<?xml version="1.0" standalone="yes"?>
<Paper uid="A00-2006">
  <Title>Encoding information on adjectives in a lexical-semantic net for computational applications</Title>
  <Section position="2" start_page="0" end_page="43" type="metho">
    <SectionTitle>
1 Adjectives in WN
</SectionTitle>
    <Paragraph position="0"> In WN adjectives are divided into two major classes: descriptive adjectives and relational adjectives. A descriptive adjective is &amp;quot;one that ascribes a value of an attribute to a noun&amp;quot; (Fellbaum et al. 1993:27). Descriptive adjectives combine with nouns to express some qualities of the thing, person or concept they designate.</Paragraph>
    <Paragraph position="1"> Typically, in this group we find adjectives that designate the physical dimension of an object, its weight, abstract values etc. Besides these referent-modifying adjectives, we also find reference-modifying adjectives (cf. Bolinger 1967; Chierchia &amp; McConnel-Ginet 1990 name them intensional adjectives). Typical examples of the latter are former, future, present.</Paragraph>
    <Paragraph position="2"> 3 ItalWordNet will be the reference lexical resource among various integrated language resources and software tools for the automatic treatment of the Italian written and spoken language which are being developed within the SI-TAL ('Integrated System for the Automatic Treatment of Language') National Project.</Paragraph>
    <Paragraph position="3"> 4 Actually, we shall only encode adverbs derived from adjectives by adding the suffix -mente, for which a derivation relation with an adjective will be encoded.</Paragraph>
    <Paragraph position="4">  Relational adjectives, on the other hand, mean something like &amp;quot;relating/pertaining to, associated with&amp;quot;, and usually have a morphologically strong link with a noun. Typical examples are musical, atomic, chemical 5.</Paragraph>
    <Paragraph position="5"> Synonymy is the basic relation encoded for all the PoSs (since it is used to build synsets).</Paragraph>
    <Paragraph position="6"> While in WN the noun and verb networks are then mainly developed around the superordinate (hyper/hyponymy) relationship, the organization of descriptive adjectives can &amp;quot;be visualized in terms of barbell-like structures, with a direct antonym in the centre of each disk surrounded by its semantically similar adjectives (which constitute the indirect antonyms of the adjectives in the opposed disk)&amp;quot; (Fellbaum 1998a: 212). The main relation encoded for these adjective synsets is antonymy, claimed to be the most prominent relation, both from a psycholinguistic point of view and from a more strictly lexical-semantic one, in the definition of the semantics of descriptive adjectives. Hyponymy is substituted by a 'similarity' relation.</Paragraph>
    <Paragraph position="7"> Relational adjectives, on the contrary, are not organized in this way, because their semantics cannot be described by using these relations.</Paragraph>
    <Paragraph position="8"> Indeed, they only point to the noun to which they pertain (e.g. atomic is linked to atom).</Paragraph>
    <Paragraph position="9"> Finally, information on the selectional preferences of both descriptive and relational adjectives is sometimes encoded (e.g., between high and degree), by using an 'is_attribute_of' relation.</Paragraph>
  </Section>
  <Section position="3" start_page="43" end_page="44" type="metho">
    <SectionTitle>
2 The IWN relations for adjectives
</SectionTitle>
    <Paragraph position="0"> As for the other lexical categories, also in IWN the basic relation encoded for adjectives is synonymy, on the basis of which synsets are built. Following EWN, we also encode a NEAR._SYNONYMY relation when two synsets are very close in meaning but their members cannot be put in the same synset (and no other relation results appropriate to link them; see Alonge et s Note that some adjectives have both a descriptive sense and a relational one. For example, musicale (musical) can modify the noun voce (voice) when we want to say that a voice is sweet-sounding and melodious, but can also be combined with the noun strumento (instrument) when we want to indicate that an instrument can be used to produce music.</Paragraph>
    <Paragraph position="1"> al. 1998 for a discussion of this relation, and Alonge et al., in prep., for a complete and detailed discussion of the linguistic design of IWN). Then, we encode a number of additional relations, which have been identified by taking into consideration i) theoretical works; ii) the EAGLES recommendations on semantic encoding (cf. Sanfilippo et al. 1999); iii) the data available in our sources; iv) possible use of data encoded in computational applications.</Paragraph>
    <Section position="1" start_page="43" end_page="44" type="sub_section">
      <SectionTitle>
2.1 Hyponymy
</SectionTitle>
      <Paragraph position="0"> Together with synonymy, the hyponymy relation constitutes the 'bone structure' of both WN and EWN. However, as we have seen, in WN the possibility of encoding hyponymy for adjectives is denied and the basic relation encoded for adjectives is antonymy, while EWN did not really deal with adjectives and a complete network for them was not built.</Paragraph>
      <Paragraph position="1"> Within IWN we have reconsidered the possibility of encoding hyponymy for adjectives. By analysing data coming from machine-readable dictionaries we find subsets of adjectives which have a genus + differentia definition, like nouns or verbs. That is, these adjectives seem to be organised into classes sharing a superordinate. This is the case, e.g., of adjectives indicating a 'containing' property (acquoso - watery; alcalino - alkaline), or a 'suitable-for' property (difensivo - defensive; educativo - educational), etc. In IWN we have decided, therefore, to encode hyponymy also for these sets of adjectives. The taxonomies which can be built on the basis of this relation are different from those built for nouns or verbs, since they are generally very flat, consisting almost always of two levels only (an exception is the color adjectives taxonomy). However, by encoding a hyponymy relation for these adjectives, we obtain classes for which it will be possible to make various inferences. For instance, it will be possible to infer semantic preferences of certain classes: e.g., all the adjectives occurring in the taxonomy of contenente (containing) will occur as attributes of concrete nouns; adjectives found in the taxonomy of affetto (affected by an illness) will never be predicated of nouns referring to objects, etc. Furthermore, it will also be possible to infer information on syntactic characteristics /I/11  of adjectives found in the same taxonomy: e.g., the hyponyrns of atto (suitable for) are always found in predicative position (and do not accept any complements); the hyponyms of privo (lacking) may occur both in attributive and in predicative position (and may take certain prepositional complements), etc.</Paragraph>
      <Paragraph position="2"> As it was done for all the relations identified in EWN, we have built substition tests or diagnostic frames based on normality judgements (cf. Cruse 1986). Inserting two words in the test sentences built evokes a 'normality'/ 'abnormality' judgement on the basis of which each relation can be determined. These tests are used by encoders both to verify the existence of relations between synsets and to encode them in a consistent way (for a complete lists of the tests built see Alonge et al., in prep.)</Paragraph>
    </Section>
    <Section position="2" start_page="44" end_page="44" type="sub_section">
      <SectionTitle>
2.2 Antonymy
</SectionTitle>
      <Paragraph position="0"> As in WN, also in IWN the antonymy relation remains an important relation to describe the semantics of various adjectives. Following theoretical work (Lyons 1977; Cruse 1986), we have further distinguished between</Paragraph>
    </Section>
  </Section>
  <Section position="4" start_page="44" end_page="44" type="metho">
    <SectionTitle>
COMPLEMENTARY ANTONYMY and GRADABLE
</SectionTitle>
    <Paragraph position="0"> ANTONYMY 6. The former relation links adjectives referring to opposing properties/ concepts: when one holds the other is excluded (alive~dead). The latter relation is used for those antonym pairs which refer to gradable properties (long~short). In case it is not clear if two opposing adjectives refer to complementary or gradable properties, we can still use an underspecified ANTONYMY relation. Also this information can be useful for computational applications since word pairs presenting one of the two kinds of opposition may occur in different contexts (cf. Cruse 1986).</Paragraph>
    <Section position="1" start_page="44" end_page="44" type="sub_section">
      <SectionTitle>
2.3 Other relations
</SectionTitle>
      <Paragraph position="0"> In WN a relation between adjectives and nouns is encoded for relational adjectives which point to a noun to which they 'pertain': atomic~atom, industrial~industry, etc. This relation will be encoded also in IWN, by using the label PERTAINS TO.</Paragraph>
      <Paragraph position="1"> Another relation 'inherited' from WN can be useful to distinguish both adjective senses and their semantic preferences7:</Paragraph>
      <Paragraph position="3"> Other relations are then being encoded which are not in WN, but are encoded for nouns and verbs in EWN. In WN each PoS forms a separate system of language-internal relations and conceptually close concepts are totally separated only because they differ in PoS. In EWN, instead of using as main classificatory criterion the traditional distinction among PoSs, drawn upon heterogeneous criteria, a purely semantic distinction was adopted (following Lyons 1977).</Paragraph>
      <Paragraph position="4"> Thus, a distinction was drawn among I st order entities (loes - referred to by concrete nouns), 2 &amp;quot;d order entities (2oes - referred to by verbs, adjectives or nouns indicating properties, states, processes or events), and 3 rd order entities (3oes referred to by abstract nouns indicating propositions existing independently of time and space). By drawing this distinction it was possible to relate lexical items that, either within a language or across different languages, refer to close concepts, although they belong to different PoSs. Thus, as said above, the possibility to encode 'near-synonymy' between synsets of the same order (but different PoSs) was provided.</Paragraph>
      <Paragraph position="5"> Furthermore, other cross-PoS relations were identified which allow to obtain a better description of word meanings. In IWN we maintain the same distinction among semantic orders and encode for adjectives some relations which can be encoded for the other 2oes. In particular, we encode the 'INVOLVED' and 'CAUSE' relations.</Paragraph>
      <Paragraph position="6"> The INVOLVED relation links a 2oes with aloe or 3oe referring to a concept incorporated within</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="44" end_page="46" type="metho">
    <SectionTitle>
6 A similar distinction is also made within the
</SectionTitle>
    <Paragraph position="0"> SIMPLE EC project (LE-8346), whose goal is adding semantic information to the set of harmonized lexicons built within the PAROLE project for twelve European languages. Of course, the sub-classification ofantonymy can also be used for nouns and verbs.</Paragraph>
    <Paragraph position="1"> 7 Furthermore, these relations being encoded between an adjectival synset and a nominal or verbal one are also useful to distinguish adjective classes as described by Dixon (1991), and reported in Sanfilippo et al. (1999). Indeed, such classes are often indicated by the nouns linked to adjectives.</Paragraph>
    <Paragraph position="2">  the meaning of the 2oe 8. Examples for adjectives are given in the following:  EWN, will be encoded for a class of adjectives indicating the possibility of some events occurring: giudicabile (= che pub essere giudicato) (triable) LIABLE_TO giudicare (to judge) inaccostabile, inavvicinabile (= che non pub essere avvicinato) (which cannot be approached) LIABLETO awicinare, accostare (to approach) negative.</Paragraph>
    <Paragraph position="3"> 8 E.g., to lapidate has as INVOLVED_INSTRUMENT stone; to work has as INVOLVED_AGENT worker (see Alonge et al. 1998).</Paragraph>
    <Paragraph position="4"> 9 As said above, in EWN various features were encoded to make implications of relations explicit: conjunction and disjunction (for multiple relations of the same kind encoded for a synset); non-factivity (to indicate that a causal relation does not necessarily apply); intention (added to a cause relation to indicate intention to cause a certain result); negation (to explicitly encode the impossibility of a relation occurring). These features are also used in IWN. The table below gives an overview of the main relations being encoded for adjectives in IWN (for the other relations being encoded for adjectives see Alonge et al., in prep.):  Aktionsart) of the situation they refer to: a basic distinction was drawn between Static and Dynamic. The Situation Components represent a more conceptual and intuitive classification of word meanings because they can be viewed like the most salient semantic components of a concept. Examples of Situation Components are: Manner, Existence, Communication, Cause.</Paragraph>
    <Paragraph position="5"> Situation Type represents disjoint features that cannot be combined, whereas it is possible to assign any combination of Situation Components to a word meaning. Here below the Top Concepts identified for 2oes are shown:  In order to be able to draw generalizations on adjective meanings by using the TO, we partially modified this scheme. First of all, we moved the PROPERTY and RELATION nodes under the SITUATION COMPONENT node. This was done for two interconnected reasons: first of all, because this distinction is not directly linked to Aktionsart (lexical aspect), while the distinctions under SITUATION TYPE are Aktionsart distinctions, i.e. they are connected with the &amp;quot;the procedural characteristics (i.e. the 'phasal structure', 'time extension' and 'manner of development') ascribed to any given situation referred to by a verb phrase&amp;quot; (Bache 1982: 70) ~deg. Secondly, adjectives may refer to PROPERTIES or RELATIONS, but they may be either stative or not (cfr. e.g. Lakoff 1966; Quirk et al. 1985; Peters et al. 1999). Thus, in our system it has to be possible to specify that an adjective expresses a PROPERTY while being DYNAMIC. In any case, since many adjectives may have both a DYNAMIC sense and a STATIC one, we have also the possibility to under-specify this information Io Of course, in EWN all 2oes (and therefore also nouns or adjectives) can be classified according to their Aktionsart.</Paragraph>
    <Paragraph position="6"> by linking adjectives directly to the SITUATION TYPE node.</Paragraph>
    <Paragraph position="7"> Adjectives may indicate many different types of properties: temporal (passeggiata mattutina morning walk), psychological (canzone triste sad song), social (uomo ricco - rich man), physical (superficie legnosa - wooden surface), physiological (bambino magro - thin child), perceptive (minestra calda hot soup), quantitative (magra ricompensa - poor reward) and intensity properties (vino forte - strong wine). In the EWN TO there are already nodes which may be used to represent these distinctions (TIME, MENTAL, SOCIAL, PHYSICAL, QUANTITY) but we needed to better specify or also add some features. For example, we have added, under the already present node PHYSICAL, the node MATERIAL, to represent, among others, some Italian adjectives ending in -oso (for example legnoso - wooden, acquoso watery) which indicate the property of containing a certain material. Moreover, we added the node PHYSIOLOGICAL (to classify adjectives corresponding to tired, hungry, sick, etc.) under PHYSICAL. For adjectives denoting an intensity, we then added the node INTENSITY directly under the SITUATION COMPONENT node.</Paragraph>
    <Paragraph position="8"> One of the main problem we had was that no Top Concept in the EWN TO could be used to classify the reference-modifying adjectives (cf. above). These are a very particular kind of adjectives, because they do not indicate a property of the referent of the noun they modify. So, aiming at showing the distinction between referent-modifiers and reference-modifiers, we created two new Top Concepts under the node PROPERTY: ATTRIBUTE and FUNCTIONAL, where the latter can be used for reference-modifying adjectives (according to the definition provided by Chierchia &amp; McConnel-Ginet 1990 for the category referred to by these adjectives: &amp;quot;a function from properties to properties&amp;quot;).</Paragraph>
    <Paragraph position="9"> Like all descriptive adjectives, also the reference-modifiers classified under the node FUNCTIONAL can be linked to other SITUATION COMPONENTS. Functional adjectives for which the temporal aspect prevails (ex former, presente - present) can be classified under the node TIME; adjectives referring to some 'epistemological' property (potenziale potential, necessario - necessary) can be linked  tO MODAL 11; etc. A particular case of functional adjectives are the 'argumental' ones. They introduce a comparison between different entities (e.g., simile - similar, diverso - different, etc.). A comparison presupposes a relation between different entities so these adjectives can be linked to both PROPERTY and RELATION.</Paragraph>
    <Paragraph position="10"> Since in the EWN TO these two Top Concepts were two different kinds of SITUATION TYPE, they were mutually exclusive; now, in the IWN revised TO they can be conjoined.</Paragraph>
    <Paragraph position="11"> Here below the IWN Top Concepts for 2 &amp;quot;a Order Entities are shown:</Paragraph>
  </Section>
  <Section position="6" start_page="46" end_page="46" type="metho">
    <SectionTitle>
2 Ndeg ORDER ENTITY
SITUATION COMPONENT
</SectionTitle>
    <Paragraph position="0"> Although the ANTONYMY, PERTAINS_TO, and ISATTRIBUTEOF relations, already encoded in the Princeton WN, are fundamental relations to describe the adjective semantics, and they can be 11 Since this node is used for situations involving the possibility or likelihood of other situations.</Paragraph>
    <Paragraph position="1"> useful for computational applications which exploit our resource, we believe that the 'new' relations being encoded may provide equally relevant information, especially because many adjectives cannot be defined by means of the WN relations. Let's take into consideration just a few examples.</Paragraph>
    <Paragraph position="2"> The adjective depresso is ambiguous in that it has (at least) three readings: 1) which has been lowered, flattened (said of a land); 2) being in bad physical or moral conditions (said of an area, a country); 3) affected by depression (said of a person). For these three senses of the adjective we would encode different relations, extractable from our sources: depressol IS_CAUSEDBY deprimere (to lower)  The relations encoded could, e.g., help disambiguate the occurrences of the adjective in contexts such as: Gianni era depresso (Gianni was depressed) or Quella regione ~ depressa (That area is depressed). Indeed, by checking the semantic information encoded for the two senses of depressione linked to depresso, it's possible to provide the right interpretation for the sentences under analysis; on the other hand, the first sense of the adjective would be excluded because of the IS A VALUE OF relation encoded. In this case, also the TO links could be helpful: actually, depresso2 would be linked to SOCIAL and CONDITION, while depresso3 would be linked to MENTAL and EXPERIENCE.</Paragraph>
    <Paragraph position="3"> Many adjectives found in our sources do not seem to have (lexicalized) antonyms, nor cannot be defined by using the PERTAINS_TO or</Paragraph>
  </Section>
class="xml-element"></Paper>