File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/97/w97-0813_metho.xml
Size: 31,393 bytes
Last Modified: 2025-10-06 14:14:42
<?xml version="1.0" standalone="yes"?> <Paper uid="W97-0813"> <Title>Inferring Semantic Similarity from Distributional Evidence: an Analogy-based Approach to Word Sense Disambiguation*</Title> <Section position="3" start_page="0" end_page="90" type="metho"> <SectionTitle> 2 Semantic Similarity and WSD </SectionTitle> <Paragraph position="0"> Most methods proposed in the literature for establishing the semantic similarity of words try to map a given word * The work reported in tills paper was jointly earned out by the authors in the framework of the SPARKLE (Shallow PARsing and Knowledge extraction for Language Engineering) project (LE2111). For the specific concerns of the Italian Academy only, S Federiei is responsible for sections 3.2, 3.4 and 3.5, S. Montemagni for 2, 3.3 and 4, and V. Pirrelli for 1, 3.1 and 5.</Paragraph> <Paragraph position="1"> in context onto the set of known usages of that word in a dictionary database: thesaural information is used as a yardstick for measuring the semantic proximity between known patterns of use and the context to be disambiguated. Eventually, the sense supported by those patterns which are semantically closer to the context in question is selected as the most likely one (see, among others, \[Dolan, 1994\], \[Resnik, 1995a, 1995b\], \[Agirre and Rigau, 1996\], \[Sanfilippo, 1997\]).</Paragraph> <Paragraph position="2"> Suppose that one wants to disambiguate the sense of accendere in the verb-object pair accendere-televlsione 'switch on-tv'. The relevant sense of accendere can be inferred-on the basis of known examples such as accendere 2-radio 'switch on-radio': this inference is supported by any seman/i'c hierarchy where both radio and television are specified for the same hyperonym, e.g.</Paragraph> <Paragraph position="3"> 'device', whether immediate or not.</Paragraph> <Paragraph position="4"> Thesaural relationships such as hyperonymy and synonymy, however, do not always capture the dimension of similarity relevant to the context in question. Consider the verb accendere in the context accendere-pipa 'light-pipe'. The table below contains typical objects of two senses of accendere, 'light' (sense 1) and 'switch on' (sense 2) as they are attested in the Collins Italian-English Dictionary \[1985\], together with the objects' corresponding hyperonyms according to a monolingual Italian dictionary \[Garzanti, 1984\].</Paragraph> <Paragraph position="5"> The word ptpa, which Garzanti describes as a smoking tool, does not match any of the immediate hyperonyms of the typical objects of accendere 1 and accendere 2 in Foutt Onbekende schakelo~ie-instructie.. By looking further up in the semantic hierarchy, some similarities are indeed found, but they are based on too general semantic features to be of avail for discriminating among senses 1 and 2 of accendere.</Paragraph> <Paragraph position="6"> We suggest that, for accendere-pipa to be understood in the appropriate sense, namely accendere 1 as in accenderel-sigaretta 'light-cigarette', semantTc proximity need be computed on different grounds. The relevant similarity with links pipes and cigarettes in this specific context relates to their both being typically smoked objects, a fact which is orthogonal to their general semantic class and can be captured on a distributional basis: p:pa and sigaretta are distributionally equivalent relative to the same verb sense, i.e. they both occur as typical objects of the verb fumare 'smoke'. Distributional equivalence correlates with semantic similarity under the assumption that nouns which bear the same syntactic relation to the same verb sense are part-of a semantically coherent class. It turns out that, in examples such as accendere-pipa, distributionally-based semantic similarities can permit more appropriate sense assignments which are specifically tailored to the context to be disambiguated. Observe further that also similarities commonly captured on the basis of thesaural information, as in the case of radio and televtsione above, can in principle be inferred from distributional evidence through relevant contexts of use (e.g. spegnere-radlo 'switch off-radio' and spegnere-televtsione 'switch offtv' in t~ example at hand).</Paragraph> <Paragraph position="7"> Summing up, we contend that thesaural relationships capture only some of the various dimensions of word sense analogy which appear to play a relevant role in the disambiugation of word co-occurrence patterns. In fact, while thesaural relationships are def'med out of context once and for all, effective analogies are to be tailored to the specific contextual pattern to be disambiguated. We showed how this can be attained on the basis of distributional evidence.</Paragraph> </Section> <Section position="4" start_page="90" end_page="94" type="metho"> <SectionTitle> 3 SENSE: a distributionally-based WSD system </SectionTitle> <Paragraph position="0"> SENSE (Self-Expanding linguistic kNowledge-base for Sense Elicitation) is a specialised version of a general purpose language-learning system (\[Federici and Pirrelli, 1994\]; \[Federici et el., 1996a\]; \[Montemagni et al., 1996\]) for carrying out WSD on the basis of distributional evidence.</Paragraph> <Paragraph position="1"> SENSE's inferential routine requires: i) a structured data set of known word co-occurrence patterns (WCPs) constituting an Example Base (EB); ii) a target context to be disambiguated (TC); iii)a best-analogue(s) function (BAF) projecting TC onto EB for the best analogue(s) to be selected and thus the most likely senses to be identified*</Paragraph> <Section position="1" start_page="90" end_page="92" type="sub_section"> <SectionTitle> 3.1 Internal architecture of EB Word co-occurrence patterns </SectionTitle> <Paragraph position="0"> WCPs are modelled here as consisting of an input and an output level of representation. At the input level, each element of the pattern is described by a set of features which are expected to be of some use for WSD: lemma, part of speech and morpho-syntactic properties (such as the syntactic function of nouns with respect to the verb)* The output representation simply consists in the expected answer, i.e. the sense of each element of the pattern in the described context. An example of this type of linguistic object, illustrating the pattern fumare 1s:garetta_l/O 'smoke-sigarette', is given in Four! (On null The input representation is a list of sets of atomic units; each feature set (which is assigned a single column in the table) describes a distinct element of the pattern. In output, the list of atomic units &quot;fumare 1&quot; and &quot;sigaretta 1&quot; indicates the senses of the elements in the specific context* Elements in the input and output lists are conventionally ordered.</Paragraph> <Paragraph position="1"> In the current version of Italian EB used for our purposes, WCPs are verb-noun pairs where the relation of the noun to the verb is either subject or object. This presupposes a preliminary stage of morpho-syntactic parsing \[Montemagni, 1995\]: co-occurrence patterns abstract away from actual word forms and are augmented with information about grammatical relations.</Paragraph> <Paragraph position="2"> Note that although availability of pre-processed input makes word sense disambiguation simpler and more accurate, it is in no way a necessary precondition for the task to be carried out.</Paragraph> <Paragraph position="3"> Pairwise analogies The Italian EB consists of WCPs of the type illustrated in Fout! Onbekende schakeloptie-instruetie, above.</Paragraph> <Paragraph position="4"> Note however that they are not used as such; rather they form part of a distributed network ~ which is constructed so as to i) factor out the optimal set of analogies shared by all WCPs in EB, and ii) link the found analogies with their corresponding complements relative to the full WCPs (so-called differing parts). To make this picture more concrete, let us consider some simple examples.</Paragraph> <Paragraph position="5"> Given a pair of word co-occurrence patterns wcp, and wcp2, they are judged to be analogous if they share some representation units at both input and output levels. Any shared collection of units of both levels is referred to as an analogical core (or simply core, written wcplnwcp2 ). Suppose that wcp, and wcp2 are fumare 1-slgaretta_l/O and fumare_l-pzpa_l/O 'smoke-pipe' respectively, defined as in Fout! Onbekende sehakeloptie-instruetie.</Paragraph> <Paragraph position="6"> Their core is identified by a function (MF) mapping one set of units in Fout! Onbekende sehakeloptieinstruetie, onto one set of units in Fout! Onbekende sehakeloptie-instruetie, through the identity relation.</Paragraph> <Paragraph position="7"> MF is order-sensitive, so that only sets which take the same relative order in the lists are mapped onto each other. Fout! Onbekende sehakeloptie-instruetie, gives a possible result of this operation in the leftmost box headed by wcptnwcp2. The core in question is a verb-noun pair where the noun element is specified only at the input level, for a subset of the features describing the noun elements in the compared patterns, while nothing being said as to the possible sense interpretation of the noun. Nonetheless, the information about the noun conveyed by the core, namely its syntactic relation to the verb, is part of the knowledge supporting the interpretation of the verb asfumare_l: i.e. the verb in this reading built from scratch every time a new TC is presented to the system However, for the sake of clarity, in what follows we Illustrate the working of our system as though the network structure were built during the acquisition of EB. See \[Federiei et al, 1996b, p.393\] for a discussion of the two alternatives.</Paragraph> <Paragraph position="8"> output l fumare I I l lsig retta_ll.l l The complements of the core relative to wcp~ and wcp2 designate those units which are specific to the compared objects: they constitute the so-called differing parts, illustrated in Four! Onbekeude sehakeloptie-instruetie.</Paragraph> <Paragraph position="9"> in the columns headed by wcp,-wcp2 and wcp2-wcp, respectively. They contain information about the lexical fillers of the noun slots of the patterns.</Paragraph> <Paragraph position="10"> Network structure of EB Cores and remaining parts are always anchored to a given pair of linguistic objects: in fact, cores cannot be extracted either from existing cores or from existing differing pans. When more than one pair of WPCs is considered, it may turn out that what is a core relative to a given pair is a remaining part relative to another pair. Suppose that MF maps fumare_l-stgaretta_l/O (wept) onto accendere_l-szgarettal/O (wep3). One of the possible results of this mapping is shown in Fout! Onbekende sehakeloptie-instructie, below: fum~e 1 &quot;'&quot; slgaretta 1 Comparison of cores and remaining parts in Fout! Onbekende schakeloptie-instructie, and Font! Onbekende sehakeloptie-instruetie, above shows that one of the remaining parts relative to wcp~ and wcp2 (namely wcpl-wcp2) is identical to the core relative to wcp, and wcp3 (wcplnwcp3).</Paragraph> <Paragraph position="11"> The informational content of Fout! Onbekende sehakeloptie-instructie, and Foat! Onbekende sehakeloptie-instructie, can be represented conveniently through the graph in Fout{ Onbekende schakeloptie-instruetie..</Paragraph> <Paragraph position="13"> The graph represents cores and remaining parts as connected nodes, each accompanied by a mnemonic label.</Paragraph> <Paragraph position="14"> For example s:garetta_l corresponds to wcp,-wcp2 = wcp,n wcp3. An (unoriented) arc connecting two nodes expresses their &quot;complementarity&quot;, i.e. the intuitive notion that the two connected nodes, taken together, cover an attested WCP in its entirety. For instance, fumare 1-0bject is connected with sigaretta 1 since they form together an attested WCP, names fumare 1stgaretta_I/O. By contrast, no direct connection is ~b- null served between accendere_l-object and plpa_l, to signify that no corresponding pattern is attested in EB. Remaining parts which are connected with the same core are said to be contrastive, since, by replacing one with the other, different WCPs are obtained. A graph like the one in Foutt Onbekende sehakeloptie-instructie, represents in our terms an analogical family (AF). Clearly, far more extended AFs than the one in Fout! Onbekende sehakeloptie-instructie, can be found.</Paragraph> <Paragraph position="15"> Among the WPCs of the AF in Fout! Onbekende schakeloptie-instructie., fumare_l-sigaretta_l/O is the only one which is made up out of two cores, namely wcpln wcp3 and wcpln wcp3. Due to its pivotal position in the graph, it is some times useful to refer to it as the &quot;hook pattern&quot;, or more simply &quot;hook&quot;, of the AF in question. Accordingly, we will call the noun collocate of a hook, i.e. stgaretta_l in the example at hand, &quot;hook noun&quot;, and the corresponding verb, i.e. fumare_l, &quot;hook verb&quot;. Note further that the hook noun stgaretta_l is functional to establishing a kinship between the verb senses fumare_l and accendere_l, since it denotes a non-empty intersection between typical patterns of their use.</Paragraph> </Section> <Section position="2" start_page="92" end_page="92" type="sub_section"> <SectionTitle> 3.2 The Best-analogue(s) Function </SectionTitle> <Paragraph position="0"> Unlike linguistic objects in EB, which are specified for two representation levels (input and output), a Target Context (TC) is specified at the input level only, since its sense is precisely what the system has to predict on the basis of the available knowledge.</Paragraph> <Paragraph position="1"> This prediction is carried out through operation of the best-analogue(s) function (BAF) which projects TC onto EB, searching for TC's best candidate analogue(s). BAF uses the notion of distributionally-driven word-sense analogy developed in the previous pages, and can be informally described through the following steps: a) if EB contains a pattern wcp, which fully matches TC . at the input level, then wcp, is the best analogue and its output is ranked first in the list of available answers; note that this step does not stop SENSE from continuing its search; b) if EB contains a single AF such that two of AF's nodes together cover TC's input representation in its entirety, the output representations associated with the matching nodes is added to the list of available answers with a ranking score, gauged as a function of type and quantity of supporting evidence (see below for more detail); c) if steps a) and b) yield no result, no output is provided by SENSE.</Paragraph> <Paragraph position="2"> BAF at work Let us look at some interesting cases of BAF at work. Note that all examples considered in this paper are representative of real test suites of SENSE, and the assumed knowledge in EB reflects the current status of an actual data base acquired from typical examples of use within verb entries of the Collins Italian-English dictionary \[1985\].</Paragraph> <Paragraph position="3"> Suppose that SENSE has to assign a verb sense in the target context accendere?-ptpa_l/O 'light-pipe'. The context being not attested in EB, TC is projected onto EB's network, for a relevant AF to be found. The AF in Fout! Onbekende sehakeloptie-instructie, above is a good instance of such a relevant family, since it contains two nodes, namely accendere_l-object and plpa_l, which fully cover TC's input. Step a) having failed, the two nodes in question are not directly connected; nonetheless, their belonging to the same family means that there exists a continuous path of complementarity arcs joining the two. This continuity allows SENSE to hypothesize an arc directly connecting accendere_l-object with pzpa_l (represented as a dashed line in Font! Onbekende schakeloptie-instructie, below): fumare_l-objeet o~./&quot;'/e pzpa l accendere l-objeet i'/'/~ s:garetta_l Figure A reconstructed connection i.e. the co-occurrence pattern accendere_l-ptpa_l/O can be reconstructed on the basis of the available distributional evidence, and supports the interpretation accendere 1.</Paragraph> <Paragraph position="4"> To sum up, SENSE identifies a distributional similarity between accenderel and fumare_l: this similarity is based on the fact that cigarettes can both be lit and smoked. This triggers the analogy-based inference that pipes, besides being smoked, can also be lit, thus supporting the interpretation of accendere 1</Paragraph> </Section> <Section position="3" start_page="92" end_page="93" type="sub_section"> <SectionTitle> 3.3 Constraints on distributionaily-based WSD </SectionTitle> <Paragraph position="0"> In the example illustrated above, nouns stand in the same syntactic relation to the verbs. However, it is often the case that clusters of nouns which function as the object of a given verb can function as typical subjects of other, somehow related, verbs. If this sort of systematic subject/object alternation is taken into account, the inferential power of distributionally-based WSD may increase considerably, as shown by the following exampies. null Consider the TC attaccare_?-fotografial/O 'hang_up-photograph'. EB contains three different senses of attaccare, each attested with the following sets of noun collocates:</Paragraph> <Paragraph position="2"> No one of the noun collocates listed above happens to be attested in EB as an object of verbs which also combine with fotografia as an object. However, if the restriction that relevant nouns must stand in the same relation to the predicate is relaxed, then relevant distributional evidence can in fact be found in EB. Fotografia and quadro 'painting' are both attested as typical subjects of the verb rappresentare_l, a fact which can be interpreted in terms of Pustejovsky's telic role \[Pustejovsky 1995\], since both nouns are normally used to &quot;show something&quot;. Furthermore, quadro is also attested as a typical object of the verb attaccare_l; on this basis, it can reasonably be supposed that also fotografia, when co-occurring as an object of attaccare, points to the sense attaccare 1.</Paragraph> <Paragraph position="3"> Inferences based on AFs involving asymmetric syntactic dependencies permit to exploit the data contained in EB to the full. Moreover, the procedure becomes essential for generalising over cases of so-called valency alternation. Consider the causative-inchoative alternation, which involves two argument structures, a transitive and an intransitive one: a verb such as aumentare 'increase' can be used in a sentence like la Fiat ha aumentato gh stipendl agh operat 'Fiat increased salaries to workers', where stipendio is the object of the verb, and in a sentence like gh snpendt aumentarono inaspettatamente 'salaries increased unexpectedly', where sttpendio is the subject. In the literature, the theoretical issue of whether alternating argument structures of the same verb should be associated with a unique sense or with different senses of that verb is still open.</Paragraph> <Paragraph position="4"> In practice, lexicographers' approaches vary considerably, depending on factors such as the dictionary's internal structure or main practical purpose: for instance, in bilingual dictionaries different but alternating argument structures often give rise to different senses, due to differences in their translation. Whatever approach is adopted by the lexicographer, however, SENSE is capable of identifying a sense alternation induced by an alternation of argument structure, or~ alternatively, of recognising two different argument structures as related to the same verb sense, thanks to its ability to deal with asymmetric syntactic dependencies in EB.</Paragraph> <Paragraph position="5"> To sum up, word sense disambiguation with verb-noun pairs involving asymmetric dependencies is more effective than when only contexts with symmetric dependencies are considered. This procedure is particularly crucial for verbs alternating between different arguments structures.</Paragraph> </Section> <Section position="4" start_page="93" end_page="93" type="sub_section"> <SectionTitle> 3.4 Beyond attested evidence </SectionTitle> <Paragraph position="0"> SENSE's inferential routine can go beyond attested evidence; in fact, the presence of an attested pattern which matches exactly TC's input does not prevent the system from exploring other hypotheses. This flexibility is often useful: when sense distinctions are fine grained and data in EB are sparse, distributional criteria get too coarse grained to be able to point to a unique sense interpretation. null Consider, for example, battere ?-mano_l/O 'hithand': in EB, this pattern is attested with the sense of clapping, as an instance of beating body parts with a regular rhythm (battere_3). However, there is at least another sense of battere which is appropriate in the context considered, namely battere_l, understood under the more general sense of hitting someone or something.</Paragraph> <Paragraph position="1"> In cases like this one, SENSE &quot;ambiguates&quot; the verb-noun pair received in input, by finding out other plausible sense assignments besides the one attested in EB. As a consequence, SENSE outputs more than one sense interpretation, while ranking the attested interpretation first. Identification of alternative sense assignments, although with lower ranking, comes in handy when the expected TC interpretation is not the attested one. This is reasonable, we believe, since WSD is often a matter of suggesting a set of more or less plausible interpretations in context rather than asserting one interpretation only; by taking attested evidence (no matter how representative) at face value one would wrongly ignore the common fact that, even in real usages, a target context can in fact be understood in more than one way.</Paragraph> </Section> <Section position="5" start_page="93" end_page="94" type="sub_section"> <SectionTitle> 3.5 Ranking multiple disambiguation re- sults </SectionTitle> <Paragraph position="0"> As just shown, distributlonally-based word sense disambiguation does not always make the system converge on a unique interpretation. This situation typically occurs when different senses of a word are close in meaning, and this closeness is reflected by their co-occurrence with distributionally similar if not identical words. When more than one sense interpretation appears to be plausible, different strategies can be followed in order to rank them from more to less likely. When the set of plausible interpretations includes a directly attested one, then the latter is always ranked first. Ranking of inferred interpretations needs to take into account a number of different factors.</Paragraph> <Paragraph position="1"> As a first approximation, different sense interpretations can be ranked according to the number of AFs supporting them.</Paragraph> <Paragraph position="2"> Suppose that SENSE has to assign a verb sense in accarezzare_?-speranza_I/O 'toy_with-hope'. Both possible sense interpretations of accarezzare (i.e. accarezzare 1 'stroke' and accarezzare2 'toy_with') are supported. In EB, the interpretation accarezzare_l is supported by one AF only, which includes the pattern perderel-capellol/O 'lose-hair'.</Paragraph> <Paragraph position="3"> On the other hand accarezzare2 is supported by four AFs, each containing the following hooks: 1. abbandonare_4-progetto_l/O 'giveup-project' 2. cullare 1-tdea 1/0 'cherish-idea' 3. nascere 2-tdea_l/S 'be_born-idea' 4. nau~agare 1-progetto_I/S 'fall~hrough-project' Hence, accarezzare2 gets score 4 and wins out over accarezzare I which scores 1.</Paragraph> <Paragraph position="4"> The sheer number of supporting AFs, however, is too gross a criterion when used on its own. Consider the target affluire?-acqua_l/S 'flow-water'. Here, the contextually more appropriate sense affluwel 'flow' is supported by three AFs, while affluwe2 'pourin' is pointed to by five different AFs: af~uire l 1. intorbtdare_l-liqutda-1/O ' cloud-liquid' 2. penetrare 2-1iquido_l/S 'percolate-liquid' 3. versare_2-liqutdo_l/O'pour-liquid' affluire_2 1. conflutre_l-persona_l/S 'join-person' 2. gettare 1-persona_1/O 'rush_in-person' 3. imbarcare_l-mercel/O 'ship-goods' 4. insinuarsi_3-persona_l/S 'creep_into-person' 5. ristagnare_l-persona_l/S 'lag_person' Nonetheless, SENSE could be &quot;persuaded&quot; to prefer the correct interpretation if also the typology of supporting evidence is taken into account. Intuitively, preference has to be given to more specific supporting semantic evidence over semantically vaguer one. In our terms, this means that supporting AFs which contain a more specific hook noun should carry more weight for WSD than AFs containing vaguer hook nouns. Usually, generality of a word is measured by referring to a semantic hierarchy. In this context we have used frequency of word occurrence in EB as a convenient measure of &quot;generality/specificity&quot; of a word: the more often a hook noun occurs as a subject/object of differents verbs in EB, the more general it can be considered. Note that EB contains only WCP types, so that word counting here is significantly different from counting token frequencies in a real text; type frequency appears to point more decisively to the general structure of lexical competence, rather than to distributional effects in language performance. On this basis, each relevant AF is assigned a specificity score, equal to the inverse ratio of the number of times its hook noun occurs in EB. The ranking score of a given sense interpretation S is then the sum of the specificity scores of all AFs = { AFt, AF~ ..... AF, } support-</Paragraph> <Paragraph position="6"> where Spec(AF,)= 1/type-frequency(hook_noun).</Paragraph> <Paragraph position="7"> ha the light of this score, :ranking of the senses of affluwe is reversed: the best disambiguation hypothesis is now aJ: flutre_l (ranking score 0.281046), against afflutre_2 whose ranking score 0.069598 is significantly lower. The hook noun supporting the sense affluire_l is hqutdo 'liquid', whose specificity score is 0.111111 when used as an object and 0.058824 when used as a subject. By contrast, the same score is significantly lower in the cases supporting the other sense: 0.007194 for persona_l/O and 0.005650 for personal~S; 0.045455 for merce 1/0.</Paragraph> <Paragraph position="8"> The specificity score tends to overrate very specific analogies, that is analogies supported by analogical families with highly idiosynractic collocates, over more general analogies.</Paragraph> <Paragraph position="9"> To counterbalance this bias, another ranking factor, called the &quot;coverage&quot; score, can usefully be exploited in our context. For each available sense interpretation of TC attested in EB, we count how many of its collocates occur as hook nouns of all AFs supporting that sense. Note that, for an AF to support a certain verb sense, it has to contain as a hook noun a collocate of the verb sense S in question. We then assign to S a coverage score S~ov~:~o~, which is proportional to the number of shared collocates: S,~w~e ~ = # nour~collocate(AF(S))/#noun collocate(S) where '# nout~collocate(AF(S))' reads &quot;cardinality of the noun collocates of the AFs supporting S&quot;, and '# nouncollocate(S)' reads &quot;cardinality of the noun collocates of S&quot;. The bigger this score, the more widely supported the corresponding sense interpretation in EB. This follows quite naturally in an analogy-based perspective, since, intuitively, two verb senses are considered more similar if they have more collocates in common. Eventually, this score is combined with the other scores considered above to yield a fmal ranking score: S lo~1 ~nS_~or~ = Ssr~:o~ x S~ov~gn_~r~ To give a concrete example, assume that SENSE has to interpret the pattern accostare?-qualcuno_l/O 'approachsomebody'. Accostare is attested in EB in three different senses: accostare_l 'bring_near' with words like chair, object and ladder, among its typical objects; accostare2 'approach' with person as typical object; accostare3 'setajar' said of shutter and door. If the coverage score is not considered, the ranking would be accostare 3 (0.428571), accostare 1 (0.316417), accostare2 (0.12~02) the latter being the appropriate sense in this context. Intuitively this is due to the fact, that, for example, the AFs supporting accostare 3 all exhibit one hook noun only, namely porta, which none~eless contributes a high specificity score, due to its poor type-frequency in EB. Yet, if the coverage score is taken into account, the ranking becomes accostare 2 (2.079136), accostarel (1.582087), accostare3 (0.4~571), with the appropriate sense ranked first.</Paragraph> </Section> </Section> <Section position="5" start_page="94" end_page="95" type="metho"> <SectionTitle> 4 SENSE: experimental results </SectionTitle> <Paragraph position="0"> Experiments have been carried out with an EB of 8,153 distinct verb-noun patterns (2,488 verb-subject, 5,665 verb-object) automatically extracted from the whole set of verb entries of the Collins bilingual Italian-English dictionary \[Montemagni, 1995\]. In these patterns only verbs are disambiguated as to their sense, whereas nouns are assigned all possible senses. These patterns exemplify 3,359 different verb senses, each illustrated, on average, through 2.42 patterns. In Font! Onbekende schakeloptie-instructie, below, verb senses are ranked per number of exemplifying patterns: Senses which are attested in ten or more patterns are a negligible part of EB; actually, most verbs are illustrated by means of a number of patterns ranging between 2 and 5. Finally, a considerable group of verb senses is attested only once. Note that this does not stop SENSE from recognising them in unseen contexts; e.g. in EB there is only one pattern exemplifying the verb sense abbassare_3 'reduce' (namely, abbassare_3-prezzo/O), but this does not prevent SENSE from recognising it in target contexts such as abbassare_?-sttpendzo/O.</Paragraph> <Paragraph position="1"> SENSE's performance has been tested on a corpus of 150 TCs randomly extracted from unrestricted texts.</Paragraph> <Paragraph position="2"> Patterns which already occur in EB were excluded from the test corpus since we wanted to focus on the reliability of inferences based on distributional evidence, rather than on EB's statistical representativity. The results of this experiment are reported below: monosemic verbs; here, recall and precision are high and refer to the topmost sense in the ranking only. In the second column, recall and precision are relative to polysemous verbs only, and in spite of an obvious decrease compare well with related work carried out with different methods (see, for instance, \[Agirre and Rigau, 1996\]), and are in fact very promising if one considers the comparatively small size of EB, and that only part of its attested words are semantically disambiguated.</Paragraph> </Section> class="xml-element"></Paper>