<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-2072">
  <Title>Modeling Adjectives in Computational Relational Lexica</Title>
  <Section position="4" start_page="555" end_page="557" type="metho">
    <SectionTitle>
2 Empirical Issues
</SectionTitle>
    <Paragraph position="0"> Adjective semantic analysis and representation is far from being a trivial issue, as adjectives show a very particular linguistic behavior, namely in what concerns sense change depending on linguistic context. Being so, there are several different typologies and classifications of adjectives in the literature: semantic based classifications, syntactic based classifications, classifications regarding the relation holding between the adjective and the modified noun, and so on.</Paragraph>
    <Paragraph position="1"> As our work on this issue progresses, it has become clear that only a combination of syntactic and semantic criteria can offer interesting insights concerning adjective linguistic behavior and the identification of relevant common features, which may set the basis for an accurate modeling of this POS in computational relational lexica. In this section we will briefly look at some of the main adjective classifications.</Paragraph>
    <Paragraph position="2"> Regarding the way adjectives relate to the noun they modify, we consider two classes: property ascribing adjectives (in (1)), which add a new restriction to the properties introduced by the modified noun; and reference modifying adjectives (in (2)), which behave like a semantic operator, taking the reference of the modified noun as its argument1.</Paragraph>
    <Paragraph position="3"> (1) o livro azul 'the blue book' (2) o diamante falso 'the fake diamond' Adjectives like falso (fake), for instance, deal with concepts instead of real or referential objects, showing how a concept applies to a particular object. These adjectives constitute a closed class with very particular properties, which makes them somewhat close to semantic operators. In this work we will therefore focus on property ascribing adjectives.</Paragraph>
    <Paragraph position="4"> 1 This distinction between property ascribing adjectives and reference modifying adjectives is basically equivalent to the one used in the SIMPLE project (Lenci et al., 2000) (extensional vs. intensional adjectives, following Chierchia and McConnel-Ginet (1990)) to address the semantics of adjectives. This distinction is also included in the EAGLES recommendations for a semantic typology of adjectives.</Paragraph>
    <Paragraph position="5"> Demonte (1999) classifies property ascribing adjectives based on their intrinsic meaning, a classification combining syntactic and semantic criteria to determine which adjectives belong to which class. Two main subclasses are considered: descriptive adjectives and relational adjectives. Each of these classes displays specific semantic and syntactic properties.</Paragraph>
    <Paragraph position="6"> In languages like Portuguese, descriptive adjectives can occur both in attributive and predicative contexts, while relational adjectives occur almost exclusively in attributive contexts2. Both prenominal and postnominal positions are possible for descriptive adjectives in attributive contexts. Relational adjectives, on the contrary, can only occur in postnominal position. Finally, descriptive adjectives are gradable, i.e. they can co-occur with degree adverbs, which is not the case for relational adjectives. However, these criteria are not always sufficient to make a clear-cut distinction between relational and descriptive adjectives. Demonte (1999) proposes some additional criteria in order to make a more accurate distinction between these adjectives: their occurrence in comparative structures, and the formation of polarity systems.</Paragraph>
    <Paragraph position="7"> (3) a. O sabor desta laranja e mais doce do que o daquela.</Paragraph>
    <Paragraph position="8"> 'this orange taste is sweeter than that one's' b. o rapaz alto / o rapaz baixo 'the tall boy / the short boy' (4) a. *Este sabor e mais mineral do que aquele.</Paragraph>
    <Paragraph position="9"> 'this taste is more mineral than that one' b. o sabor mineral / *o sabor amineral 'the mineral taste / the amineral taste' But most of all, and besides all the syntactical contrasts we have mentioned above, there is a clear contrast in the way these two adjective classes relate to the noun they modify. Descriptive adjectives ascribe a single property, setting a value for an attribute, whereas relational adjectives introduce a set of properties.</Paragraph>
    <Paragraph position="10"> (5) o predio alto 'the high building' 2 Predicative contexts with relational adjectives are generally ruled out in Portuguese. Nonetheless, some specific contexts, like contrastive contexts, for instance, seem to license predicative uses of relational adjectives: (I) As proximas eleicoes sao autarquicas, nao sao presidenciais.</Paragraph>
    <Paragraph position="11"> 'next election will be autarchic, not presidential'  (6) a industria alimentar  'the alimentary industry' Looking at (5) and (6), we see that, while alto (high) sets the value of the height attribute of predio (building) to high, alimentar (alimentary) does not ascribe a single property, but a set of properties to industria (industry). Moreover, this set of properties corresponds to the main features describing another noun - alimento (food) in the example above. In fact, the way properties are ascribed to the modified nouns in (5) and in (6) are quite different. Ascribing a singular property usually corresponds to an incidence relation of this property in the nominal referent, while ascribing sets of properties usually entails more complex and diversified semantic relations.</Paragraph>
    <Paragraph position="12"> However, despite the relevance of the descriptive/relational dichotomy, it cannot account for the following contrasts: (7) a. *Ele viu a Maria alta.</Paragraph>
    <Paragraph position="13"> 'He saw Mary tall' b. Ele viu a Maria triste.</Paragraph>
    <Paragraph position="14"> 'He saw Mary sad'.</Paragraph>
    <Paragraph position="15"> Both alta and triste are descriptive adjectives, but they do not behave in the same way regarding secondary predication.</Paragraph>
    <Paragraph position="16"> We can refine the classification, considering, for instance, the opposition between accidental properties and permanent or inherent properties (this distinction goes back to Milsark (1974; 1977) and Carlson (1977)). According to this distinction, the property denoted by alta (tall) belongs to the latter class and the property denoted by triste (sad) to the former one. However, as pointed out by Marrafa (2004) and previous work, the characterization of adjectives on the basis of this dichotomy is not straightforward, since certain adjectives are ambiguous with regard to those properties, as it is the case of triste (sad). In the example above triste (sad) denotes an accidental property, but in an expression like um livro triste (a sad book) it denotes a permanent property.</Paragraph>
    <Paragraph position="17"> Intuitively, we can say that triste (sad) expresses a state of tristeza (sadness), but we let the discussion of the status of this relation out of the scope of this paper.</Paragraph>
    <Paragraph position="18"> Nevertheless, this kind of adjectives is of great importance to model telic verbs. The semantics of telic verbs involves a change of state of their theme argument, i.e. the subevent that closes the whole event is an atomic event, (a state) that affects the theme and is different from its initial state. As argued in Marrafa (2005) and previous work, by default, verbs like lavar (to wash) are associated to the following Lexical-Conceptual Structure (LCS' in Pustejovsky (1991)):  (8) [T [P act(x,y)and ~ Q(y)], [eQ(y)]] T:transition, P:process, e: event, Q: atomic event When syntactically realized, the telic subevent generally corresponds to an adjectival constituent, like in the example below: (9) Ele lavou a camisa bem lavada.</Paragraph>
    <Paragraph position="19">  'He washed the shirt well washed' In (9) the absence of the telic expression bem lavada (well washed) does not induce ungrammaticality. However, in the case of verbs like tornar (to make), it seems impossible to assign a value to Q independently of the telic expression.  (10) a. Ele tornou a Maria triste.</Paragraph>
    <Paragraph position="20"> 'He made Mary sad' b. *Ele tornou a Maria.</Paragraph>
    <Paragraph position="21"> 'He made Mary'  Along the lines of Marrafa (1993) and further work, verbs like tornar (to make) are assumed here to be LCS deficitary, the telic expression filling the gap of the LCS of the verb.</Paragraph>
    <Paragraph position="22"> As shown below, the troponyms of these verbs incorporate the telic state: (12) a. Ele entristeceu a Maria.</Paragraph>
    <Paragraph position="23"> 'He saddened Mary' b. *Ele entristeceu a Maria triste.</Paragraph>
    <Paragraph position="24"> 'He saddened Mary sad' The grammaticality contrast above is due to the fact that entristecer (to sadden) incorporates the telic state. This justifies that this verb can be paraphrased by tornar triste (to make sad).</Paragraph>
    <Paragraph position="25"> In this section we have mainly focused on property ascribing adjectives. We have considered two main subclasses, descriptive and relational adjectives, briefly presenting their syntactic and semantic behavior with regard to gradability, formation of polarity systems and their occurrence in predicative and attributive (both pronominally and postnominally) contexts and comparative structures. We have also addressed the issue of adjective relation with the noun they modify. Different adjective behavior regarding secondary predication is also discussed and analyzed in terms of the opposition between acci- null dental and permanent properties. The properties discussed in this section should be encoded in computational relational lexica such as wordnets.</Paragraph>
  </Section>
  <Section position="5" start_page="557" end_page="558" type="metho">
    <SectionTitle>
3 Adjectives in WordNet and in Eu-
</SectionTitle>
    <Paragraph position="0"> roWordNet Hyponymy is the main structuring relation both in WordNet and in EuroWordNet. However, the semantic organization of adjectives is entirely different from that of other POS: nothing like the hierarchies of hyponymic (in the semantic organization of nouns) and troponymic relations (in the semantic organization of verbs) is available for adjectives. Even if it is possible to find some small local hierarchies, hyperonymy/hyponymy is far from being the crucial semantic relation in the organization of adjectives in relational lexical databases such as wordnets.</Paragraph>
    <Paragraph position="1"> However, some authors working within the EuroWordNet framework have reconsidered the possibility of encoding hyponymy for adjectives. Hamp and Feldweg (1998), in the development of GermaNet, abandon the cluster organization of WordNet in favor of a hierarchical structuring of adjectives, arguing for a uniform treatment of all POS. Even though taxonomic chains of adjectives yield rather flat in comparison to those of nouns and verbs, these authors claim to derive more structural information from these small taxonomies than from clusters, as they seek to eliminate what they consider to be the 'rather fuzzy concept of indirect antonyms'. Even though the concept of indirect antonymy is not completely clear, it is not obvious to us why this fact should entail that adjectives must show a hierarchical organization instead.</Paragraph>
    <Paragraph position="2"> In ItalWordNet, Alonge et al. (2000) also organize adjectives into classes sharing a superordinate. These classes correspond to adjectives sharing some semantic features, and are generally rather flat. These authors argue for the possibility of inferring semantic preferences and syntactic characteristics of adjectives found in the same taxonomy. The SIMPLE project addresses the semantics of adjectives in a similar way, identifying a set of common features relevant for classifying and describing adjective behavior. However, as noted by Peters and Peters (2000), even though similarities exist &amp;quot;adjectives belonging to the same semantic class may differ from each other in numerous ways&amp;quot;, i.e. the classes established in this way are not homogeneous. null In WordNet, descriptive and relational adjectives are distinguished, first, by being encoded in separate files, and second, by the relations holding between synsets.</Paragraph>
    <Paragraph position="3"> Descriptive adjectives are organized in clusters of synsets, each cluster being associated by semantic similarity to a focal adjective which is linked to a contrasting cluster through an antonymy relation. Therefore, antonymy is the basic semantic relation used in WordNet to encode descriptive adjectives. As argued for in Miller (1998), this cluster organization of adjectives seems to mirror psychological principles. In fact, this organization is clearly motivated if we recognize that these adjectives main function regards the expression of attributes, and that an important number of attributes are bipolar.</Paragraph>
    <Paragraph position="4"> Relational adjectives, on the other hand, do not have antonyms. Therefore, they cannot be organized in opposite clusters. As pointed out by Levi (1978), the intrinsic meaning of these adjectives is something along the following lines: 'of, relating/pertaining to, associated with' some noun. The way these adjectives are encoded in WordNet mirrors this as it links relational adjectives to the nouns they relate to.</Paragraph>
    <Paragraph position="5"> In GermaNet a distinct treatment of relational and descriptive adjectives is abandoned, as the distinction between these two classes is considered to be 'not at all clear'. Nonetheless, the WordNet strategy for distinguishing between different adjective classes is maintained: listing lexical items in different files3.</Paragraph>
    <Paragraph position="6"> As pointed out in the previous section, even if the distinction between these two classes is not always clear-cut, testing adjectives against the set of syntactic and semantic criteria presented in section 2 allows us to distinguish descriptive from relational adjectives. We consider that this distinction can be mirrored in the database via the semantic relations expressed in the network, adjective listing in different files not being therefore necessary. In order to do this we propose several cross-POS relations, since in the EuroWordNet model, unlike what happens in WordNet where each POS forms a separate system, it is possible to relate lexical items belonging to different POS. Such an approach has the 3 GermaNet classifies the adjectives into 15 semantic classes, following the classes proposed by Hundsnurscher and Splett (1982), with some minor changes: perceptional, spatial, temporality-related, motion-related, material-related, weather-related, body-related, mood-related, spirit-related, behaviour-related, social-related, quantityrelated, relational and general adjectives. One special class is added for pertainyms.</Paragraph>
    <Paragraph position="7">  advantage of coping with adjective representation in lexical semantic databases without using strategies external to the lexical model, such as a priori semantic classes or separate files corresponding to different classes.</Paragraph>
  </Section>
  <Section position="6" start_page="558" end_page="559" type="metho">
    <SectionTitle>
4 Relating adjectives, nouns and verbs
</SectionTitle>
    <Paragraph position="0"> It is undeniable that important structural information can be extracted from the hierarchical organization of lexical items, namely of nouns and verbs. However, extending wordnets to all the main POS involves a revision of certain commonly used relations and the specification of several cross-POS relations.</Paragraph>
    <Paragraph position="1"> We previously mentioned that adjectives show a very particular semantic organization. Thus, encoding adjectives in wordnets calls for the specification of a number of cross-POS semantic relations. Here we use these cross-POS semantic relations to mirror adjectives main features in wordnet-like databases, which allows us to make adjective classes emerge from the relations expressed in the network.</Paragraph>
    <Paragraph position="2"> According to the strategies discussed in Mendes (2006), we present here the relations we argue are appropriate to encode adjectives and show how they conform to some complex phenomena. null</Paragraph>
    <Section position="1" start_page="558" end_page="559" type="sub_section">
      <SectionTitle>
4.1 Relating Adjectives and Nouns
</SectionTitle>
      <Paragraph position="0"> To put it somewhat simplistically, descriptive adjectives ascribe a value of an attribute to a noun. We link each descriptive adjective to the attribute it modifies via the semantic relation characterizes with regard to/can be characterized by4. Thus, instead of linking adjectives amongst themselves by a similarity relation, following what is done in WordNet, all adjectives modifying the same attribute are linked to the noun that lexicalizes this attribute. This way, and in combination with the antonymy relation, we obtain the cluster effect argued to be the basis of the organization of adjectives (Miller, 1998; Fellbaum et al, 1993), without having to encode it directly in the database.</Paragraph>
      <Paragraph position="1"> As shown by word association tests, antonymy is also a basic relation in the organization of descriptive adjectives. Nonetheless, this relation does not correspond to conceptual opposition, which is one of the semantic relations used for 4 This semantic relation is very close to the is a value of/attributes relation used in WordNet. We have changed its label in order to make it more straightforward to the common user.</Paragraph>
      <Paragraph position="2"> the definition of adjective clusters. We argue that conceptual opposition does not have to be explicitly encoded in wordnets, since it is possible to infer it from the combination of synonymy and antonymy relations (see Mendes (2006) for more details).</Paragraph>
      <Paragraph position="3"> Concerning relational adjectives, even though they are also property ascribing adjectives, they entail more complex and diversified relations between the set of properties they introduce and the modified noun, often pointing to the denotation of another noun (cf. section 2). We use the is related to relation to encode this.</Paragraph>
      <Paragraph position="4"> Therefore, the characterizes with regard to/can be characterized by and the antonymy relations, for descriptive adjectives, and the is related to relation for relational adjectives, allows us to encode the basic features of these adjectives in computational relational lexica such as wordnets, while making it possible to derive membership to these classes from the relations expressed in the network.</Paragraph>
      <Paragraph position="5"> Another issue regarding adjectives is that they have a rather sparse net of relations. We introduce a new relation to encode salient characteristics of nouns: is characteristic of/has as a characteristic to be. These characteristics are often expressed by adjectival expressions. Although in terms of lexical knowledge we can discuss the status of this relation, it regards crucial information for many wordnet-based applications, namely those using inference systems, allowing for richer and clearer synsets.</Paragraph>
      <Paragraph position="6"> Also, it may allow for deducing semantic domains from the database, as it makes it possible to identify the typical semantic domains of application of adjectives. Research on the classes and semantic domains emerging from the relations expressed in the database is still ongoing. Thus, the combination of these relations allows us to encode a less sparse net of adjectives. Besides the importance of having a more dense net from the point of view of wordnet-based applications, as mentioned above, this is also crucial with regard to relational lexica such as wordnets themselves, as the meaning of each unit is determined by the set of relations it holds with other units. Thus, a denser network of relations allows for richer and clearer synsets. Fig. 1 illustrates this idea, presenting an example of the way adjectives are being encoded in WordNet.PT. null</Paragraph>
    </Section>
    <Section position="2" start_page="559" end_page="559" type="sub_section">
      <SectionTitle>
4.2 Relating Adjectives and Verbs
</SectionTitle>
      <Paragraph position="0"> We also introduce new semantic relations to encode telic verbs in the database (on this issue see also Marrafa, 2005; Amaro et al., 2006).</Paragraph>
      <Paragraph position="1"> As shown in section 2, the facts render evident that the representation of LCS deficitary telic verbs has to include information regarding the telic expression. Obviously, it would not be adequate to overtly include in the synset all the expressions that can integrate the predicate, among other reasons, because they seem to constitute an open set. Rather, we claim that we can capture the telicity of these verbs by including a new relation in the set of internal relations of wordnets: the telic sub-event relation, as exemplified  below.</Paragraph>
      <Paragraph position="2"> (13) {make} has_telic_sub-event {state} {state} is_telic_sub-event_of{make}5  Relating make to state by means of this relation, we capture the telic properties of the verb and let the specific nature of the final state underspecified. This way, we also account for the weakness of the verb selection restrictions. As expected, we can also use this relation to encode telicity in the case of the troponyms of the class of verbs discussed in section 2.</Paragraph>
      <Paragraph position="3"> 5 Word senses presented here correspond to Princeton WordNet synsets (2.1 version).</Paragraph>
      <Paragraph position="4"> 6 The relation is not obligatory in this direction. In these cases, we use the telic sub-event relation to relate the verb to the expression corresponding to the incorporated telic information:  As shown, the telic sub-event relation straight-forwardly allows the encoding of lexical telicity in wordnets, in accordance with the empirical evidence.</Paragraph>
      <Paragraph position="5"> It should be noticed that the existing sub-event relation in the EuroWordNet framework is different from the relation proposed here. It only stands for lexical entailment involving temporal proper inclusion. Therefore, it does not account for the geometry of the event. On the contrary, the telic sub-event relation regards the atomic sub-event that is the ending point of the global  As previously mentioned, the proposal presented in this paper is mainly concerned with the specification of appropriate cross-POS relations to encode adjectives in computational relational lexica.</Paragraph>
      <Paragraph position="6"> In order to test whether the set of relations presented here is appropriate and allows the encoding of adjectives in wordnet-like lexica, we have introduced a selection of Portuguese adjectives in WordNet.PT.</Paragraph>
      <Paragraph position="7"> In the first phase of the WordNet.PT project mostly nouns were encoded in the database.</Paragraph>
      <Paragraph position="8"> Thus, we have mainly focused on the encoding of relations between adjectives and nouns7. Table 1 presents the number of entries and relations specified at the present stage.</Paragraph>
      <Paragraph position="9"> total number of adjectives 1462 synonymy relation 252 antonymy relation 134 near-antonymy relation 40 is related to relation 331 is characteristic of relation 1293 characterizes with regard to relation 261 total number of relations 2311 Table1. Statistics concerning the encoding of adjectives in WordNet.PT Besides the discussion presented above, the implemented data, being already a representative sample, show that the cross-POS relations proposed here effectively allow for a fine-grained encoding of adjectives in relational lexica (specifically in wordnet-like lexica) through the specification of a denser network of relations.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>