Wordnet As a Base Lexicon Model for the Computation of Verbal Predicates

WordNet as a base Lexicon Model for the Computation of Verbal Predicates

Raquel Amaro Group for the Computation of Lexical and Grammatical Knowledge Center of Linguistics of the University of Lisbon [email protected]

Abstract cates’ valence, reflected in restrictions on argu- Assuming that the lexicon is a complex and ment selection: intransitive verbs are expected dynamic knowledge system, the lexicon model to form a group, transitive verbs another group, and the information that is stated in it become ditransitive verbs yet another group, and so crucial for the computation of meaning. In on. More still, it is expected that the verbs this paper we discuss how the WordNet model in these groups have the same resctrictions, reflected in the syntax. However, the analysis of can support a decompositional approach to 1 troponymy, the enrichment of the lexical entries Portuguese verbs within WordNet.PT , focus- with Qualia information, and the establishment ing mainly on verbs of movement, allowed us of a lexical inheritance device, without much to notice that verbs conceptually and semanti- additional work. More specifically, we intend cally connected – in a hypernymy / troponymy to show how the choice of the model and relation – have different selection restrictions, its further enrichment is motivated by the and that the predicates’ valence is not enough reflex of the semantic content on the syntactic to determine groups: behaviour of the lexical items, namely through (1) a. tirar (take) is mover (move) from a co-occurrence restrictions, and address issues given location on argument structure, Aktionsart shifts, tirar is troponym of mover co-troponym’s incompatibilities and shadow arguments’ realization. We present a level of b. Ele moveu o caixote. inheritance structure in the semantic repre- ‘he moved the box’ sentation of the lexical items and propose a c. Ele tirou o caixote da rua. semantic representation for prepositional argu- ‘he took the box from-the street’ ments, using the Generative Lexicon notation. We also address the ability of the WN model The examples in (1) show that the transitive to be used as a semantic type hierarchy. verb mover and the ditransitive verb tirar are part of the same semantic domain (in (1a)), although displaying a different number of argu- 1 Introduction ments (cf. (1b) and (1c)). The lexicon can be seen as a complex and dy- The Aktionsart properties of the verbs are namic knowledge system that determines struc- also desirably mirrored in the organization of ture and meaning. In this perspective, the lexi- the lexicon, being state denoting verbs or activ- con model and the information that is stated in ity denoting verbs naturally close to each other. it are crucial for the computation of meaning. Nonetheless, there are cases such as the ones in The organization of the lexical entries within (2), where the troponym verb does not share the lexicon and the type of information that the Aktionsart values of its hypernym: the verb is introduced in the lexical entries should also voltar (to return) is an accomplishment denot- be motivated by processing issues, namely re- ing verb, hyponym of the activity denoting verb strictions on the (co-)occurrence of the lexical mover-se (move oneself). These Aktionsart dif- items. For instance, the model for the verbal ferences reflect on the co-occurrence restrictions lexicon must necessarily deal with the predi- with certain adverbials. Accomplishment denoting verbs can occur with in-adverbials but This research was supported by Funda¸cãopara a Ciênciae a Tecnologia (grant SFRH/BD/13875/2003). 1See Marrafa (2002). cannot occur with for-adverbials (see (2c)), However, these different particular manners whereas activity denoting verbs can occur with can be reflected on the argument structure of for-adverbials but not with in-adverbials (see verbs (number and / or type of arguments se- (2b)). lected, as exemplified in (1)), motivating a decompositional approach to troponymy. (2) a. voltar (return) is mover-se (move The hierarchichal organization of a relational oneself) back to the start location. net seems also to be able to account for phenom- voltar is troponym of mover-se ena such as lexical shadowing, namely the restrictions on the syntactic realization of shadow b. Ele moveu-se *em meia hora. arguments, as exemplified bellow: ‘he moved *in half an hour’ Ele moveu-se durante meia hora. (5) a. *Ele serrou a árvore com a serra. ‘he moved for half an hour’ ‘he sawed the tree with the saw’ c. Ele voltou em meia hora. b. Ele serrou a árvore com a serra de ferro. ‘he returned in half an hour’ ‘he sawed the tree with the iron saw’ Ele voltou *durante meia hora. ‘he returned *for half an hour’ c. Ele serrou a árvore com a motosserra. ‘he sawed the tree with the chainsaw’ Also, it was possible to observe that co- troponym verbs – verbs that have the same di- The construction of a relational lexicon al- rect hypernym – are not necessarily incompat- ready points out most of the crucial similarities ible, i.e. co-troponym co-occurrence is not al- and differences that are established in a decom- ways ruled out: positional analysis, being the enrichment of the WN model not highly costly. Moreover, the (3) a. Ele subiu a rua correndo. ontological nature of the WN model, besides ‘he moved up the street running’ accounting for already known phenomena such b. *Ele subiu a rua descendo. as lexical shadowing, seems to allow for the ‘he moved up the street moving down’ semantic representation of lexical items without recurring to a hierarchy of semantic types. c. Ele desceu a rua correndo. ‘he moved down the street running’ Section 2 describes how troponymy reflects the predicate’s meaning and structure. Sec- The verbs subir (move up), descer (move tion 3 focuses on the decompositional analysis down) and correr (run) are all troponyms of of troponymy, namely on how it is reflected on the verb mover-se (move oneself), but are not argument structure and Aktionsart properties. uniformely incompatible. The verbs subir and Section 4 discusses the relevance of introducing descer can co-occur with the verb correr ((3a) Qualia information in the lexical entries, consid- and (3c)) but cannot co-occur with each other ering verbal Qualia unification and co-troponym (see (3b)). compatibility. Section 5 is dedicated to the se- In this paper we propose to account for these mantic representation of the verbal predicates, data by extending the information in the lexical specifically in what concerns lexical inheritance 2 entries in a WordNet (WN) model lexicon with and the semantic content of prepositional ar- 3 Generative Lexicon (GL) structures. guments. Section 6 refers how the WN model Within a strict WN context – and to deter- can be used as a semantic type hierarchy, and mine the base relation of troponymy – it is not how the organization of this model can account necessary to consider or distinguish between dif- for the restrictions on phonetic realization of ferent types of manner (Fellbaum, 1998): shadow arguments. Finally, Section 5 presents our conclusions. (4) To V1 is to V2 in some particular manner. (Fellbaum, 1998, 79) 2 Troponymy and predicate’s 2The extension of the information in lexical entries, meaning particularly of verbs, is the focus of several works, namely the recent Koenig et al. (2002), Agirre and Mar- It is well known that the WN model shares tinez (2002), Dang et al. (2000), among others. some aspects with decompositional models. For 3See Pustejovsky (1995). instance, the correspondence between Lexical Conceptual Structures4 (LCS) primitives (GO d. ground, object with respect to which in (6a)) and verbs that are top nodes in the hi- the figure undergoes the event: erarchy (mover-se in (6b)) is quite transparent: aumentar (to increase) = make bigger with respect to a previous dimen- (6) a. LCS for andar (walk) sion. V e. cause, what brings about the event: NPj sufocar (to suffocate) = die due to lack of access to air or oxygen. [Event go [T hing ]j ] b. andar (walk) is mover-se (move f. source, figure initial location: oneself), slowly, resting one foot after remover (to take from place) = to take the other on a surface from the usual place. andar is troponym of mover-se g. goal, figure final location: encaixotar (to box) = to put in a box. In this way, the hypernym relation already determines the core meaning of the troponym, h. path, medium location(s) between the playing a role similar to that of semantic prim- source and the goal: itives in the event structure representation, as retroceder (to backtrack) = to move informally exemplified in (7). back through the same path. i. direction, way in which the motion (7) andar(e1, x, z): event occurs: e1 = move oneself, descer(to move down) = to move (in x = animate entity, which direction?) down. z = slowly, resting one foot after the other This decompositional analysis identifies what on a surface is specific of each troponym. The consequences are more or less visible, i.e. the lexicalization The troponymy relation definition, stated in of the semantic elements affects the predicate’s (4), allows us to explicitly “translate” the mean- inherited content in different degrees, mainly in ing specificity denoted by the troponym into se- what concerns argument structure, but also re- mantic components, since the core meaning is garding Aktionsart properties. already identified. 3.1 Argument structure 3 Decompositional analysis of The argument structure is the structure for troponymy the representation of logical arguments, inde- In Amaro (2005), a new set of semantic ele- pendently of their exact syntactical realization. ments, using the notion of incorporation as de- Here, we assume the types of arguments defined fined in Talmy (1985, 60), is proposed in order in Pustejovsky (1995, 63-64): to describe the meaning specificities that distinguish hypernyms from troponyms in a relational (9) a. True arguments (argn) are necessarily net. The resulting set is presented in (8): realized syntactically. b. Default arguments (d-argn) pertain to (8) a. manner, how the event develops: the semantic content of a lexical item, sussurrar (to whisper) = speak softly. but are not necessarily realized syntac- b. intention, purpose/ intended goal of tically. the event: c. Shadow arguments (s-argn) are seman- sacrificar (to sacrifice) = kill to tically incorporated in the meaning of please/ honor divine entities. the lexical item, and can only be re- c. figure, object that undergoes the alized syntactically through discourse event: specification devices. anestesiar (to anesthetize) = give anes- According to these types of arguments, the thesia. incorporation of goal in the synset {pôr,colo- 4See Jackendoff (1983). car} (put), represented a la GL in (11), is reflected by an increase of the list of true argu- tree, without artificially isolating lexical items ments inherited from the hypernym, {mover, conceptually related for not sharing the same deslocar} (move, change location), in (10). The argument structure. Also, it becomes possible expression of the final location, the goal – the to characterize lexical inheritance using the lex- arg3 introduced by a preposition – becomes ical semantic structures available through the obligatory. hypernymy relation. 3.2 Aktionsart shifting (10){mover, deslocar}(move):  ... The analysis of semantic incorporation also al-  " # lows us to account for Aktionsart differences be-  arg = 1 x:entity  arg-str = 1  tween hypernyms and troponyms. arg2 = 2 y:entity We use the four Aktionsart classes defined in Vendler (1967), following Móia(2000)’s set of (11) {pôr,colocar}(put):   criteria: ...      (14) state: non-punctual (i.e. temporally un-  arg1 = 1     bounded), homogeneous, and simple (non- arg-str = arg2 = 2  complex) event. arg = 3 z:goal 3 activity: non-punctual, relatively homogeneous (i.e. homogeneous considering Again, this argument structure is inherited subintervals), and simple event. by the immediate troponym, {meter} (put inside of), in (12), that expresses a more specific accomplishment: non-punctual, hetero- direction of the event, replacing the prepo- geneous, and complex event (preparatory sition5. The lowest troponym, {enjaular}, in- process, culmination point and consequent corporates a specific goal location, a cage, state). through lexical shadowing, thus changing the achievement: punctual, heterogeneous, predicate type of argument from true argument and complex event (culmination point and to shadow argument. consequent state).

(12) {meter}(put inside of): These ontological properties are reﬂected   ... (and therefore testable) in terms of distribu-      tional properties. As follows,  arg1 = 1     arg-str = arg2 = 2  (15) state denoting predicates can not oc- arg3 = 3 cur in the present continuous tense form (is V-ing) (13) {enjaular}(cage):   *John is being tall; ...      activity denoting predicates occur in  arg1 = 1     the present continuous tense form (is V- arg-str = arg2 = 2  ing), entailing that if X is (now) V-ing than s-arg = z:cage 3 X has already V In this fashion it is possible to account for dif- John is running → John has already ran, ferent argument structures within a troponymy occur with for-adverbials, but not with in- adverbials 5Prepositions, although not the focus of the current paper, will eventually be treated within the same frame- John run for 10 minutes /*in 10 minutes; work, in order to account for these data. Since the prepositions at stake have clear semantic content, i.e., they are more than theta markers, it is possible to assume a re- accomplishment denoting predicates lational net in which the concept denoted by dentro de occur in the present continuous tense form (inside of) is in a hyponymy-type relation with the con- (is V-ing), entailing that if X is (now) V- cept denoted by em (in) – if x is inside of z then x is in ing than X has not yet already V z. In this way, it would not be necessary to consider a preposition replacement in (12), but rather a speciﬁca- John is building the house → John has not tion through a hyponym. yet built the house, occur with in-adverbials, but not with for- c. Ele saiu da casa adverbials ‘he moved out of-the house John built the house in 10 months /*for *durante 10 min. / *em 10 min. 10 months; *for 10 min. / *in 10 min.’

achievement denoting predicates can The regularity exhibited in the cases of se- not occur with in-adverbials, nor with for- mantic incorporation of source, goal and adverbials manner and Aktionsart shifting, as well as the argument structures’ differences presented, pro- John died *in 10 months /*for 10 months. vide strong motivation for some level of decom- The cases observed within the class of verbs of positional analysis in the representation of the movement involve the incorporation of source, lexicon. goal, and manner. Moreover, it is clear that this level of de- Intuitively, the incorporation of source and compositional analysis is, for the most part, re- goal establishes a limit to the event, shifting an coverable from a relational lexicon construction activity type to an accomplishment type event, process. in (16) and (17), whereas the incorporation of manner, in (18), may be described as the addi- 4 Qualia information tion of a process to an achievement type event6. The association of Qualia information (Puste- jovsky, 1995) to other lexicon models is nowa- (16) a. {tirar accomplish.} (take) [moveactivity days a commonly accepted strategy to enhance from source] the lexicon model expressiveness, namely in b. Ele tirou a caixa da rua what concerns the construction of lexica for ‘he took the box from-the street computational purposes.7 em 10 min. / *durante 10 min. According to Mendes and Chaves (2001), the in 10 min. / *for 10 min.’ association of Qualia information to synsets enables the model to account for issues such as c. Ele moveu/deslocou a caixa co-hyponyms co-occurrence. By recognizing the ‘he moved the box different aspects of the meaning specification *em 10 min. / durante 10 min. involved in the hypernymy/hyponymy relation *in 10 min. / for 10 min.’ through the Qualia values of each synset, it is possible to predict the compatibility in (19a) as

(17) a. {voltar, regressar accomplish.} (come well as the incompatibility in (19b): back / return) [move oneselfactivity to the start point] (19) a. This dog is a pitbull[formal] and a police-dog [telic]. b. Ele voltou / regressou ‘he came back / returned b. *This dog is a pitbull[formal] and a em 10 min. / *durante 10 min. Saint-Bernard[formal]. in 10 min. / *for 10 min.’ This contrast refutes the assumption that all c. Ele moveu-se *em 10 min. / co-hyponyms, as specifications of the same con- ‘he moved *in 10 min. / cept, are incompatible, since the data show that durante 10 min. this incompatibility depends on the type of spe- for 10 min.’ cific information each co-hyponym denotes. 4.1 Qualia unification (18) a. {esgueirar-seaccomplish.} (sneak out) [move-out-of furtively] To account for this phenomenon, Mendes and activity Chaves (2001) assume a unification operation. b. Ele esgueirou-se da casa Incompatibility is expressed considering that ‘he sneaked out of-the house co-hyponyms are incompatible if their Qualia em 10 min. / *durante 10 min. structures do not unify: in 10 min. / *for 10 min.’ 7See, for instance, Fabre and Sébillot(1999), Busa et 6These Aktionsart shifts are predicted in Moens al. (2001), De Boni and Manandhar (2002), O’Hara and (1987, 45). Wiebe (2003), Veale (2003), among others. (20) Two qualia structures do not unify if there b. Ele subiu a rua, is a quale Q from two nominal qualia struc- ‘he moved up the street tures [Q=R1] and [Q=R2] afastando-se da casa. where roles R1 and R2 exist such that moving away from the house’ ¬(R1=R2) ∧¬ subsumes(R1,R2). {subir}(move up)direction:up, {afastar-se}(move away)direction:away The second conjunct in the last restriction (¬ subsumes(R1,R2)) assures that cases of two Note that, as stated before, the semantic in- co-hyponyms with different values for the same corporation is reflected in the argument struc- Qualia role are compatible, since one of these ture of the verbs, and not directly in the Qualia values is a subtype of the other, such as the role’s values. Verbal Qualia structure’s values ones in (21): are somewhat different from those of nouns and the verbal semantic structure seems to intro- (21) a. This animal is a predator and a mam- duce several difficulties. mal. 4.2 Verbal Qualia structure b. {predator" }: # formal=animal The first issues to consider relate to the quale Q= roles values licensed for verbs. Verbal Qualia telic=hunting structure is fulfilled with semantic predicates c. {mammal}: that establish the relations between the argu- h i ments of a verb. Q= formal=mammal Typically, only two Qualia roles are used in d. {mammal} @-> {animal}. this description: the formal role, for activity and state type events; and the agentive Here the values of R1 and of R2, animal and (bearing the causal chain) and formal (sta- mammal of the formal quale are not equal, ting the final state) roles for accomplishment 8 but R1 is a subtype R2, accounting for the com- and achievement type events , as exemplified in patibility of these co-hyponyms. (24): According to the decompositional analysis presented in section 2, it is possible to consider (24) a. eathactivity: i a version of this operation referring the seman- Q= formal=eat act(e1,x,y) tic elements incorporated in the verbs. Thus, two co-troponym verbs are incompatible if some b. kill : value of the semantic elements incorporated in accomplishment" # formal=kill result(e ,y) their semantic structures do not unify: Q= 2 agentive=kill act(e1,x,y) (22) a. Ele saiu da casa, correndo ‘he exited the house, running’ In this way, we have the following represen- {sair}direction:out, {correr}manner:... tations of the verbs in question: b. *Ele saiu da casa, entrando (25) a. correr (run): ‘he exited the house, entering’  " # {sair} , arg =x:ind direction:out  1  arg-str=  {entrar}direction:in  s-arg1=z:manner   h i  At a first glance, this hypothesis seems to q= formal=run act(e1,x,z) cover the same kind of co-occurrence restrictions on verbs co-hyponyms. However, cases such as (23b) are not accounted for, since the values of the element incorporated refer to direction but are, all the same, compatible.

(23) a. Ele saiu da casa, recuando. ‘he exited the house, moving back’ 8See Pustejovsky (1995, 183-188). Note that, how- {sair}(exit)source, ever, some authors also use other Qualia roles, namely {recuar}(move back)direction the Telic, as in Marrafa and Moura (2005). b. sair (exit, move out): count for co-troponym compatibility via indi-    rect Qualia uniﬁcation: arg1=x:ind     arg =y:loc  arg-str= 2  (26) Two co-troponym verbs are incompatible    s-arg1=z:out  iﬀ the arguments in their argument struc-  " # tures refer to incompatible co-hyponyms,  formal=¬at(e ,x,y)  q = 2  i.e. if the Qualia structures of these argu- ag.=exit act(e1,x,z) ments do not unify.

c. entrar (enter, move in): Let us reconsider the verbs in (19), here in (27):     arg1=x:ind     (27) a. sair (exit):  arg =y:loc    arg-str= 2     arg1 = x: ind  s-arg1=z:in     " # arg-str =arg2 = y: location  formal=at(e ,x,y)  q= 2  s-arg1 = z: out ag.=enter act(e1,x,z) b. entrar (enter):   arg1 = x: ind Although the incompatibility of the co-   troponyms sair (exit) and entrar (enter), in arg-str =arg2 = y:location (25b) and (25c) is covered, since these verbs s-arg1 =z:in have comparable values in their Qualia roles, c. correr (run): to assume the unification operation as directly " # arg = x: ind applying to verb Qualia structures is to pre- arg-str = 1 dict that correr (run) is incompatible with sair s-arg1 = z:manner (exit) or entrar (enter) since the values of their formal role are different and none subsumes Indirect Qualia structure unification enables the other – the formal quale role value is ac- us to predict that the co-troponym verbs sair tually contrary, since the verb sair (exit) incor- (exit) and entrar (enter) are incompatible since porates source, denoting a final state of not the Qualia structure of out and in - which are co-hyponyms - do not unify. Conversely, co- being at a given location,(¬at(e2,x,y)), whereas the verb entrar (enter) incorporates goal, de- troponym verbs sair (exit) and correr (run) are noting a final state of being at a given location, compatible since there is no co-hyponym arguments in their structure. (at(e2,x,y)). On the other hand, to allow for unification disregarding the Qualia roles’ values To consider Qualia information in a relational is to ignore evidence relating the Qualia roles’ model of the lexicon proved once again of rele- values with the different meaning specifications vance. Also, the determination of the values for among co-troponyms. This seems to show that the Qualia roles does not require any additional Qualia structure unification is not feasible for work, since it only involves the statement of the the verbal lexicon. meaning specifications that lead us to consider a troponymy relation between two items. 4.3 Indirect Qualia unification 5 Semantic representation The data presented so far regarding compatibility issues among co-troponyms, and the fact 5.1 Lexical Inheritance that it is possible to relate this with the se- In the previous section we focused on how to mantic elements incorporated in the troponyms, use Qualia information to determine which co- constitute evidence that Qualia structure unifi- troponyms are compatible and which are not. cation is occurring at some level. And yet, we have not considered the hierarchical Considering that it is in the argument nature of the WN model and it’s major relation structure that the logical arguments of a – the hypernymy relation. predicate are listed, and that verbs’ argument The compatibility issues arise among co- structure necessarily reflect the incorpora- troponym verbs, i.e. verbs that have the same tion of the semantic elements responsible for the hypernym. In fact, co-troponyms are predicted meaning specificities between hypernyms and to be incompatible since they can denote contra- troponyms (see section 2), we propose to ac- dictory specifications of a shared core meaning – that of their hypernym. In other terms, it is inheritance-structure – enables us to possible to consider that the core meaning of a reformulate the indirect Qualia unification troponym verb is inherited from its hypernym. clause, since inheritance prevents the shared Following the four-level semantic representa- arguments, recovered from the hypernym tion presented in the GL, we propose to consider lexical semantic structure, from being checked an Inheritance structure (corresponding for incompatibility: roughly to the Lexical Inheritance Struc- ture of Pustejovsky (1995)). The hypernym (29) Two co-troponym verbs are incompatible lexical items are referred to in this structure, iff the non-inherited arguments in their their semantic content becoming available, and argument structure refer to incompatible being only necessary to add the specific infor- co-hyponyms. mation that concerns the meaning specifications 5.2 Prepositional arguments of the troponym. In this way, we can represent the verb andar (walk), informally stated in (7), It is clear in Pustejovsky (1995) that the type or the verb sair (exit), as follows: of arguments (see section 3.1) in the argument structure also accounts for some level of syntactic mapping. Moreover, arguments are (28) a. andar(walk)     listed from the less oblique one (correspond-  mover     h i  ing, roughly, to the subject position) to the       ev-st= e1= 1 e1:act.  more oblique one (corresponding, most of the    h i     times, to arguments expressed by prepositional inh-st=hyp.=arg-st= arg = 2 x:ind     1  phrases) (Pustejovsky, 1995, 62-67). However,    " #      the information of which prepositional argu-    formal=   q=  ment is selected by which verb is not visible in  move act(e1,x)    the GL lexical entries.  h i    Observe, for instance, the sentence in (30): ev-st= e1= 1     " #   arg = 2  (30) Ele tirou o livro da mesa. arg-st= 1    ‘he took the book from-the table’  s-arg1=z:manner   h i  q= formal=walk act( 1 , 2 , z) The true arguments in the argument structure of this verb (arg-str=[arg1= x:ind, arg2= y:object, arg3= z:loc]) do not reveal b. sair(exit) that the arg3 is of the prepositional phrase     form, necessarily headed by the preposition de mover   (from), source marker. More still, it should be    h i     e = 1 e :act.  possible to associate this prepositional phrase –   ev-st= 1 1     h i also present in the argument structure of the     inh-st=hyp.=arg-st= arg1= 2 x:ind  verb sair in (28b)– to the resulting state e :     2    " #     formal=      (31) tirar (take)  q=     move act(e1,x)  ...    " #  " #     e = 1  formal=¬at(e2, y, z)  1  q=  ev-st=  ag.=take act(e , x, y, w)  e2=e2:state  1        arg1= 2      It can be assumed that in the cases of the arg-st= arg =y:location    2   verbs sair (exit) and tirar (take) the seman-    s-arg1=z:out  tic content attributed to the source denoting  " #  preposition de is not individuated since it is part  formal=¬at(e , 2 , y)  q= 2  of the semantic content of the verb itself. How- ag.=exit act( 1 , 2 , z) ever, although consisting of true arguments in the previous cases, these prepositional phrases can occur as adjuncts with other verbs – namely within a directional construction – maintaining This level of representation – the its meaning: (32) Ele empurrou a cadeira daqui. (34) sair(exit)  ‘he pushed the chair from-here’     mover     h i      Following the approach in Verspoor (1997),   ev-st= e1= 1 e1:act.     h i we assume that the semantic contribution of     inh-st=hyp.=arg-st= arg = 2 x:ind  prepositional phrases is consistent across uses,    1     " #  regardless of their status as complements or ad-    formal=      juncts9, and propose the following lexical entry  q=   move act(e1,x)  for the preposition de (from):    " #     e1= 1  (33) de (from) ev-st=   h i   e2= 3      ev-st= e1=e1:state    " #   arg1= 2          arg =x:obj      arg-st= 1    de        h i   arg2=y:loc        h i   ev-st= e1= 3 e1:state        q= formal=¬at(e1, x, y)    " #     arg = 2   arg-st=arg = 1     2 arg-st=      arg2=y:loc   The association of an event to the semantic    " #        content of a preposition is proposed in Puste-    formal=     q=   jovsky (1995, 126), for the representation of the   4 ¬at(e , 2 , y)    1  PP into the cave. The author considers that       the semantic content of the preposition into in-  s-arg1=z:out   " #  cludes, besides a state, a motion event that  formal= 4  causes that state. Here we feel that this as- q=  ag.=exit act( 1 , 2 , z) sociation is not motivated enough and that it is awkward to propose a transition representation for a preposition. Note, however, that the 6 WN and semantic types’ consequences of this view on the treatment of hierarchies directional constructions are yet to be studied. With this representation, it is possible to 6.1 Semantic types account for the semantic contribution of the One of the interesting consequences of consid- prepositional phrase in sentences such as the ering a relational model of the lexicon together one in (32), as well as in the semantic content of with generative representations is, perhaps, the verbs as sair (exit) or tirar (take), stating at the possibility of not having to build a semantic lexical level the exact prepositional complement type hierarchy to describe the lexical items se- selected by the verb: mantics, since it is possible to use the available lexical semantic structures in the hierarchy. Our proposal is to consider the nodes in the lexicon as the semantic types themselves. In fact, as a model of the mental lexicon, the WN can be seen as a reflex of the relation between lexical knowledge and our ability to orga- nize and classify the concepts in the world (cf. Pustejovsky (2001)). 9The analysis in Verspoor (1997) concerns mainly As a consequence we would have that, when dative prepositional phrases, divided in three types referring to an argument of type human, we of prepositional datives: complement PPs, pseudo- would access not the type and the relations complement PPs, and adjunct PPs. According to this account, the prepositional phrases considered in this established between this type and the other paper fall under the definitions of complement PPs or types in the hierarchy of types, but the synset pseudo-complement PPs. Note, however, that the se- {human} and the relations established be- mantic content of all the types of PPs is the same, between this node and all the other nodes in ing the type of prepositional modification determined via lexical rules. If this proposal accounts for the construc- the lexicon. Phenomena such as regular pol- tions in which these particular PPs enter is an issue for ysemy for instance, covered by complex types further development. such as information•physical object, are already accounted for by considering that synsets In this way, the model seems able to account like {book} have two hypernyms, {text} and for the syntactic behavior of a set of lexical {object}. items without resorting to external devices – Also, the hypernym relation enables lexical namely considering also function relations –, al- inheritance through the use of available lexical though further study on this issue is in order. structures. All that is required is that the semantic specification of the arguments and the 7 Conclusion values of the Qualia roles on a lexical entry do In this paper, we showed that the construction not refer to types pertaining to an extra-lexicon of a WN, adding semantic structure to the lexi- hierarchy, but to other lexical items. Note that cal entries, is a non-costly task that greatly en- lexical gaps – concepts that are not lexicalized hances the computational potentialities of the in a given language – might raise important is- model. sues to this proposal. Yet, it seems reasonable The use of the troponym relation to deter- to consider that the ontological and hierarchical mine both the core and specific meaning of nature of a hand-build relational lexicon should the lexical items, through the decompositional be enough to ensure the conceptual categoriza- analysis of this relation, identifies essential tion needed to describe the lexical items mean- elements of the semantic representation of ing. Nonetheless, we are naturally working to verbal predicates in the GL framework: the test further this hypothesis. argument and Qualia structure values. The 6.2 Shadow arguments different argument types in the argument structure account for different selection restrictions The ontological nature of a WN seems also of verbs, whereas the association of Qualia to shed light on lexical shadowing phenomena. information to the lexical structures and the Fellbaum (1998, 81) states that the variety of assumption of an indirect Qualia unification nouns that a verb can select as arguments is operation predict compatibility among co- contrary to the level in which it is embedded in troponym verbs. Also, the WN as base model the hierarchy, i.e., as the meaning of verbs gets for the lexicon allows for lexical inheritance more and more specific, their arguments are also without resorting to external resources. Finally, selected in lower levels of the hierarchy. Let us the ontological nature of this relational model consider the following hierarchy fragment: seems to allow for the replacement of semantic (35) {serrar}(saw) @-> {cortar}(cut) @-> types by referring to other lexical items in the {dividir}(separate) structure.

It is possible to verify that the lowest node This paper refers on-going work that we in- has more selection restrictions, namely, that it tend to continue, namely in what concerns fur- incorporates an instrument argument through ther exploring the verbs’ properties that ac- lexical shadowing, i.e. it incorporates an ar- count for differentiated syntactic behavior. gument that can only be syntactically realized through discourse specification devices (see 8 Acknowledgements Pustejovsky (1995)). Note, however, that ac- My thanks to Palmira Marrafa, Rui Chaves and cording to the hypothesis that there is a straight Sara Mendes for their comments, reviews and connection between what is semantically deter- suggestions. mined and what is syntactically realized, it is possible to account for the syntactic occurrence References restrictions of shadow arguments by using its E. Agirre and D. Martinez. 2002. Integrating hypernymy relations: shadow arguments can be selectional preferences in wordnet. In Pro- syntactically realized with their hyponyms. ceedings of the First International WordNet Conference. Mysore, India. (36) a. *The lumberjack sawed the trees with R. Amaro. 2005. Semantic incorporation in a a saw. portuguese wordnet of verbs of movement: b. The lumberjack sawed the trees with a on aktionsart shifting. In Third International chainsaw. Workshop on Generative Approaches to the Lexicon, pages 1–9. Ecole´ de Traduction et (37) {chainsaw} @-> {saw} d’Interpretation, University of Genéve. F. Busa, N. Calzolari, A. Lenci, and J. Puste- T. Móia.2000. Identifying and computing tem- jovsky. 2001. Building a semantic lexicon: poral locating adverbials: with a particular fo- Structuring and generating concepts. In H. cus on Portuguese and English. Ph.D Thesis, Bunt, R. Muskens, and E. Thijsse, editors, University of Lisbon. Computing Meaning, volume vol. II, pages T. O’Hara and J. Wiebe. 2003. Classifying 29–51. Dordrecht: Kluwer. functional relations in factotum via word- H. Dang, K. Kipper, and M. Palmer. 2000. In- net hypernym associations. In Proceedings tegrating compositional semantics into a verb of CICLing-2003. CIC, IPN, Mexico City, lexicon. In Proceedings of the 17th Confer- url://www.cs.nmsu.edu/ tomohara/ ence on Computational Linguistics, pages 1– ohara factotum Roles cicling03.pdf. 9. Saarbrücken, Germany. J. Pustejovsky. 1995. The Generative Lexicon. M. De Boni and S. Manandhar. 2002. Au- MA: The MIT Press. tomated discovery of telic relations for J. Pustejovsky. 2001. Type construction and wordnet. In Proceedings of the first In- the logic of concepts. In P. Bouillon and ternational WordNet conference. India, F. Busa, editors, The Syntax of Word Mean- url://www-users.cs.york.ac.uk/ ings. Cambridge University Press, url: mdeboni/publications.html. www.cs.brandeis.edu/ jamesp/articles/ C. Fabre and P. Sébillot. 1999. Semantic type-construction.pdf. interpretation of binominal sequences and L. Talmy. 1985. Lexicalization patterns: information retrieval. In Proceedings of Inter- semantic structure in lexical forms. In national ICSC Congress on Computational T. Shopen, editor, Language typology and Intelligence: Methods and Applications. syntactic description, volume vol. III, pages CIMA’99, Symposium on Advances in In- 57–149. Cambridge University Press. telligent Data Analysis,AIDA’99, New York, T. Veale. 2003. Qualia extraction and url://www.irisa.fr/texmex/Papers/1999/ creative metaphor in wordnet. In Pro- sebillotFS99.pdf. ceedings of the 3rd International Work- shop on Creative Systems. IJCAI 03, C. Fellbaum. 1998. A semantic network of eng- url://afflatus.ucd.ie/papers/ lish verbs. In C. Fellbaum, editor, WordNet: ijcai2003.pdf. an electronic lexical database, pages 69–104. MA: The MIT Press. Z. Vendler. 1967. Linguistics in Philosophy. New York: Cornell University Press. R. S. Jackendoff. 1983. Semantics and Cogni- C.M.˜ Verspoor. 1997. Contextually-Dependent tion. Cambridge: The MIT Press. Lexical Semantics. Ph.D Thesis, University J. Koenig, G. Mauner, and B. Bienvenu. 2002. of Edinburgh. Class specificity and the lexical encoding of participant information. Brain & Language, 81:224–235. P. Marrafa and Moura. 2005. Portuguese telic causative verbs. In Proceedings of the Third International Workshop on Generative Ap- proaches to the Lexicon, pages 144–148. Ecole´ de Traduction et d’Interpretation, University of Genéve. P. Marrafa. 2002. The Portuguese WordNet: General Architecture and Semantic Internal Relations. DELTA. S. Mendes and R. P. Chaves. 2001. Enriching wordnet with qualia information. In Proceed- ings of the Workshop on WordNets and Other Lexical Resources, pages 108–112. NAACL 2001 Conference, Pittsburgh. M. Moens. 1987. Tense, Aspect and Temporal Reference. Ph.D Thesis, University of Edin- burgh.