File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/92/c92-1018_metho.xml
Size: 18,745 bytes
Last Modified: 2025-10-06 14:12:55
<?xml version="1.0" standalone="yes"?> <Paper uid="C92-1018"> <Title>A Lexicalist Account of Icelandic Case Marking</Title> <Section position="3" start_page="0" end_page="0" type="metho"> <SectionTitle> 1 Introduction: Case Marking </SectionTitle> <Paragraph position="0"/> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> and Defaults </SectionTitle> <Paragraph position="0"> The Icelandic case marking system I has often been taken as evidence for a distinction between structural (or regular or default) and lexicai (irregular, quirky) case marking. This distinction is introduced to explain the following two facts. First, most verbs select nominative case marked subjects (1), but a number of verbs select accusative, genitive or dative subjects (2)(4). Similarly, most transitive or ditransitive verbs select an accusative direct object (1), but some select a nominative, dative, or genitive (4) null (6) object.</Paragraph> <Paragraph position="1"> (1) Stfilkan(N) kyssti drengina(A) the-girls kissed the-boys 1See Andrews (1982) for an extensive overview. (2) Drengina(A) vantar mat(A) the-boys lacks food (3) Verkjanna(G) g~etir ekki the-pain is-noticeable not (4) Barninu(D) batnac~i veikin(N) the-child ~,eeovered-from the-disease (5) 13g hjPS1paSi honum(D) I helped him (6) Eg mun sakna hans(G) I will miss him The fact that the vast majority of subjects and objects is nominative and accusative, respectively, is accounted for by assigning default status to these cases (that is, if nothing is said about the case of a subject or object, assume it must be nominative or accusative, respectively). Second, passive sentences (7)-(8) and so-called 'raising' constructions (9)-(12) exhibit a distinction between regular and irregular case marked NPs. In transformational terms, the case of regular case marked NPs corresponds with their surface position ((7), (9), and (11)), whereas the case of irregular case marked NPs corresponds to their underlying position ((8), (10), and (12)). (In these examples, tc represents a lexicaily case marked NP-trace and t a trace which has not been assigned case (assunfing that structural case is only assigned to surface NP positions)).</Paragraph> <Paragraph position="2"> (7) Drengirnir(N) voru kysstir t the-boys were kissed (8) IIonum(D) var hjPS1pa5 tD him was helped (9) lIann(N) virSist It elska hana\] he seems to-love her (10) liana(A) virSist IrA vanta peninga\] her seems to-lack money Preservation of irregular case is particularly striking is a sentence such as (13). Although, normally, a verb which selects an accusative object Mways selects a nominative subject when passivized, this is not so in (13), where a raising to object verb has been passived. The lexicai case assigned by the embedded verb vanta is preserved even if raising to object and subsequent passivization takes place.</Paragraph> <Paragraph position="3"> (13~ig(h) er tMi3 tA \[tA vant;~ peninga\] I is believed to-lack money The interaction of case assignment, raising, and passivization is summarized in the table below (where ~ and \[ refer to the underlying and surface NP position, respectively). The distinction between structural and lexical accusative in passives is needed to account for a sentence suetl as (13). Transitive verbs with nominative objects do not passivize (Yip et al., t987, p. 225). All previous accounts of these facts ~ have adopted the following two assmnptions: * Lexical case is assigned in the lexicon (that is, before the lexical and syntactic l)rocesses responsible for passive formatiml and raising have appfied) attd structural case is assigned in syntax (after idl other syntactic processing has been completed).</Paragraph> <Paragraph position="4"> * Structural case assignment is a default process which can never override lexical case assignments.</Paragraph> <Paragraph position="5"> 2Except Sag et at. (to appear), whose analysis is in many respects similar to the one developed below. Andrews' (1982, p. 485) LFG-based account, for instance, assumes two default conventions, applical)le to functional structures (that is, after syntactic processing has been cmnI)leted). First, if the subject is unmarked for case, (direct and indirect) objects are accusative by default (i.e. only if they are not lexically ease marked), and, second, an argmnent must be nominative if no other ease constraint has been imposed upon it. The application of the first convention nmst precede that of the second. Zaenen et al. (1985, p. 466) adopt a default case marking convention which appears to be restricted to the lexicon and which assigns nominative case to the highest available grammaticM function (i.e. a grammatical function which is present in the argument (or thematic) structure of a verb attd which has not been assigned lexical case), and accusative to the next highest available grammatical fnnc tiou. Zaenen et al. concentrate primarily on passives and do not present an explicit account t)f raising constructions. Yip ctal. (1987), inspire(\[ by i(\[eas froln autosegmental phonology, propose a system in which a case tier&quot; \[NOM ACC\] is associated with verbal arguments in syntatx in a left-to-right fashion, in such a way that NOM is associated with the leftmost argument titat has not been assigned lexical case and hcc is associated with the next available argument. Their account of raising to object assumes that case association is a cyclic rule and that syntactic (as opposed to lexieal) case assignments may be overwritten by syntactic case assignments in higher cycles.</Paragraph> <Paragraph position="6"> These proposals are prol)lematic for standard unification-based formMisms of the PATR-II variety (Shieber, 1986a). The default constraints for nominative and accusative case ntarking must not apply if :t lexical case constraint is present already. This ilnlflies that tit(., default constraints cannot be added to the syntactic rules or lexical entries which introduce subjects and objects, as this would make the constraints absolute. Rather, syntactic default case marking principles would have to be added as a filter on syntactic structures or as seperate default rules (much like the syntactic fi.~ature specification defaults of GPSG) whose application must be intertwined with syntactic processing a. Admitting such extensions, however, is potentially damaging to the descriptive adequacy of 3Note that the defaults cannot be 'compiled away' in this case, as is proposed for the feature specification defaults of GPS( / in Shieber (1986b).</Paragraph> <Paragraph position="7"> Acrv.s DE COLING-92, NA~rrEs, 23-28 Aotrr 1992 9 5 PROC. OF COL1NG-92, NArCrES, AUO. 23-28, 1992 unification-based formalisms, as the. addition of syntactic tilters a<hls a powerful and seemingly unrestricted level of <lescription to the grammar formalism arrd the a<ldition of syntlt<:tic defaults iirtroduces a procedural aspect to syntactic descriptions wbich is absent in tire standard forrealism. null The proposals of Andrews (1982) and Yip ~:t al. (1987) are also a challenge for lexicalist ttleories such as I\[I'SG and CUC,. The probleni for lexicatist theories is that the default principles above must make use of information which is only available a.fler a certain amount of syn |actic processing has taken place. In raising to object constructions, for instance, the questi<m whether the object must be assigned accusative case or not Call olrly be answered after it is known wbethcr the sul>ject of tire embedded VP is assigned lexical case or not. These con siderations have even been used (in ,lacobson, 1990) as arguments against lexicalist accounts of case marking in generM.</Paragraph> <Paragraph position="8"> Below, I will demonstrate that the Icelandic data can ire a.ecounted for declaratively alid in a lexicalist fashion. Ill the next section, I in: troduce a fragment of (;Ill\] including raising verbs and passives. Next, the distinction betweet |lexicM and structural case is accounted for by assuming that case is encoded as a featurecolnplex and by introducing a system for non-monotonic iiflmritanee of lexical information. In section 4 I show that this deeomlrosed case system makes it possible to analyze the case patterns found in passive and raising constructions as a reflection of a partial agreement relation. The final section romllares the present proposal with recenl, work by Sag ctal. (to appear).</Paragraph> </Section> </Section> <Section position="4" start_page="0" end_page="0" type="metho"> <SectionTitle> 2 Categorial Unification Gram- </SectionTitle> <Paragraph position="0"> mar In this section, I will l)r'esent a brief outline of Categorial Unification Gralnmar ((~U(',), a unitication-based version of Categorial Grammar (see, ainong others, Uszkoreit (f986), Karttunen (1989), Bouma (1988), alld Zeevat (1988), for details). In particular, the treatment of raising verbs and passives will be adressed.</Paragraph> <Paragraph position="1"> The categories of categoriM grammar can be encoded as feature-structures using tire feature CAT(EGORY) to represent basic categories and the Datures VaL(UE), ARG(UMENT) and DIR(E(:TIONALITY) to represent complex care-AcrEs DE COLlNG-92, NAbrI'ES, 23-28 Ao\[rr 1992 gories. The category NP\,S', for instance, is translated as: <1,,/ira, {c t s 1 / dir left |arg \[cat np\]</Paragraph> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> Lcat - </SectionTitle> <Paragraph position="0"> The specification \[CA\]' -\] is needed only to make complex categories non-unifilthle with basic categories. \[ will omit this specification in tile examples below. The template system ill (15) <lefines the basic inwmtory of categories that is itssumed in the sequel 4.</Paragraph> <Paragraph position="2"> t assume a categorial grammar which provides (at least) the following combinatory rules: Right Application: ofl.,X ~, t~,X/Y fl, Y Left Application: ties, X => fl, Y ~, Y\X Right Wrap: cqfl(~,X => c*la2, X/Y fl, Y These rules can ire thought of as licencing certain operations on strings arid feature structures, q'his is illustrated for leftward application in (16)below.</Paragraph> <Paragraph position="3"> (16) Arldy dreams, \[~ \[cat s\] All eombinatory rules unify the value of the \[unctor daughter with the feature structure of the mother aild unify the argument of the funetor with the argument daughter.</Paragraph> <Paragraph position="4"> Raising verbs subcategorize for a VP-complement (i.e. a category NP\,5 ~) the subject of which is controlled by an NP-argument of the raising verb. Following Bach (1979) I assume that the controller is always the so-called ti.extargur~er~t-itl. That is, a 'raising to subject' verb such as seem receives the category (NI>\,5)/VP,</Paragraph> <Paragraph position="6"> whereas a 'raising to object' verb such ,as cxpect is categorized ms (VP/NP)/VP. This implies that such verbs combine with their object using right wrapS: (17) expects Bob to go, VP expects to go, VP/ NP Bob, NP expects, (VP/NP)/VP to go, VP The semantic and syntactic implications of the relation between the controlling NP and the sub-ject of the VP-complement can be implemented by making these two arguments reentrant. I will ignore tits aspect of the analysis below (but see Pollard &, Sag (to appear) for an aimlysis which makes a similar assumption).</Paragraph> <Paragraph position="7"> Passive auxiliaries subcategorize for a passive TVP (i.e. a category VP/NP with passive morphology) and a subject NP (that is, I adopt a lexicalized version of the analysis in Bach (1980)). The fact that the subject is interpreted as the object of the '\['VP cat) again be implemented by making these two argument positions reentrant. Since raising to object verbs are functors which are reducible to TVI', this analysis predicts that such verbs may be passivized: null a feature complex. 'l?bat is, instead of using a single fi~atnre CASE, i will use three distinct features. The feature STItU(:T distinguishes between structural and lexical case, and the features Lt;X and NOM distinguish between the various lexical cases and between nominative and accusative structurM case, respectively. The data in section 1 show that genitive and dative case are always lexical (i.e. non-structural), that accusative may be lexical (if assigned to subjects) as well as structural (if assigned to olljects), and that nominative is always structurM. Therefore, I assume that the templates NOM, Acc, I)AT, and (JEN assign the values specified in table 2 to tim features lnentioned M)ove. The fact that A(:c is unspecilied for STRUCT i8 important in the account of case preservation presented below. Note that these four template definitions denote incompatible feature structures and thus can be used to defiim the morphologi- null cal case of NPs in lexicM entries: (18) w~ expected to go, VP was, VP/TVP expected to go, 'I'VP 3 Distinguishing Lexical and</Paragraph> </Section> <Section position="2" start_page="0" end_page="0" type="sub_section"> <SectionTitle> Structural Case </SectionTitle> <Paragraph position="0"> The most important tool for expressing linguistic generalizations in unification-based forntafisms is the template meclmnism. In this section, I iutroduce the template system for case and show how nonmonotonic template inheritance call be used to account for the default character of structurM case.</Paragraph> <Paragraph position="1"> The distinction between lexical and structural application or wrap). It is possible to add a feature which encodes this information explicitly.</Paragraph> <Paragraph position="2"> Note that tile fact that accusative acts a,s a structural case for objects has been added explicitly. null Both structural and lexical case assigning verbs are defined in terms oftlm templates ill (20). For structurM case assigning verbs only their category needs to be specified. The detinitions of lexicM case assigning verbs, on the other hand, will contain explicit case constraints. To avoid feature-clashes, I assume that these irregular verbs inherit nonmonotonically from the tern- null 9 7 PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 plates above: (21) kyssti :(TV ).</Paragraph> <Paragraph position="4"> In these definitions, the '!'-prefix (comparable to the overwrite operator of PATR-H) indicates that the following constrMnt contains non-default information which may surpress the inheritance of incompatible information from other sources (i.e. the template TV). Thus, hjalp~i denotes the feature struture (where IVis left unexpanded): null dir right</Paragraph> <Paragraph position="6"> Note also that vantar assigns \]exicM (i.e. nonstructural) accusative case to its subject.</Paragraph> <Paragraph position="7"> The definitions above capture the intuitive difference between structural and lexical case.</Paragraph> <Paragraph position="8"> If a verb assigns structural, default case, only its category needs to be specified, the relevant case constraints follow by inheritance. If a verb assigns lexical case, however, case must be specified explicitly. Although tids analysis introduces a limited form of nonmonotonicity, this does not endanger the declarative nature of the grammar.</Paragraph> <Paragraph position="9"> The definitions in (21) still denote feature structures~ and thus, the effect of combining default and non-default information can be computed at compile time. That is, there are no formal difl'erences between a grammar using nonmonotonic inheritance and agrammar which does not. Furthermore, in Bouma (to appear) it is demom strated that the inheritance operation itself can be defined declaratively. What is gained by using nonmonotouic inheritance is the fact that it supports a direct and natural implementation of certain linguistic observations.</Paragraph> </Section> </Section> <Section position="5" start_page="0" end_page="0" type="metho"> <SectionTitle> 4 Case Preservation </SectionTitle> <Paragraph position="0"> In this section I argue that the case preservation pheuomena which can be observed in raising and passive constructions are a reflection of a (partial) case agreement relation between two NP-argument positions.</Paragraph> <Paragraph position="1"> Ac'r~ DE COLING-92, NANTES, 23-28 aotrr 1992 98 ~br raising to subject verbs, for instance, the case of controlling subject is Mways identical to the subject case constraint specified by the embedded VP (that is, I assume that infinitival verbs may specify structural, nominative, case for their subject). Thus, raising to subject verbs can be defined as in (23). The paths (val arg) and (arg arg) refer to the controlling and controlled NP, respectively.</Paragraph> <Paragraph position="2"> (23) R-to-S : (xv/yp (val arg struct) = (a,:q arg struct) (,,.1 ..j lex) : (.,g .,9 l~x) (..t .,v nora) = (.-j &quot;'V ,,ore) ).</Paragraph> <Paragraph position="3"> The feature structures denoted by this template and by the templates defined below are shown in figure 1.</Paragraph> <Paragraph position="4"> Raising to object verbs, on the other hand, do not take nominative objects and thus require that their object is \[NOM -\]. The other case features again agree:</Paragraph> <Paragraph position="6"> (val arg struct) = ( an2 arg struct) (val a~ lex) : ( ar 9 arg lex) ).</Paragraph> <Paragraph position="7"> Note that if the VP-complement assigns nominative case to its subject argmnent, the object is specified as \[STItUCT +, NOM -\], This impfies that the object must be accusative, as only the case template Ace will unify with this specification. In all other cases, the lexical case assigned by the Vl'-complement will be preserved. In particular, a lexical accusative subject gives rise to the following derivation: Passive auxiliaries, finally, are detined as follows: null</Paragraph> <Paragraph position="9"> In passives, the question whether a nomina~ tire subject will appear or not is determined by the question whether the object of tim TVPcomplement receives structural case or not. If the object was marked for one of the lexical cases GEN or I)AT, this case will also appear on the subject. A structural accusative object gives rise to a \[NOM -~-, LEX ACC\] specitication for the subject, which is unitiable with the case template NOM only. A lexieai accusative object (as in (25)), on tit(.' other hand, gives rise to a \[NOM -,LEX ACC\] sul)ject, a specilicati(m which is only unifiable witil Acc. Note that this accounts for example (13), which showed that lexical accusative case imlst be preserved in passive constructions.</Paragraph> </Section> class="xml-element"></Paper>