented perspective, there is the further 
problem of how it is possible to regiment their role and 
relevance as a function of context variation. 
As a somewhat radical alternative to taxonomical 
relationships, other ways of measuring semantic 
similarity based on distributional evidence have been 
put forward in the literature (see, among others, Brown 
et al. 1991, Gale et al. 1992, Pereira and Tishby 1992), 
which emphasise the role played by context in this 
game. These approaches compute the semantic 
similarity between W z and W, on the basis of the extent 
to which W,/W,'s average contexts of use overlap. 
Here, the context is generally defined as an n-word 
window centred on Wl/W:. The method rests on the 
assumption that words entering into the same 
syntagmatic relation with other words are perceived as 
semantically similar. The method has a potential for 
capturing word similarities grounded on contextual 
effects of the sort sketched out above, although it may 
often happen that, given two instances of the same 
word W in a text and their corresponding context 
windows, very few token words are found in both 
windows. Strategies to alleviate this sparse data 
problem have been described for word sense 
disambiguation (e.g. Schiitze 1992): they def'me the 
context no longer in terms of the immediate 
neighbouring words, but rather as the set of words that 
neighbourmg words normally consort with. An 
interesting issue here is whether "context cascades" of 
this sort are still constrained enough to be able to 
capture effects of context=sensitive similarity. The 
amount of data that this method requires is also an 
issue. 
Be that as it may, it is still to be shown 
conclusively that any of the NLP tasks listed at the 
outset really requires such a f'me grained measure of 
context-sensitive semantic similarity. In this paper, we 
tend that an ideal lexical resource aimed at being 
as a yardstick for measuring word sense similarity 
I 
I 
i 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
at the level of granularity required by most NLP 
applications should strive to complement the lexico- 
semantic knowledge typically embedded in a WordNet- 
like resource with distributional evidence of some kind. 
This is argued on grounds that: i) contextual factors 
play an important role in assessing the semantic 
similarity between words and ii) this is what most 
applications require. Both points will be dealt with in 
some detail in the context of the problem of classifying 
the typical complements lexically selected by a given 
verb sense. We will show that verbs' selectional 
preferences cannot always be neatly expressed in terms 
of taxonomy nodes/classes, but rather cut across the 
taxonomy in a seemingly erratic way, straggling for 
several relatively unrelated nodes. Close examination 
of real data shows that different verb senses select 
different classes of complements according to different 
dimensions of semantic similarity, to such an extent 
that it soon becomes impossible to provide an effective 
account of these dimensions independently of the verb 
sense in question. 
2. Taxonomy-based semantic similarity: general 
background 
Different methods have been put forward in the 
literature to assess semantic similarity in relation to a 
hierarchically structured lexical resource such as 
WordNet. In most of them (see among others Rada et 
al. 1989 and Lee et al. 1993), assessment of semantic 
similarity is carried out on the basis of hyperonymy 
(IS-A) links. More concretely, semantic similarity is 
evaluated by measuring the distance between the 
taxonomical nodes corresponding to the items being 
compared: the shorter the path from one node to 
another, the more similar the corresponding items. 
Given multiple paths, the shortest path is taken as the 
one involving the stronger similarity. 
A number of criticisms have been levelled at this 
approach. Some scholars pointed out that IS-A links are 
simply not sufficient. Nagao (1992), for instance, uses 
both hyperonymy and synonymy links to compute 
semantic similarity, and assigns higher similarity 
scores to synonymy relationships. Other scholars have 
attempted to further widen the range of relationships on 
the basis of which semantic similarity is computed; see, 
among others, Niremburg et al. (1993) who also use 
morphological information and antonyms. 
A more technical problem faced by the path- 
length similarity method has to do with the underlying 
assumption that links in a taxonomy represent uniform 
distances between nodes. As often pointed out, this is 
not always the case: in real taxonomies, the "distance" 
covered by individual taxonomic links is variable, since 
certain sub-taxonomies can be much denser than 
others. To overcome the problem of varying link 
distances, Agirre and Rigau (1996) propose a semantic 
similarity measure (referred to as "conceptual density") 
which is sensitive to i) the length of the path, ii) the 
depth of the nodes in the hierarchy (deeper nodes are 
ranked closer) and iii) the density of nodes in the sub- 
hierarchies (concepts involved in a denser subhierarchy 
are ranked closer than those in a more sparse region). 
In a similar vein, Resnik (1995) defines a taxonomic 
similarity measure which dispenses with the path 
length approach and is based on the notion of 
information content- Under his view, semantic 
similarity between two words is represented by the - 
log P(C) value of the most informative concept C 
subsuming both words in a semantic taxonomy, where 
P(C) is a maximum likelyhood estimate of C's 
probability of occurrence in a reference corpus. 
Despite their differences, all these methods 
address the issue of how lexico-semantic hierarchies 
like WordNet should best be exploited, but do not 
question their suitability for measuring word semantic 
proximity. This issue will he dealt with in some detail 
in the following section. 
3. Taxonomy-based semantic similarity at work: an 
illustrative example 
In this section, the problem is tackled of how and to 
what extent a WordNet-like lexical resource can 
provide the information needed to assess semantic 
similarity of words in context, in connection with the 
task of of semantically characterising the class of 
typical collocates of a given verb sense. In section 3. I a 
taxonomy-based account of selectional preferences of 
different senses of the same verb is illustrated. This is 
complemented with a comparative study of intersecting 
sets of typical collocates of different verb senses 
(section 3.2). 
3.1 A taxonomy-based account of selectional 
preferences of verbs 
This section illustrates the modelling of the selectional 
preferences of different senses of a verb according to a 
taxonomy-based view. To exemplify, we consider here 
the different senses of the Italian verb accendere 
together with the sets of their typical object collocates. 
These typical objects are projected onto a semantic 
hierarchy to evaluate whether and to what extent the 
verb's selectional preferences are captured through 
taxonomical generalizations of some kind. 
According to the Collins Italian-English 
dictionary (1985), the Italian verb accendere has, in its 
transitive reading, the following four senses, each 
accompanied by an illustrative set of its typical objects: 
1) light when it takes as a direct object nouns like fiammifero, 
candela, sigaretta, caraino (respectively, 'match, candle, 
cigarette, fireplace') 
2) mrn on. switch on, when the object is a device such as radio, 
luce, lampada, gas, motore (respectively, "radio. light, lamp, gas 
cooker, engine') 
88 
I 
I 
1 
I 
I 
i 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
3) raise, if the object is some kind of feeling such as speranza, 
desiderio ('hope, desire') 
4) open, if the object is a bank-related entity such as conto, debito, 
ipoteca 'bank account, debt, mortgage' 
Given the source of lexical information considered 
here, each sense is characterised in terms of its 
appropriate English translation equivalent. The tree- 
like structure reported below illustrates the result of 
projecting Collins' typical object collocates of each set 
onto WordNet. 
I '| 
Figure I Objects of accendere: semantic hierarchy 
Shadowed boxes (typically but not necessarily tree 
leaves) represent the actually occurring collocates, 
which are accompanied by an indication of the sense of 
accendere with which they are associated in the 
dictionary. Dotted lines in the tree show that the link 
between the connected nodes is not direct, i.e. that the 
taxonomical path includes intermediate nodes. 
The fast thing to note in this context is that 
collocates of different senses exhibit a different 
propensity to cluster together in the semantic hierarchy. 
The selectional preferences of senses 3 ('raise') and 4 
('open') nicely fall into distinct branches of the 
taxonomy. The class of typical objects of sense 3 can 
appropriately be described as a <feeling>, while 
<possession> being a suitable hyperonym of all and 
only objects of sense 4. Yet, the same taxonomy fails to 
part the selectional preferences of sense 1 ('light') from 
those of sense 2 ('switch on'). In the latter case, object 
collocates of both senses are categorised as an 
<artifact>, a notion which is far too general to tell the 
collocates of sense 1 of accendere from those of its 
sense 2, as illustrated by the internal structure of the 
sub-hierarchy of artifact objects of accendere 
diagrammed in Figure 2 below. For senses 1 and 2 of 
accendere, the clustering of nodes in the subtaxonomy 
of 'artifacts' does not help to identify the semantic 
"glue" that keeps together the object collocates for each 
relevant sense. 
Although we are working on Italian examples, we will 
use hereafter, for illustrative purposes, WordNet 1.5 as 
a reference taxonomical resource, due to its 
completeness, the Italian WordNet being still under 
development in the framework of the EuroWordNet 
project (LE2-4003). This decision is not arbitrary, 
since, for the words considered in this paper, the Italian 
WordNet shows a similar taxonomical organization as 
Wor&~qet 1.5. 
• 
| ,. N 
Figure 2 Hierarchy of 'artifact' objects of accendere 
Sense 1 of accendere selects for artifact objects which 
can bum; sense 2 basically selects for devices which 
are activated through making electric contact. 
The problem here is not simply that it is 
impossible to identify one single upper node covering - 
say - all and only burning artifacts as opposed to 
devices making electric contact. The classical 
assumption that one scrnantic class should be made to 
contain all and only the collocates of one sense is 
clearly too strong in this context, if it is a workable 
cognitive hypothesis at all. One could nonetheless fall 
back to the weaker assumption that a class of collocates 
be expressed in terms of a disjunction of the 
taxonomy's nodes/subclasses, provided that each such 
node/subclass defines a proper subset of the typical 
objects of the verb sense in question. In fact, our 
diagram above shows that even this weaker 
characterization is not viable in all cases. Consider the 
taxonomy chain formed by luce-larapada-candela 
corresponding to the class of objects having to do with 
<light, source of illumination>. Whereas luce 'light' 
and its hyponym larapada 'lamp' both point to the 
'switch on' sense (sense 2), candela 'candle' (the 
terminal node of this chain) is associated with the 
'light' sense (sense 1), due to its being a typically 
burning object. Here the same taxonomy chain includes 
objects related to different senses of the verb. This is 
tantamount to saying that the dimension of semantic 
similarity captured through the taxonomical structure is 
not appropriate but rather misleading if one wants to 
unambiguously characterise the different senses of the 
verb through thei~ selectional preferences. The property 
of burning, on which the preference is based, cannot 
possibly be percolated from higher to lower nodes 
through the taxonomy chain. Rather, it represents a 
property peculiar of some nodes only, either 
intermediate or terminal ones. Hence, given the 
taxonomy illustrated above, one can do little more than 
disjunctively listing all nodes corresponding to the 
collocates in question, with the further stipulation that 
the property of being a collocate does not necessarily 
percolate further down in the taxonomy chain. This is 
fine, but it boils down to saying that the taxonomy in 
question can do very little to generalize over the 
selectional preference classes. 
89 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
,II 
I 
To sum up, this simple example shows that 
taxonomy-based semantic similarity is not always 
sufficient to justify the belonging of a given lexical 
item to a specific selectional preference class. Very 
granular distinctions may be needed to characterise any 
such class. Moreover, some of the distinctions required 
are orthogonal to the distinctions conveyed by a 
taxonomical organisation of the lexicon. 
3.2 Comparing overlapping seleetional preferences 
of different verbs 
So far, we focussed on the difficulty of neatly 
characterising verb selectional preferences in terms of 
taxonomical classes. It turns out that the semantic glue 
pasting together the object collocates of senses 1 and 2 
of the verb accendere is given by distinctions which are 
not directly reflected in the semantic taxonomy. 
Taxonomical relationships seem to capture only some 
of the various dimensions on which semantic similarity 
is grounded. This is not accidental, we believe, since 
taxonomical dimensions are typically def'med i) 
independently of context, and ii) once and for all. It is 
thus not surprising that they may fail, in some cases, to 
reflect the similarity dimension appropriate in a 
specific context. In this section, this issue is explored in 
more detail by comparing the selectional preferences of 
different verbs exhibiting a non empty intersection of 
the sets of their typical collocates. 
Among the typical collocates of sense 1 of 
accendere 'light' there is sigaretta 'cigarette' which, in 
WordNet 1.5, is the terminal node of the following 
taxonomical path: 
sigarctla 'cigarette' 
=> roll of tobacco 
=> tobacco, baccy 
=> narcotic 
=> drug 
=> artifact, artefact 
=> object, inanimate object, physical object 
=> entity 
Let us look now at some of the typical verbs with 
which sigaretta occurs, together with other possible 
collocates of these verbs, as they are attested in the 
Collins Italian-English Dictionary (1985), in both 
example sentences and the semantic indicators field. In 
these examples, the sequences "/S" (short for 
"Subject") and "/O" (short for "Object") specify the 
grammatical relation of the noun relative to the verb: 
• ACCENDERF..$0_I/V {SIGARETTA./O CAMINO/O 
CANDELMO FIAMMIFERO/O} 
light {cigarette/O, fireplace/O, candle/O, match/O} 
• ARROTOLARE$O l/V {SIGARETTPdO CARTAJO 
STOFFA/O} 
roll up {cigarette/O paper/O fabric/O} 
• FUMARES0_I/V {SIGARETTMO PIPA/O} 
smoke {cigarette/O pipe/OI 
• OFFRIRES0_I/V {SIGARETrA/O AIUTO/O LAVORO/O 
M ERCE/O PREGHIERAiO} 
offer {cigarette/O help/O job/O goods/O prayer/O } 
• RIACCENDERE$0 I/V {SIGARETTA/O FUOCO/O GAS/O 
\[NTERESSE/O LUCE/O RADIO/O SENTIMENTO/O} 
light/switch on/revive {cigarerte/O fire/O gas/O interest/O 
light/O radio/O feeling/O} 
• SPEGNERE$0 I/V {SIGARETTA/O APPARECCHIO/O 
DEBITO/O l~tJOCO/O GAS/O LUCE/O PASSIONE/O 
S UO NO/O } 
extinguisWswitch off7stifle/rnufl\]e {cigarette/O device/O debt/O 
fire/O gas/O light/O passion/O sound/O} 
• SPEGNERSI$O 2./V {SIGARETTA/S APPARECCHIO/S 
FUOCO/S LUC-E/S PASSIONE/S RICORIX)/S SUONO/S} 
be extinguished/stop/fade away {cigarette/S deviceJS fire/S 
light/S passion/S memory/S sound/S} 
Careful consideration of these examples shows that 
different types of semantic glue are at work in different 
eases. With the verb accendere (sense 1) the glue is, as 
we saw, the property of burning. A similar analogy is at 
work in the case of riaccendere, spegnere and 
spegnersi, with the main difference that this case also 
includes figurative usages. As to the verb arrotolare, 
the semantic similarity of its object collocates is 
grounded on their being made of material whose 
texture makes them rollable. The relevant similarity 
which links pipes and cigarettes relative to the context 
offumare rather hinges on their telic role, their both 
being typically smoked objects. Finally, the 
collocational set of offrire includes words denoting 
typical human needs and/or desires ranging from 
cigarettes and goods to more abstract things such as 
help and prayers. 
These examples confirm the difficulty of 
assessing the semantic similarity of words when they 
are considered outside their actual contexts of use, 
difficulty which already emerged in relation to a 
characterization of the selectional preferences of the 
verb accendere. By projecting these collocational sets 
onto WordNet, appropriate generalisations can hardly 
be found. A general semantic class subsuming some or 
all members of each set may exist, but often it is not 
specific enough to avoid undesired intersection of 
classes, as in the case of senses 1 and 2 of accendere. 
On the other hand, semantic features such as 
"lightability", "enjoyability", "smokability" or 
"rollability" seem to be at work here: they strike us as 
hardly amenable to a global consistent taxonomical 
rendering. 
3.3 Implications 
In the previous sections, we discussed whether and to 
what extent taxonomical relationships as actually 
implemented in WordNet-like lexical resources can be 
used to measure the semantic similarity of typical 
collocates associated with a given verb sense. We 
showed that one can hardly fred a unique taxonomy 
node subsuming all and only the collocates bearing the 
same grammatical relation to a given verb sense. A 
weaker but more realistic hypothesis was also 
considered, namely that a class of verb collocates be 
g0 
I 
l 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
expressed in terms of a disjunction of taxonomy's 
nodes, provided that each such node defines a proper 
subset of the typical collocates of the verb sense in 
question. It turned out that even this weaker 
characterization of selectional preferences is not always 
viable since it is often the case that selectional 
preference information is not disjunctively distributed 
over taxonomy nodes. When this is the case, a 
taxonomy provides virtually no means of generalising 
over the set of typical collocates of a given verb sense. 
In our view of things, such an inadequacy of 
taxonomical information cannot be got around by 
letting finer grained distinctions slip in the semantic 
type model. Rather, it bears upon one inherent property 
of most taxonomies as they arc currently built up: 
monodimensionality. In fact, taxonomies are often 
anchored to a fixed classificatory dimension (e.g. 
perceptual features as opposed to functional ones). By 
contrast, real data suggest that different verb senses 
select different classes of complements according to 
different dimensions of semantic similarity. This is the 
reason why taxonomies do not always capture locally 
salient common features, which are needed to 
appropriately account for the semantic similarity of 
verb complements. 
Our examples showed that multidimensional 
classifications are indeed required to dynamically 
capture locally salient features. Although in WorNet 
1.6 provision is made for concepts to be cross- 
classified with respect to different dimensions, it is not 
clear how many and what dimensions should be added 
to the original WordNet structure to comply with real 
NLP application requirements. These considerations 
are, in our view, compelling enough to prompt the 
investigation of different and more workable ways to 
complement the taxonomical structure of WordNet-like 
resources. A simple but effective source of knowledge 
which can nicely complement WordNet for capturing 
locally salient semantic similarity is represented by 
distributional information about words, under the 
assumption that words which bear the same syntactic 
relation to the same word sense form a somehow 
semantically coherent class. 
In the following section, we illustrate this point 
by describing a measure of semantic similarity based 
on distributional evidence and we show how helpful 
this is in capturing locally salient semantic similarity. 
4. Distributionally-based semantic similarity 
A semantic similarity measure computed on the basis 
of distributional evidence is at work in SENSE, an 
example-based word sense disambiguation (WSD) 
system carrying out the task on the basis of a 
representative set of typical patterns of use (Federici et 
al. 1997). In particular, SENSE presupposes prior 
availability of verb-noun pairs where the contextually 
relevant sense of the verb token is assigned. At the 
same time, the accompanying noun is provided with its 
grammatical function. This set of verb-noun pairs 
constitutes the knowledge base of examples (or 
example base for short) on the basis of which SENSE is 
able to draw its inferences. 
Given an Input Pair IP to be disambiguated 
where the grammatical relation of the noun relative to 
the verb is specified, SENSE searches its example base 
looking for the set of examples which are most similar 
to IP. If an identical pair is found in the example base, 
then the usual assumption is made that the verb token 
in IP is used in the same reading of the verb in the 
known example) The key notion used by SENSE to 
compute similarity between non identical pairs is 
proportional analogy. To illustrate, if the verb sense in 
the pair accendere-pipa/O 'light-pipe' has to be 
inferred, this can be done through the following 
proportion, involving three disambiguated verb-object 
pairs attested in the example base plus the input pair 
accendere.pipa/O as the fourth term: 
fumare l- : ~umare l- = accendere_l- : accendere~- 
sigarettatO pipalO sigaretta/O pipalO 
'smoke- : 'smoke- = "light- : 'light- 
ci~arette/O' pipedO' ci~arettc/O' pile-dO' _ 
Intuitively, the proportion says that the sense of 
accendere in accendere-pipa "light-pipe' is likely to be 
the same as in accendere-sigaretta 'light-cigarette' 
since both pipa and sigaretta can typically be smoked, 
or - in more linguistic terms - since they ate both 
typical objects of sense 1 of fumare 'smoke' 
(fumare l). 
It is important to point out here that this 
inferential strategy is "local" in two senses: i) relative 
to the example base, and ii) relative to the input pair. 
First, it neither presupposes nor relies on a preliminary 
classification of all known examples. In this respect, 
the system simply memorizes all examples, with no 
attempt to generalize over them in any optimal global 
way. Generalizations are only made to interpret new 
unknown evidence. Hence, the resulting classification 
does not reflect general properties of the example base 
as such, but only associations which are triggered by 
the specific input pair in question. In this sense the 
hypothesis search space is constructed on the fly, every 
time the system is confronted with a new unknown 
pair. 
The second notion of "locality" we intend to 
emphasize here is related to the issue of what 
constitutes a relevant analogy, given the input pair IP 
considered. The similarity between an IP and some 
known examples is not simply based on a a-priori 
2 In fact, SENSE is also able to go beyond the evidence 
provided by an attested example as illustrated in 
Federici et al. 1997. 
91 
I 
I 
I 
I 
I 
I 
I 
I 
! 
! 
li 
I 
global similarity of some. of its constituent elements 
(i.e. the verb and the noun), which, as we just saw, is 
not available. An analogical proportion enforces a 
much more constraining relation. The interpretation 
'light-pipe' of accendere-pipa is not simply based on 
the piecemeal analogy with accendere-sigaretta (where 
accendere is found in common, and pipe and cigarette 
are sufficiently similar). The conclusive element of the 
analogy is that both fumare and accendere in their 
respective senses of 'tight' and 'smoke' are 
systematically related in the example base through a set 
of shared objects, and that pipa occurs with fumare in 
the required sense. This is exactly what the proportion 
is able to capture. 
We contend that, for the notion of context- 
sensitive word sense similarity to adequately be 
modeled, both notions of locality play an important 
role. 
The example base used so far for testing the 
effectiveness of the distributionally-based semantic 
similarity measure for WSD purposes was 
automatically acquired from both semantic indicators 
and example sentences of the Collins Italian-English 
Dictionary (Montemagni 1995). Each acquired verb- 
noun pair can thus be said to represent a typical pattern 
of use of a given sense of a verb. The choice of a 
bilingual dictionary was also motivated by the practical 
interest that the resulting sense subdivisions have for 
purposes of Machine Translation. 
The derived example base contains 8,153 verb- 
noun pairs (either verb-subject or verb-object patterns) 
which exemplify 3,359 different verb senses. All pairs 
are acquired from verbentries, and thus provide sense 
information only about the verb; each accompanying 
noun, ifpolysemous, is not disambiguated. On average, 
a verb sense is illustrated through 2.42 patterns. Senses 
which are attested in ten or more patterns are a 
negligible part of the training set, whereas most verbs 
are illustrated through a number of patterns ranging 
between 2 and 5. Finally, a considerable group of verb 
senses is attested only once. Note that the latter 
circumstance does not stop SENSE from recognising 
• hapax senses in novel unknown contexts. 
SENSE performance was tested on a corpus of 
150 IPs randomly extracted from unrestricted texts. 
Since the test was intended to evaluate the reliability of 
distributionally-based inferences, the test corpus did 
not contain any pattern already present in the example 
base. Only verbs were disambiguated. The results of 
this experiment are reported in the table below: 
Overall Polysemous 
RECALL 79.3% 66.3% 
PRECISION 89,9% 80.4% 
Figures in the first column refer to both polysemous 
and monosemic verbs. In the second column, recall and 
precision are relative to polysemous verbs only. These 
figures are very significant if one considers i) the 
comparatively small size of the lexical database used 
for training, ii) the distribution of patterns per verb 
sense, and iii) the fact that only some of its attested 
words (namely verbs) are semantically disambiguated. 
The results reported above were computed on the 
basis of distributional evidence only. On closer 
analysis, it turned out that some of the input contexts 
which were lefr ambiguous by SENSE could have been 
successfully disambiguated if also taxonomical 
information was taken into account. Consider the 
following three cases: 
verb sense object 
input abbattere ? pianta 
context 'cut down' 'plant' 
known abbattere 1 albero 
example 'cut down' '~'ee' 
hyponym: 
albero 
input abbassare 
context 'hang' 
known abbassare 
example 'hanl\[' 
input accarez=are 
context 'stroke' 
known acearezzare 
example 'stroke' 
? capo synonym: 
'head' testa 
1 testa 
'head' 
9 barba hyperonym: 
'beard' pelo 'hair' 
1 capello hypcronym: 
'hair' pelo 'hair' 
In the first case, the object in the target context is the 
WordNet hyperonym of the object in the known 
example, as shown in the rightmost column of the 
table. In the second case, the objects of both input and 
known pairs are synonyms. Finally, the last case 
illustrates a typical instance of hyperonym sharing. 
This indicates that distributionally-based and 
taxonomy-based inferences can nicely be 
complemented. In practice, this can be done in more 
than one way. In some experiments of syntactic 
disambiguation (subject/object assignment in Italian, 
Montemagni 1995, Montemagni et al. 1996), we tried 
to combine both taxonomical and distributional 
measures in such a way that the system retied on 
taxonomical information first, to turn to distributional 
evidence only when the first step was not conclusive. 
This strategy, however, did not seem to be successful, 
as the system was frequently led astray by irrelevant 
similarities. Our experience seems to suggest that a 
more promising way to integrate distnbutionally-based 
and taxonomy-based information is arguably to use 
distributional evidence fast, so as to exploit the 
context-sensitivity (or locality) of proportional analogy 
as a filter of irrelevant similarities. Taxonomical 
information is to be relied on only at a second stage, as 
a fall back solution to outstanding ambiguities. 
5. Conclusions 
Semantic similarity is not simply a relation between 
two words in isolation, but rather a relation between 
two words in their context. This context-sensitive view 
of semantic sLrnJlarity makes its identification more 
problematic. In principle, semantic similarity of words 
92 
can be captured in a number of different ways, ranging 
from their taxonomical relationships to their actual 
distribution in a corpus. It would be very difficult to 
argue that one such a way is more plausible than 
another; nonetheless, it should be observed that their 
practical utility in well-known interesting NLP 
applications can vary considerably. 
We noted that taxonomy-based measures of 
semantic similarity are to an extent inadequate, as they 
capture only some of the classificatory dimensions 
which play a relevant role in NIP applications. We 
showed that relevant similarities need to be grounded 
on the specific context to be processed (e.g. 
disambiguated, retrieved or summarised) and that 
different contexts call for different classificatory 
dimensions. Distributional evidence can be used to 
model this sort of context-sensitive multidimensional 
classification, so as to induce semantic associations 
between words that nonetheless belong to different 
places in a taxonomy. We also showed that 
distributionally-based semantic similarity has a 
considerable impact on crucial NLP tasks such as word 
sense disambiguation. All this provides evidence that 
WordNet-like lexical resources should strive to 
integrate taxonomical and distributional information, 
by combining both paradigmatic and syntagmatic 
dimensions. 
As already mentioned, Word,Net has a potential 
for doing that, through extended implementation of so- 
called pointers from nouns to verbs and from verbs to 
nouns, to represent functions, typical semantic 
preferences etc. Within the EuroWordNet project (LE2- 
4003), some steps in this direction have already been 
taken in developing multilingual WordNets for Dutch, 
Italian and Spanish. Among the additions to the 
original set of relations borrowed from WordNet 1.5, 
syntagmatic relations feature prominently: e.g., one 
finds verb-to-noun relations denoting the typical 
entities involved in a given event, or noun-to-verb 
relations referring to the typical events in which a given 
entity play a role (Alonge et al. forthcoming). 
This certainly provides the information needed to 
capture context-sensitive semantic similarities. We also 
showed that local inferential engines such as SENSE 
can demonstrably tap this type of information with the 
degree of flexibility, noise-tolerance and input- 
relevance required, among others, by WSD. 

References 
Alongc A., N. Calzolari, P. Vosscn, L. Bloksma, \[. 
Castellon, A. Marti, W. Peters, forthcoming, 'The 
Linguistic Design of the EuroWordNet Database', in 
Computers and the Humanities: Special Issue on 
Euro WordNet. 
Agirre E., G. Rigau, 1996, 'Word Sense 
Disambiguation using Conceptual Density', in 
Proceedings of COLING-96, Copenhagen, pp. 16-22. 
Brown P.F., S.A.D. Pietra, V.J.D. Pietza, R.L. Mercer, 
1991, 'Word Sense Disambiguation Using Stastical 
Methods', in Proceedings of A CL-1991, pp. 264-270. 
Collins Giunti Mamocco, 1985, English-Italian Italian- 
English Dictionary, London Firenze. 
Federici S., S. Montemagni, V. Pirrelli, 1997, 
'Inferring semantic similarity from Distributional 
Evidence: an Analogy-based Approach to Word 
Sense Disambiguation', in Proceedings of the 
ACL/EACL Workshop "'Automatic Information 
Extraction and Building of Lexical Semantic 
Resources for NLP Applications'; Madrid, Spain. 
Gale W.A., K.W. Church, D. Jarowsky, 1992, 'A 
Method for Disambiguationg Word Senses in Large 
Corpora', Computers and the Humanities. 
Goldstone R.L., Medin D.L., Halberstadt J., 1997, 
'Similarity in context', Memory and Cognition, 25(2), 
pp. 237-255. 
Lee J.H., M.H. Kim, Y.I. Lee, 1993, 'Information 
Retrieval based on conceptual distance in IS-A 
hierarchies', Journal of Documentation, 49(2), June 
1993, pp. 188-207. 
Montemagni S., 1995, Subject and Object in Italian 
Sentence Processing, PhD Dissertation, UMIST, 
Manchester, UK. 
Montemagnl S., S. Federici, V. Pirrelli, 1996, 'Resolving 
syntactic ambiguities with lexico-semantic patterns: an 
analogy-based approach', in Proceedings of COLING- 
96, Copenhagen, August 1996, pp. 376-381. 
Nagao M., 1992, 'Some Rationales and Methodologies 
for Example-Based Approach', in Proceedings of 
"International Workshop on Fundamental Research 
for the Future Generation of Natural Language 
Processing", 30-31 July 1992, Manchester, pp. 82-94. 
Nirenburg S., C. Domashnev, D.I. Grannes, 1993, 
'Two Approaches to Matching in Example-Based 
Machine Translation', in Proceedings of TMI-93, pp. 
47-57. 
Pereira F., N. Tishby, 1992, 'Distributional Similarity, 
Phase Transitions and Hierarchical Clustering', 
Working Notes, Fall Symposium Series, AAAI, pp. 
108-112. 
Rada IL, M. Hafedh, E. Bicknell, M. Blettner, 1989, 
'Development and application of a metric on 
semantic nets', IEEE Transactions on System, Man, 
and Cybernetics, 19(I), pp. 17-30. 
Resnik P., 1995, 'Using Information Content to 
Evaluate Semantic Similarity in a Taxonomy', in 
Proceedings of IJCAI-95. 
Schiitze, H., 1992, 'Word Sense Disambiguation with 
Sublexical Representations', Workshop Notes, 
Statistically-based NLP Techniques, AAAI, 109-113. 
