keenan com rie

8/13/2019 Keenan Com Rie

1/38

Noun Phrase Accessibility and Universal GrammarAuthor(s): Edward L. Keenan and Bernard ComrieSource: Linguistic Inquiry, Vol. 8, No. 1 (Winter, 1977), pp. 63-99Published by: The MIT PressStable URL: http://www.jstor.org/stable/4177973

Accessed: 22/12/2009 12:04

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at

http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless

you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you

may use content in the JSTOR archive only for your personal, non-commercial use.

Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at

http://www.jstor.org/action/showPublisher?publisherCode=mitpress.

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed

page of such transmission.

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of

content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms

of scholarship. For more information about JSTOR, please contact [email protected].

The MIT Pressis collaborating with JSTOR to digitize, preserve and extend access toLinguistic Inquiry.

http://www.jstor.org
http://www.jstor.org/stable/4177973?origin=JSTOR-pdfhttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/action/showPublisher?publisherCode=mitpresshttp://www.jstor.org/action/showPublisher?publisherCode=mitpresshttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/stable/4177973?origin=JSTOR-pdf


2/38

Linguistic Inquiry Volume 8 Number I (Winter, 1977) 63-99.

Noun PhraseAccessibilityandEdward L. KeenanBernard Comrie Universal Grammar*

In section 1 we present the Accessibility Hierarchy, in terms of which we state threeuniversalconstraintson Relative Clause Formation. In addition, we present the data insupport of these constraints and discuss certain partialcounterexamples. In section 2we proposea partialexplanation or the hierarchyconstraintsand present furtherdatafromRelative Clause Formation upportinghese explanations.Finally, in section 3 werefer briefly to other work suggesting that the distributionof advancementprocessessuch as Passive can be described in terms of the Accessibility Hierarchy; hese factsshow that the proposedexplanation or the Hierarchyneeds to be generalized

1. The Accessibility Hierarchy1.1. Two Methodological PreliminariesWe are attemptingto determine the universalpropertiesof relative clauses (RCs) bycomparingtheir syntactic form in a large number of languages. To do this it isnecessaryto have a largely syntax-freeway of identifyingRCs in an arbitrary anguage.Oursolutionto this problem s to use an essentially semanticallybased definitionof RC. We consider any syntactic object to be an RC if it specifies a set of objects(perhapsa one-memberset) in two steps: a largerset is specified, called the domain ofrelativization, and then restricted to some subset of which a certain sentence, the

* This article is a developed version of Keenan and Comrie (1972). We attemptin this version toaccount for several objectionsto the earlier ormulation hat have since been brought o ourattention.We would like to acknowledge he followingsources for significanthelp withthis article:

(i) One LinguisticInquiryreviewer or constructivecriticismsof an earlierdraft;(ii) Support rom an NSF postdoctoralresearchgrant andtwo Wenner-Gren rants,2384 and2944,for work on the Malayo-Polynesian anguagesreportedherein;for supportof the psycholinguisticworkonthe Accessibility Hierarchywe are indebted o a grantfrom the NuffieldFoundation;(iii) The willing and substantive help from the following linguistsconcerning languages they haveworked with: G. Brettschneider Basque), P. Brown (Tzeltal), M. Butar-ButarToba Batak), S. Chung(Indonesian,Tongan), 0. Dahl and J. Allwood (Swedish), K. Ebert (North Frisian, Kera), J. Hawkins(German,Slovenian),A. Janhunen Finnish), M. Perera Catalan),H. van RiemsdijkZurichGerman), . deRooij (Dutch), A. Salmond (Maori), P. Sgall (Czech), R. Tanaka(Japanese),B. Vattuone (Genoese), N.Vincent(Italian);(iv) The willing and substantivehelp of our informants.


3/38

64 EDWARD L. KEENAN AND BERNARD COMRIE

restricting sentence, is true.' The domain of relativizationis expressed in surfacestructureby the head NP, and the restrictingsentence by the restrictingclause, whichmay look more or less like a surfacesentence dependingon the language.For example, in the relative clause the girl (that) John likes the domain ofrelativization s the set of girls and the head NP is girl. The restricting entence is Johnlikes her and the restricting clause is (that) John likes. Clearly, for an object to becorrectly referred to by the girl that John likes, the object must be in the domainofrelativizationand the restricting entence mustbe true of it. We shall refer to the NP inthe restrictingsentence that is coreferentialwith the head NP as the NP relativizedon

(NPrei); in our example, this is her, i.e. the direct object of John likes her.Note that we only consider definiterestrictive RCs in this study. The role of thedeterminer he is held constant and ignored,and the termRC is used to apply to thecollocation of the head NP and the restrictingclause.Note further hat our semanticallybased notion of RC justifies consideringas RCscertain constructions that would perhapsnot have been so considered in traditionalgrammar. Thus, in German, alongside the traditional RC in (1) we also count theparticipialconstruction n (2):(1) der Mann,der in seinemBuro arbeitetthe man who in his studyworks'the man who is working n his study'(2) der in seinem Buro arbeitendeMann

the in his study working man'the man who is working n his study'As the German data above illustrate, not only do differentlanguages vary withrespect to the way RCs are formed, but also within a given language here is often morethan one distinct type of RC. We shall refer to distinct ways of forming RCs asdifferentrelative clause forming strategies. Different strategies differ with regard towhich NP positions they can relativize. Thus, the participialstrategyin (2) above canonly relativizesubjects (thatis, the head NP can only be understood o function as thesubject of the main verb of the restrictingclause), whereas the strategy in (1) above

functions to relativize almost any major NP position in simplex sentences. Conse-quently, generalizationsconcerningthe relativizabilityof differentNPs must be madedependenton the strategiesused. It will be criticaltherefore o providesome principledbasis for decidingwhen two different RCs have been formed with different strategies.There are many ways RCs differ at the surface, and hence manypossible criteriafor determiningwhen two strategies are different. We have chosen two criteria thatseem to us most directlyrelated to our perception of how we understand he meaningof the RC-that is, of how we understandwhat properties an object must have to be1 A more formal statement of this semantic notion in which the logical structure of an RC is representedas a pair, a common noun phrase and an open formula, can be found in Keenan (1972a).


4/38

NOUN PHRASE ACCESSIBILITY AND UNIVERSAL GRAMMAR

correctly referredto by the RC. The first concerns the way the head NP and therestricting clause are distinguishedat the surface, and the second concerns how theposition relativizedis indicated.

In the first case we consider two RCs to be formedby differentstrategiesif therelative position of the head NP and the restrictingclause differs. There are threepossibilities: the head occurs to the left of the restricting clause, as in (1) above(postnominal RC strategy); the head occurs to the right, as in (2) (prenominal RCstrategy);or the head occurs within the restrictingclause (internal RC strategy),as in(3) and(4), fromBambara Bird (1966))and Diguefio(Gorbet (1972)), respectively:

(3) a. ne ye so ye.I Past horse see'I saw a horse.'b. ne ye so min yeI Past horse which see'the horse that I saw'c. tye ye ne ye so min ye san.manPast I Past horse which see buy'The manbought the horse that I saw.'

(4) a. tOnay ?awa:+ 0 ?awu:w.yesterdayhouse DO I-saw'I saw the house yesterday.'b. ?owa: + pu + LI ?ciyawx.house Def in I-will-sing'I will sing in the house.'C. [NP[S tanay ?;wa: + 0 ?3wu :w]] + pu + LU?ciyawx.yesterdayhouse DO I-saw Def in I-will-sing'I will sing in the house that I saw yesterday.'

In the second case we consider two RCs to be formed from different strategies ifone presents a nominalelement in the restrictingclause that unequivocally expresseswhich NP position is being relativized, and thus we know exactly what the restrictingclause is saying about the head NP (that is, we can recover the restrictingsentencefrom surface) (+case RC strategy). For example, the English strategythat formsthegirl who John likes is not case-coding since who, the only relevant particle in therestrictingclause, can be used as well if the role of the head NP in the restricting lauseis different, e.g. the girl who likes John (-case RC strategy). On the other hand, incomparablesentences in Russian, (5a) and (5b), the form of the relative pronoundoesunequivocally tell us the role of the head NP, so that strategy in Russian is case-coding:(5) a. devuska, kotoruju Dzon Ijubitgirl who (accusative)John likes'the girlwho John likes'


5/38


b. devuska, kotoraja ijubitDzonagirl who (nominative) ikes John'the girlwho likes John'Note, however, that RCs in English like the chest in which John put the money areconsideredcase-coding, since the preposition n, which indicatesthe role of the headNP, is present in the restrictingclause.In additionto the use of relativepronouns,case can be coded in anotherway inthe languagescovered in our study. Namely, a personal pronouncan be presentin theNP position relativized, as in Hebrew:(6) ha-isha she- Yon natan a et ha- seferthe womanthat John gave to-her DO the book'the woman that John gave the book to'

1.2. The Accessibility Hierarchy and the Hierarchy Constraints1.2.1. Statement of the Hierarchy and the Constraints. On the basis of data fromabout fifty languages, we argue thatlanguagesvary with respect to which NP positionscan be relativized, and that the variation s not random. Rather,the relativizabilityofcertainpositions is dependenton thatof others, and these dependencies are, we claim,universal.The Accessibility Hierarchy AH) below expresses the relative accessibilityto relativizationof NP positions in simplex main clauses.

Accessibility Hierarchy (AH)SU > DO >LO > OBL > GEN > OCOMPHere, > means 'is more accessible than'; SU stands for 'subject', DO for 'directobject', 10 for 'indirectobject', OBLfor 'majorobliquecase NP' (we intend here NPsthat express arguments of the main predicate, as the chest in John put the money in thechest rather than ones having a more adverbial unctionlike Chicago in John lives inChicago or that day in John left on that day), GEN stands for 'genitive' (or'possessor') NP, e.g. the man in John took the man's hat, and OCOMP stands for'object of comparison', e.g. the man in John is taller than the man.The positions on the AH are to be understoodas specifying a set of possiblegrammaticaldistinctions that a languagemay make. We are not claimingthat any givenlanguagenecessarily distinguishesall these categories, either in terms of RC formationor in terms of other syntactic processes. For example, some languages (e.g. Hindi)treat objects of comparison like ordinaryobjects of prepositions or postpositions. Insuchcases we treat these NPs as ordinaryOBLs, and the OCOMPpositionon the AHis unrealized. Similarly, in Gary and Keenan (1976) it is argued that the DO and 10positions are not syntactically distinguished in Kinyarwanda, a Bantu language.Further,it is possible that in some language RC formationmightdistinguishbetween


6/38


two types of DOs. If this were so, we would have to expandthe AH at thatpoint andsay that languages ike Englishdo not make the distinction. For the moment,however,we take the AH as specifying the set of possible grammaticaldistinctionsto which RCformation(fromsimplexmainclauses) may be sensitive, since our data do not appearto justify any furtherrefinement n the categories.In terms of the AH we now give the HierarchyConstraints HCs):

The Hierarchy Constraints (HCs)1. A languagemust be able to relativizesubjects.2. Any RC-forming trategymustapplyto a continuoussegmentof the AH.3. Strategiesthatapplyat one pointof the AH may in principlecease to applyat

any lowerpoint.The HCs define conditions that any grammarof a humanlanguagemust meet. HC1says that the grammarmust be designed to allow relativization on subjects, theuppermostend of the AH. Thus, for example,no languagecan relativizeonly DOs, oronly locatives. It is possible, however, for a language to allow relativizationonly onsubjects(and this possibilityis in fact realized;see 1.3.1for examples).HC2statesthat,as far as relativization s concerned,a language s free to treatadjacentpositionson theAH as the same, but it cannot skip positions. Thus, if a given strategy can apply toboth subjects and locatives, it can also apply to DOs and 1Os. And HC3 states thateach pointof the AH is a possible cut-offpointfor any strategythatappliesto a higherpoint. This means that in designingthe grammar or a possible humanlanguage, oncewe have given it a strategy that applies at some point on the AH, we are free toterminate ts applicationat any lower point.Note that it is HC2that justifies the actual orderingof terms in the AH. Further,HC2allows as a special case that a particularRC-forming trategy mayapply to only asingle position. Thus several languages e.g. Javanese (see Table 1)) have recourse to acase-codingstrategyfor positions low on the AH (e.g. genitives), whereas the strategyfor major NPs is not case-coding.On the other hand, HC3 states thateach pointon theAH is relevant.Thus, if no languagecouldhave an RC-forming trategythatappliedtoDOs but not to 1Os, then the datawould notjustify making his distinction n the AH.In section 1.3 we substantiate he claim that each pointon the AH is in fact a possiblecut-off point by showing that for each point on the AH some language has a strategythatcuts off at thatpoint.1.2.2. The Primary Relativization Constraint. In 1.2.1 we stated that the AHdetermines,universally, the degree of accessibilityto RC formation.But it is perhapsnot obvious that the HCs actuallymake that intuitionexplicit. However, the PrimaryRelativizationConstraint(PRC) below does make that intuitionmore explicit, and itlogicallyfollows from the HCs. To state it we first define an RC-forming trategyin a


7/38


8/38


(8c):(8) a. *abit na manussiboru-boruclothes thatwash woman the

'the clothes that the womanis washing'b. disussi ni boru-boru bit i.wash(passive)by woman clothes the'The clothes were washedby the woman.'c. abit na nisussi ni boru-boruclothes thatwashedby woman the'the clothes thatwere washedby the woman'(Note that the passive prefixdi- becomes ni- in subordinateposition.)Thus Toba, like many Malayo-Polynesian anguages (Keenan (1972b)), cannotdirectly relativize NPs thatcan be promoted o subject. It mustpromotethemandthenrelativize them as subjects. However, in distinction to Philippine languages andMalagasy, for instance, Toba has a rather imitedpromotion( voicing ) system. NPsgovernedby prepositions, ncluding1Os,cannot be systematicallypromoted o subject.To relativizethem, a second strategy, this time +case, can be used. The relativizationmarker s different, and a personalpronoun s retained n the position relativized:

(9) dakdanak, ima- na nipaboani si Rotuaturi-turian-i tu ibanachild the namelythat told by Art Rotuastory the to him'the child that Rotuatold the story to'Clearly, then, Toba has gapped RC-formingstrategies. One applies only tosubjects, the otherto 1Os, OBLs, and GENS. The NPs in the gap , in this case theDO, must be promoted o subject to be relativized.All the languages n oursample thatpresent gapped strategies(see Table 1) are similarto Toba in this respect. That is,unrelativizableNPs can be systematicallypromoted to higherpositions on the AH,whence they can be relativized. The fact that NPs lying in strategygaps can alwaysbepromoted to accessible positions does justify the following form of the strongconstraint: If a languagecan relativizeany position on the AH, then it can relativizeany higherposition either directly or by promoting t to a position that can itself berelativizeddirectly.We considernow, in section 1.3, the data that support the HierarchyConstraints.In particular,we show that each point on the AH represents a cut-offpoint for someprimarystrategy. Then, in 1.4, we considersome of the problematiccases andpossiblecounterexamples.

1.3. Justification of the Hierarchy Constraints1.3.1. Subjects Only. In many WesternMalayo-Polynesiananguages,only subjectscanbe relativized;consider for instanceMalagasy,whichhas basicwordorder V 0 X Sand a developed system for promotingany major NP to subjectposition. RCs place


9/38


10/38


examples:(12) a. Poydallatanssinut poikaoli sairas.on-table having-dancedboy was sick

'The boy who had danced on the table was sick.'b. Nakemani poika tanssi poydalla.I-having-seenboy danced on-table'The boy that I saw dancedon the table.'As a final case in this category, consider Malay. The primary strategy usespostnominalrestricting clauses introducedby the invariableparticleyang with nopronounretained n the position relativized:2(13) Ali bunohayam yang Aminahsedang memakan.

Ali kill chicken that AminahProg eat'Ali killed the chicken that Aminah s eating.'But this strategyapplies only to subjectsanddirectobjects. Thus from(14a) we cannotform(14b) or (14c), either strandingor moving the preposition:

(14) a. Ali beri ubi kentang tu kapadaperempuantu.Ali give potato the to woman the'Ali gave the potato to the woman.'b. *perempuan ang Ali beri ubikentang tu kapadawoman that Ali give potato the toc. *perempuankapadayang Ali beri ubikentang tuwoman to who Ali give potato thatThe only systematicallyelicitablepossibility was (15):

(15) perempuankapadasiapaAli beri ubi kentang tuwoman to who Ali give potato the'the woman to whom Ali gave the potato'Here, the interrogativepronoun iapa precededby the preposition s used to introducethe relative clause. Clearly, the use of the interrogativepronoun, which can takeprepositions(i.e. +case), is a different strategyfrom that using the invariableparticleyang. In fact, this use of the interrogativepronoun was elicited only underpressure,andwas felt to be bookishandclumsy. The preferredalternativewas (16), in which theoriginalsentence has been reorganized exically so that the position relativizedis thesubject:

(16) perempuanyang menerimaubikentang tu daripadaAliwoman that received potato the from Ali'the womanthatreceived the potato from Ali'2 Our two informantswere from Malaya. MacDonaldand Soenjono (1967)present a slightlymorerestricted trategy or Indonesian.


11/38


1.3.3. Subject-Indirect Object. The indirect object position is perhaps the mostsubtle one on the AH. For purposesof relativeclause formation, t appearsthatmanylanguageseither assimilate indirectobjects to the other oblique cases (e.g. English,Malay)or to directobjects (e.g. Shona, Luganda).Nonetheless, Basquedoes appear odiscriminate ndirectobjects fromboth its immediateneighborson the AH.In Basque, subject,direct object, andindirectobjectarerepresented n the verbbyverbal affixes (Lafitte (1962, 193-194)). Relativization on any of these positions iseffected naturallyby deletingthe position relativized, puttingthe restrictingclause inprenominalposition, and markingthe juncture with an invariablemarker -n. Thus,from(17a)we can generateany of the three relative clauses (17b)-(17d):

(17) a. Gizon-a -k emakume-a -ri liburu-a eman dio.man the SU woman the 10 book the give has'The man has given the book to the woman.'b. emakume-a -ri liburu-a eman dio-n gizon-awoman the 10 book the give has Rel man the'the man who has given the book to the woman'c. gizon-a -k emakume-a -ri eman dio-n liburu-aman the SU woman the 10 give has Rel book the'the book that the manhas given to the woman'd. gizon-a -k liburu-a eman dio-n emakume-aman the SU book the give has Rel woman the'the womanthat the manhas given the book to'

However, once we attempt to relativize on positions that are not explicitly codedin the verb, a variety of difficultiesarise. For speakers of what de Rijk(1972)has calledthe restricted dialect, no further relativization is possible. For other speakers asomewhatgreater varietyof positions may be relativized, but often a differentstrategyis used. For instance, a pronoun may be retained in the position relativized and therelative clause may occur postnominally.Consequently, for speakersof the restricteddialect, as well as for certain other speakers, the primarystrategyworks only on thesubject, direct object, and indirectobject.Anotherlanguage n this categoryis Tamil, a Dravidian anguageof southern Indiaand Sri Lanka (Ceylon). One RC-forming strategy puts the restricting clause inprenominalposition, with the participial (nonfinite) ending -a on its verb, and noindication in the restricting clause of the syntactic function of NPrei. This strategyapplies to subjects, directobjects, and indirect objects:

(18) a. Jan patu-kir -a penmani(y)-ai kan-t -an.Johnsing Pres Partwoman DO see Past Sg-3rd-Masc'Johnsaw the womanwho is singing.'


12/38


b. Anta manitanati-tt -a penmani(y)-ai jan kan-t -an.that man hit Past Partwoman DO John see Past Sg-3rd-Masc'John saw the woman that that manhit.'c. Jan puttakatt-ai(k) oti-tt -a penmani(y)-ai nankan-t -en.John book DO give Past Part woman DO I see Past Sg-lst'I saw the woman to whom Johngave the book.'

It does not work on other positions, e.g. instrumentals,where a differentRC-formingstrategyis required,retainingNPrel in the restrictingclause:(19) Enna(k)katti(y)-dl kori(y)-ai antamanitanwhich knife withchickenDOthat mankolaippi-tt -an antakatti(y)-ai jan kan-t -an.kill Past Sg-3rd-Maschat knife DO Johnsee Past Sg-3rd-Masc

'John saw the knife with which the man killed the chicken' (literally: withwhich knife the man killed the chicken, John saw that knife).Roviana, a Melanesian language spoken in New Georgia, Solomon Islands, alsoprovides some support or discriminating he indirectobjectposition. There, in simplexsentences, indirect objects are treated like other oblique case NPs in that they arepreceded by a preposition,whereas directobjects are not:(20) Ele ponia Jone koe Mere sa buka.Pastgive John to Marythe book

'Johngave the book to Mary.'(21) Vekoa Jone sa bereti pa tevelo.put Johnthe bread at table'Johnput the bread on (the) table.'On the other hand, in relative clauses indirectobjects are relativizedjust like directobjects and subjects-a postnominal strategy, in which the case of the relativizedposition is not marked, as the relativizationmarker apu is morphologicallynvariableand no pronounis retained in the position relativized. But in the oblique cases thefunction of NPrel s coded, either in the form of a stranded adverb or in the variableform of the relativizationmarker:

(22) a. sa bukasapu ele ponia Jone koe Merethe book that Past give John to Mary'the book thatJohn gave to Mary'b. sa barikalege apu ele ponia buka Jonethe woman that Past give book John'the woman thatJohn gave the book to'(23) sa tevelo vasina vekonia Jone sa beretithe table where put (by?) John the bread'the table where John put the bread'


13/38


1.3.4. Subject-Oblique. The primarystrategy in Korean (Tagashira(1972)) placesthe restrictingclause to the left of the head NP, separatedfrom it by the suffix -()n,-nin. NPrei is simply deleted for all NPs on the AH down to and including obliques:

(24) hy;nsik -i ki la -lil ttali-n maktakiHyensik SU the dog DO beat Rel stick'the stick with which Hyensik beat the dog'However, where genitives are relativized, a pronounmust be retained:

(25) chaki-ij la -ka chongmyongha-n ki salamhe of dog SU smart Rel the man'the manwhose dog is smart'In fact, many languages besides Korean change RC-formingstrategiesat the genitiveposition by presentinga pronominalelementin the position relativized;see Table2 insection 2.2.2.

Furthersupportfor distinguishing enitives on the AH comes fromlanguages ikeCatalan and North Frisian(Fering dialect), where genitives (and objects of compari-son) are simply not relativizableat all, although all NPs higher on the AH thangenitives are relativizable,e.g. North Frisian:(26) a. Johnklanddet wuf's henk.'Johnstole the woman's chicken.'

b. det henk wat kland John'the chicken thatJohnstole'c. *det wuf wat's henk Johnkland'the woman whose chickenJohnstole'1.3.5. Subject-Genitive. A great manywell-known anguageshave primary trategiesthat permit relativizationon all the positions on the AH except that of objects ofcomparison.Thus in French we cannotrelativizelejeune homme in (27a):

(27) a. Marie est plus grande que le jeune homme.'Marieis biggerthanthe young man.'b. *lejeune hommeque que Marie est plus grande'the young man than whom Marieis bigger'The same situation holds in Spanish, German,and Romanian nonparticipial trategy,in each case).1.3.6. Subject-Object of Comparison. Few languages that distinguish objects ofcomparisonfrom direct objects or oblique NPs permit them to be relativized. InEnglishwe do have phrases like (28), thoughsome find them ratheruncomfortable:

(28) the manwho Mary is taller than


14/38


And in Urhobo, whose primarystrategy is postnominal,and pronoun-retainingor allpositions on the AH, we have:(29) oshale na 1- i Maryrho n- o

man the that Marybig thanhim'the manthat Maryis biggerthan'This completes the argument hat primaryRC-forming trategies can discriminateall the positions on the AH. Thus, the PRC is established, and we may concludegenerally that the Accessibility Hierarchy determines the relative ease of relativeclause formation rom unmarked implex sentences across languages.Table 1 (p. 76) summarizesour data concerningthe relativizingpower of the RC-formingstrategieswe considered n the languages n our sample. A key to the entries inthe row-column intersections is given at the end of the table. Needless to say, such a

summarycannot adequately representall the language-particularroblems nvolved indetermining he nature of an RC-forming trategy.1.4. Problems and Possible CounterexamplesWe shall consider here two types of difficultieswith the analysis we have proposed.The first concern methodologicalproblems nvolved in identifyingsubjectsand relativeclauses (RCs). The second concern specific counterexamplesto the HierarchyCon-straints.1.4.1. Methodological Problems1.4.1.1. IdentifyingSubjects. We are using a largelytraditionalnotionof subject. Anattempt to make explicit the large numberof specific propertiesthat comprise thisnotion is given in Keenan(1976b)and will not be discussed here. It is clear from thatinvestigation, however, that the NPs we call subjects in some languages are moresubject-like han those of other languages.Thatis, they possess a greaternumberof theproperties hatare characteristicof subjectsin general.A language n which the subjectpropertieswere systematicallydistributedacross two or moreNPs thenmightarguablybe said not to have a single category of subject. In such a case the AH for thatlanguagewould lack the subject position, much as the AH appliedto other languagesmay lack an OCOMP position (see 1.2.1), and the predictionsmadeby the HCs wouldbe reduced and furtherprincipleswould have to be found to account for the degree ofrelativizabilityof NPs not on the AH.Fortunately, many languagesdo appearto presentNPs that conformfairly well tothe traditional oncept of subject. Nonetheless, three categoriesof possibly subjectlesslanguages have been recently discussed in the literature: Tagalog and Philippinelanguages generally (Schachter(1976)), topic oriented languages ike Lisu and otherSino-Tibetan languages (Li and Thompson (1974; 1976)), and ergative languages(Tchekhoff (1973)) (see 1.4.2 below).


15/38


Table 1Language Relativizable positions

Relative clause formingstrategy Subj DObj IObj Obl Gen OComp

Aoban (North-East)1. postnom, -case + - - - - -2. postnom, +case - + + + + +

Arabic (Classical)1. postnom, -case +2. postnom, +case - + + + + +

Basque1. prenom, -case + + +

Batak (Toba)1. postnom, -case +2. postnom, +case - - + + +

Catalan1. postnom, -case + + + - - -2. postnom, +case - - - +

Chinese (spoken Pekingese)1. prenom, -case + +2. prenom, +case - + + + + +

Czech (colloquial)1. postnom, +case + + + + + +?Dutch

1. postnom, -case + +2. postnom, +case - - + + +

English1. postnom, -case + +2. postnom, +case - - + + + +

Finnish1. postnom, +case + + + + +2. prenom, -case + + - - -

French1. postnom, +case + + + + +

North Frisian (Fering dialect)1. postnom, -case + + +2. postnom, +case - - -(?) +

Fulani (Gombe dialect)1. postnom, -case + + *2. postnom, +case - - * + +


16/38


Table 1 (Continued)Language Relativizable positions

Relative clause formingstrategy Subj DObj JObj Obi Gen OComp

Genoese (Zeneyze)1. postnom, -case + + - - - -2. postnom, +case - + + + + +

German1. postnom, +case + + + + +2. prenom, -case + - - - - -

Gilbertese1. postnom, -case + - - - - *2. postnom, +case - + + + + *Greek (Modern)1. postnom, -case + + - - - -2. postnom, +case - - + + + +

Hausa1. postnom, -case + +2. postnom, +case - - + + +

Hebrew1. postnom, -case + + - - -2. postnom, +case - + + + + +/*

Hindi1. postnom, +case + + + + + *2. internal, +case + + + + + *

Iban (Sea Dayak)1. postnom, -case + - - - - -2. postnom, +case - - - +/-

Italian1. postnom, -case + + - - -2. postnom, +case - - + + +

Japanese1. prenom, -case + + + +1- +1- -?2. prenom, +case - - - - +1- -?

Javanese1. postnom, -case +2. postnom, +case - - - - +

Kera1. postnom, -case + - - - - *2. postnom, +case - + + + + *


17/38


Table 1 (Continued)Language Relativizable positions

Relative clause formingstrategy Subj DObj IObj Obl Gen OCompKorean

1. prenom,-case + + + +2. prenom, +case - - - - +

Luganda1. postnom, +case + + * - - *

Malagasy1. postnom, -case +

Malay1. postnom, -case + +/- - - - *2. postnom, +case - - + + + *

Maori1. postnom, -case +

Minang-Kabau1. postnom, -case + - - - - *2. postnom, +case - - - + + *

Persian1. postnom, -case + + - - - *2. postnom, +case - + + + + *

Polish1. postnom, +case + + + + +

Romanian1. postnom, +case + + + + + *

Roviana1. postnom, -case + + +2. postnom, +case - - - + +

Russian1. postnom, +case + + + + +Sinhala

1. prenom, -case + +/+? +? +/-Shona

1. postnom, -case + +? + - - *2. postnom, +case - - - + + *

Slovenian1. postnom, +case + + + + + +


18/38


19/38

80 EDWARD L. KEENAN AND BERNARD COMRIEOf these categories, Schachter's claim that the focus NP in Tagalog cannot be

regarded as a subject is the most damaging to our claim, since Tagalog was one of thelanguages that justified HC1 and HC3 (a language may relativize only subjects).However, it has been shown in Keenan (1976a) that others of the Western Malayo-Polynesian languages, notably Malagasy, do not present the evidence that supportsSchachter's claims for Tagalog, and consequently there is still sufficient data to supportHC1.

The evidence Li and Thompson present from Sino-Tibetan is less damaging, sincethe NPs that are most subject-like do not present any Hierarchy violations. What theyargue, however, is that the syntactic category subject has a very low functional loadin these languages in that few if any syntactic processes need to be made sensitive tothe NP that is the subject. Further research in those languages might then reveal thatcategories other than those on the AH, e.g. topic, are relevant in determining RC-forming possibilities.

With the possible exception of Dyirbal (and perhaps Eskimo; see Woodbury(1975)) (see section 1.4.2.1), the claim that ergative languages lack subjects has beenrefuted by Anderson (1976), who shows that with respect to many major syntacticprocesses transitive subjects (the ergative NP) and intransitive subjects behave alike, indistinction to DOs, thus establishing the fact that there is a category of subject in thoselanguages.1.4.1.2. Identifying Relative Clauses. We have been considering an RC to be anysyntactic structure that designates an object (or set of objects) in a certain way,namely, by first specifying a larger domain of objects and then restricting it to a subset,perhaps a one-member subset, of which a certain sentence, the restricting sentence, istrue. However, many languages present sentence types that appear to designate objectsin this way but in which there is no surface constituent with which we can associate thedesignating properties. One such example concerns extraposed RCs, as in (30a) below:

(30) a. The student finally arrived who we had been waiting all morning for.b. the student who we had been waiting all morning for

Clearly, to evaluate the truth of (30a) it is necessary to determine that the mainpredicate holds of an object that is, first, a student, and second, is such that thesentence we had been waiting all morning for him is true of him. That is, the object inquestion is clearly the one designated by the RC in (30b), and in this sense the logicalstructure of (30a) contains the RC in (30b). But (30a) itself does not present in surface aconstituent that has the referential properties of (30b) and so does not contain an RC.

In this case, however, it is easy to argue that at a fairly shallow level of underlyingstructure (30b) does occur in the syntactic structure of (30a), and that a late rule of RCExtraposition moves the restricting clause around the main verb. Thus the syntacticstructure of (30a), considered as the sequence of phrase markers representing its


20/38


derivation, does contain an RC, andit is clear, then, that our definitionof RC does notrequirethat, in surface, the head NP and the restrictingclause be a constituent.It is clear, furthermore, hat the most ordinary ype of RC in Englishis expressedby structures like (30b), those in (30a) being more markedand of more restricteddistribution. But in many languages the usual translations of English sentencescontaining RCs do not present the head NP andthe restrictingclause as a constituent.(31)from Hindi and (32) from Walbiri Hale (n.d.)) illustratethis type:(31) Mai us aurat ko jantahoonjis- ko Ramne kitabdiya.I thatwoman DO know who IO Ram Erg bookgave'I know the womanthat Ram gave the book to.'(32) ijatjulu-Ju 'k-payankiripantu-nu, kutja- pa tapa ia- nu.I Erg Aux emu spear Past CompAuxwater drinkPast

'I spearedthe emu that was drinkingwater/while t was drinkingwater.'In these cases it is not clear that what corresponds to the restrictingclause in Englishever occurs as a constituentwith the head NP in underlying tructure.Both Hale (n.d.)and Andrews(1975) argue that it does not.However, languageswith this type of constructionvery regularlyexhibita relatedtype of construction n which there is a constituent that meets our semantic conditionsof relative clause-hood. The correspondingversions of the Hindi ((31)) and Walbiri((32)) are given below in (33) and (34), respectively:(33) [Ramne jis aurat ko kitabdiya]us (aurat) ko maijantahoon.RamErg which woman IO book gave that woman DO I know'I know the woman that Ramgave the book to.'(34) [yankiri-li kutja- pa ijapa Da- nu] ula pantu-nu ijatjulu-lu.emu Erg CompAux water drinkpast that-onespear Past I Erg'I spearedthe emu that was drinkingwater.'

It is not clear that the bracketed constituents above should be considered NPs.Certainly hey presentthe normalsyntaxof full sentences, except that they contain themarkersjis and kutja. Whatever the grammaticalcategory of these constituents,however, it does appear thatthey meet our definitionof RC-they specify a domainofobjects, those markedby jis and kutja, and they give the restrictingclause, the onedeterminedby the entire clause in which they occur. Thus, like the examples (3) fromBambara and (4) from Diguenlomentioned earlier, they illustrate an RC-formingstrategy in which the head NP occurs within the restrictingclause. Our definition ofRC then does not requirethat an RC be an NP, nor does it requirethat the head NPcommandthe restrictingclause (a use of head that differsfrom the usual one in thelinguisticliterature).The RC-forming trategies llustratedabove in Walbiriare perhapsresponsibleforthe claim sometimes heard that, in violation of HC1 (which states that subjects are


21/38

82 EDWARD L. KEENAN AND BERNARD COMRIEuniversally relativizable), Australianlanguages do not have RCs. Walbiri at leastclearly does not present embedded RCs in which the head NP commands therestrictingclause and in which the two form an NP constituent. But it certainlydoespresent constituents that meet our semantically-based riteriafor what constitutesanRC, so HC1 is not violated here.1.4.2. Some Possible Counterexamples1.4.2.1. Possible Counterevidence to HC1. We consider first the best documentedcounterexample o HC1(subjectsarealways relativizable).The counterevidencecomesfrom Dyirbal (Dixon (1969; 1972)),an Australian anguage. Dyirbalappearsergativeinthe standardsense that full NP subjectsof intransitivesentences and full NP DOs oftransitive sentences are case-markedin the same way (zero, in this case), whereassubjectsof transitive sentences carrya special marker,the ergative. In the analysistofollow we shall refer to intransitive subjects and transitive DOs collectively asabsolutives. Examplesare from Dixon (1972, 100-101).

(35) bayi yara banagajiu.Det-Abs man-Abs return'The manis returning.'(36) bayi yuri baijgul yataggu bagan.Det-Abs kangaroo-AbsDet-Erg man-Ergspear'The man speared a kangaroo.'Dixon demonstrates that absolutives can be relativized from both transitive andintransitive sentences, but the ergative NP cannot be relativized. To talk about theman who speared a kangaroo, it is necessary to apply a transformation hat promotesthe ergativeNP to absolutive status (37a), whence it can be relativized (37b):

(37) a. bayi yara bagalUajiu bagul yurigu.Det-Abs man-Abs spear-' Pass Det-Instr kangaroo-Instr'The manspeared the kangaroo.'b. [Npbayi yara bagal-rja -iju bagulDet-Abs man-Absspear Pass Rel Det-Instryurigu] banagajiu.kangaroo-Instr eturn'The man who speared the kangaroo s returning.'

On the basis of the Dyirbaldata, Johnson (1974a) has suggested that the uppersegment of the AH is inapplicable o ergative languagesgenerally, and that an ErgativeHierarchy (EH), ABS > ERG > 10 > OBL, etc. more adequately describes theaccessibilityto RC formation n those languages. The EH differs significantly rom theAH only in that it predicts the possibility of RC-forming trategies that relativize onlyabsolutives, such as in Dyirbal. But general support from ergative languages for thisprediction is lacking. The majorityof ergative languages known to us permit both


22/38


ergativeand absolutive NPs to relativize. We have alreadyillustratedrelativizationonergatives in Basque (17b) and Walbiri (34). Examples (38)-(42) below illustraterelativizationon ergativesfrom a few other languages.Hindi(38) Ram us larki ko pasandkartahe, is- ne kapre dhoye.Ramthat girl DO likes Rel Ergclothes washed'Ramlikes the girlwho washed clothes.'Tongan (Anderson(1975))(39) te mo fetaulakimo e tangata oku ne fua 'a e siokivai.Fut 2 Du meet with Artman Pres3Sg carryPart Artjug water'You will meet a manwho is carryinga jug of water.'Gahuku(Deibler (1973))(40) izegipa get -a vechild begot 3Sg man'the man who had begottenthe child'Jacaltec (Craig(1974))

(41) x- 0- w- 'il [naj x- 0- watx'e-n hun ti']Asp him-AbsI-Ergsee manAsp him-Abs make Rel one this'I saw the man who made this.'Greenlandic Eskimo (Woodbury (1975))(42) qimmi-O -a tuquk-kiga nalu Nil- saNa.dog Abs-Sg Poss-3Sgkill Trans gnorantNeg Indic-3Sg-lSg'He knows aboutme, who killed his dog.'

Woodbury (1975)does argue, however, that absolutives are more relativizable nGreenlandic han are ergatives, on the grounds that (1) RCs formed on ergatives aresomewhatmore restricted in the distribution n matrix clauses (p. 21) than are thoseformed on absolutives, and (2) for certain verb classes ergatives cannot be relativizedout of the active participle(p. 27). To relativize that NP from an active participle, theergativemustfirst be promotedto an absolutive via Antipassive. Nonetheless, it is thecase that ergatives can in general be relativized, even if not quite so freely asabsolutives, and so the AH does appear to apply, although more work would need tobe done to distinguishaccording o our criteriawhat the differentRC-forming trategiesin Greenlandicare.The general claim, then, that in ergativelanguagesabsolutives are morerelativiza-ble than ergatives receives little support. This in turn suggests that the Dyirbal datamight be analyzed differently. One reanalysis that would have the effect of makingDyirbal conformto the AH would be simplyto regardthe absolutive NP in transitivesentences (as well as in intransitiveones) as the subject. Such an analysis again callsinto question the definingcriteria or subjecthoodand wouldrequiremuch independent


23/38


support. Perhapssurprisingly,however, Dyirbaldoes providesuch support.Below wesummarizethis evidence, citing page numbers from Dixon (1972) for the supportingdata.1. Absolutives are the most essential NP in the sentence. Thus an unspecifiedagent (ergative) may simply be eliminatedfrom a sentence, much like unspecifiedagents of passive sentences in English. But unspecified absolutives cannot be soeliminated. If we want to eliminate an unspecifiedpatient, we must first passivizethe sentence so that the ergative becomes a surface absolutive and the formerabsolutivebecomes oblique, whence it can be deleted when nonspecific (p. 70).2. Absolutives are the only target of advancementrules. Dyirbal has severalrules, rather like Passive in English, that promote e.g. ergativesand instrumentals othe absolutivepositionof underlying ransitivesentences, as we illustrated n (37a).3. Absolutivesnormallyprecede ergativesin simplex sentences (p. 291).4. Only absolutives can be coreferentiallydeleted by operations like Equi andConjunctionReduction(pp. 73, 67). To coreferentiallydelete an ergative NP, it mustfirst be advanced to an absolutive.5. It appears that certain demonstratives, always definite and presupposingreference-a characteristicpropertyof subjects-are restricted to absolutivepositions(p. 218).6. In the few cases where we can tell, absolutives appear to control reflexives.Usuallyreflexivization s done by verbalaffixing,so the derivedverbis intransitiveandthe subject is absolutive on independent grounds. However, in a few examples theobject is an inalienablebody part and is expressedin surface structureeven though theverb is reflexivized. And in these cases the controller s absolutive (p. 153):(43) bayi yara mala da1gaymarijiu.Det-Abs man-Abs hand-Abs eat-Refl-Pres/Past'The manchewed his finger.'If we adopt the analysis of Dyirbal n which absolutives in transitivesentences arethe subjects, then Dyirbal will be typologically anomalous in only one majorrespect:namely, in the least markedtype of transitive sentence the NP with the referentialor

topic propertiesof subjectsdoes not express the agent. And since the predominanceoftopic propertiesover agency properties n this case is quite large (veryfew transforma-tions mention ergatives, but very many mentionabsolutives), it turns out that the mostsubject-likeNP in basic transitivesentences is the absolutive and does not express theagent.This analysis at least has the (metalinguistic)advantage of isolating at a singlepoint the anomaly of Dyirbal-namely, transitivesubjects in basic sentences are notagents. The ergativeanalysis on the other hand makes Dyirbal anomalousin a greatmany respects: the word order is OSV; DOs are the most essential constituent oftransitive sentences, subjects being freely deletable when unspecified;etc. Further-


24/38


more, it is possible that when the lexicons of more languageshave been analyzed indetail, it will be seen thatmany languageshave classes of verbs that select goal subjectsbut allow agents present in oblique cases. Biggs (1974) has recently argued, forinstance, that Fijian has one large class of verbs that select goal ratherthan agentsubjects, whereas another class selects agent subjects, as is more usual. PerhapsDyirbalwill then be seen to be merelynear the end of a continuum n that most of itstransitive verbs select nonagent subjects.1.4.2.2. ApparentExceptions to HC2. A second type of counterexamplewe considerare apparent exceptions to HC2, that a given RC-formingstrategy must apply to acontinuous segment of the AH. Hausa (Abrahams (1959); Schachter (1973)) andYorubaappearto use RC strategies hat applyto discontinuoussegmentsof the RC.Thus Hausa seems to use a +case strategyto relativizesubjects, since a personalpronounseems present in the positionrelativized:

(44) dokin da ya mutuhorse Rel it died'the horse thatdied'Relativizationon DOs, however, does not permitretentionof a pronoun n the positionrelativized, and as no other nominal particle is present to code the case of therelativizedNP (the relativizerda is invariable), RC formation on DOs appears to beeffected by a -case strategy:

(45) mutuminda na gani (*shi)person Rel I saw him'the person that I saw'However, when we relativize on OBLs in Hausa, a pronoun is again present. (Inrelativizingon 1Os, the presence of a pronoun s optional.)

(46) yaron da sukagaya {wa/masa}child Rel they said to to-him'the child whom they told'(47) wuqad da ya kashe ta da itaknife Rel he killed her with it'the knife with whichhe killed her'It appears, then, that the postnominal,+case strategy in Hausa does not apply to acontinuoussegmentof the AH, in violationof HC2.On examining further the pronoun that appears in the RC when a subject isrelativized, however, we note that this same pronoun is also required in simplex


25/38


sentences with a subjectNP:(48) a. Yusufuya zo.Joseph he came

'Joseph came.'b. *Yusufu zo.Quite generallyin Hausa, full subject NPs must be accompanied by clitic pronouns.This suggests a different line of analysis for such pronouns.Insteadof regardinghemas constituent parts of a +case strategy, we may regardthem as an instanceof verbagreement.In many languagesverb-agreement ffixes are knownto derivefrom(clitic)pronouns(Givon (1976)), and it is oftendifficultto draw a precise dividing-linebetweenclitic pronouns and agreement affixes. We are therefore suggesting that where thepresence of a pronoun is requiredby the presence of a full NP, then the pronounberegardedas an instance of verb agreement,andnot as an instanceof NP case marking.In Hausa, such internal pronouns (for this terminology, see Keenan (1972a, 447-450)) occur only with subjects;with DOs, etc., we have (49a) and (49c) for instance,but not (49b)and (49d):

(49) a. Na gani mutumin.b. *Na gani shi mutumin.'I saw the person.'c. Ya kashe ta da wuqad.d. *Ya kashe ta da ita da wuqad.'He killed her with the knife.'Another similarpossible counterexample, rom Tongan (pointed out in Anderson(1975)),is not so easily disposed of, however. Here it appearsthatpronounsmay, andin some cases must, be retainedwhen subjects are relativized. (39) already illustratesthis. Similarly, relativizationon 1Os, OBLs, and GENs obligatorily eaves behind apronominal race (examples from Anderson (1975), Chung (personal communication)):(50) ko e 'eiki eni na'e langa mo'ona 'a e fale lahi.PartArt chief this Past buildfor-3Sg PartArthouse big

'This is the chief for whom the big house was built.'(51) fakahamai kiata au 'a e tamasi'ina'eshow towards to-Pers me Abs Art child Pastngalo 'ene polosi fulunifo.disappearhis brush toothbrush'Show me the boy whose toothbrushdisappeared.'However, it is not possible to present pronominal races when DOs are relativized:

(52) 'oku 'ikai 'ilo 'e ha taha 'a e tangatana'a ku taa'i (*ia).Pres not know Erg any one Abs Art man Past I hit him'Nobody knows the manwho I hit.'


26/38


Furthermore, he type of explanationused above for Hausa and Yorubadoes notin generalseem applicable.The subjectpronounspresentwhen subjectsarerelativizeddo not in general function as agreementparticles,althoughthe pronounpresentin theRC is the clitic one, not the independentone. Thus it occurs preverbally,whereastheindependentpronounsand full NPs functioningas subjectsnormallyoccur postverbally(Tonganis VSO).

(53) na'e 'ave 'e Sione ho'o telefone.Past take Erg John yourtelephone'Johntook away yourtelephone.'(54) *na'ane 'ave 'e Sione ho'o telefone.Pasthe take Erg John yourtelephoneFurthermore, he clitic pronouns can be present when the subject is an independentpronoun,althoughthe meaninghere is emphatic:

(55) na'a ne taa'i 'e ia pe ia.Past he hit Erg he Emph him'He hit himself.'This use of the clitic could conceivablydevelop so that the clitic would cooccur withfull NPs in subject position without the emphaticmeaning.This at least would be anormalway for an agreement o arise (see Givon (1976)forjustification of this claim).But this would be to anticipate the development of Tongan and cannot be used tojustify its currentstatus as a counterexample o HC2.Perhaps the best explanationwe can offer for the patternof RC-forming trategiesin Tongan s a historicalone. As we have mentioned,Tongan s an ergative anguage nwhich the ergative marker 'e is cognate with the passive agent marker in relatedPolynesian languages. Hohepa (1969) has presented evidence that the ergative para-digm present in Tonganand the closely related Niuean has evolved from an originalnominative-accusativeparadigm n which the productivedistinction between activeand passive verb forms was lost and the passive morphologyretainedon the majorNPs. Thus, historicallyspeaking,the ergativesubjectof Tongan s a passiveagent,andthe absolutive DO in Tongan is historically the subject. We suggest here that theevolution of the agent NP to subject status in Tongan is not quite complete. WhileAnderson(1976)has shown that it does possess manygeneralpropertiesof subjects,itmay retaina few traces of its originaloblique case status. For example, in simpleactivesentences the ergative NP can be omitted when unspecified, like the ergative NP inDyirbal and the passive agent in English, as illustrated n (56) and (57) below (bothfrom Churchward1953)):

(56) na'e tamate'i 'e Tevita 'a Koliate.Past kill Erg David Abs Goliath'David killed Goliath.'


27/38

88 EDWARD L. KEENAN AND BERNARD COMRIE(57) na'e tamate'i 'a Koliate.Past kill Abs Goliath'Goliath was killed.'Now, since OBLs in general in Tonganrequireretention of a clitic pronounwhenrelativized (in one case a nonclitic pronoun can be retained, but this is quiteexceptional), the ergative NP in Tongan is behaving like the other OBLs in thisrespect. On the other hand, the DO does not allow a pronounto be retainedunderrelativization. But, as we show in 2.2.2, this is characteristicof subjects in general.That is, even in languages like Hebrew in which pronounretentionis normal underrelativization,subject pronouns are normallynot retained.So, in this way, the DO inTongan is behaving like a subject. In these very minor respects, then, perhaps theergative and absolutive in Tongan betray their earlier status as passive agent and

subject, respectively.2. Towards an Explanation of the Hierarchy Constraints2.1. The AH as a Psychologically Valid EntityGiven that the HCs do makecorrectpredictionsabout RC formation n a wide rangeoflanguages, it is natural to wonder why this should be so. We propose the followingexplanation: The AH directly reflects the psychological ease of comprehension. Thatis, the lower a position is on the AH, the harder t is to understandRCs formed on thatposition.If the AH does reflect the psychological accessibility of NPs to RC formation,then we can use this fact, in conjunctionwith certainotherassumptions,to explaintheHCs in the following way. First, it would be natural hat a way of relativizinga certainposition might not be applicable at the next lower position (HC3) on the generalassumption that syntactic processes are ways of encoding meanings; and, if onemeaning is inherently more difficult to encode than another, then a strategy forencoding the first need not apply to the second. By the same token, a strategythatapplies to one position but fails to apply to the next lower position would not beexpected to apply to a still lower position (HC2). For, if a given strategy is used toencode a fairly easy meaningand that strategy is strong enough to encode a ratherdifficultmeaning,then it is surely strongenough to encode the meaningsof intermedi-ate difficulty. However, this rather nformal ine of reasoningcan be slightlyextendedto yield an argument hat languagesshould not have gapped RC-forming trategies.Thus, if speakerscan in general encode a meaningof a certaindifficulty in some way,then they shouldcertainlybe able to encode meaningsof lesser difficulty (possibly in adifferentway). Yet we have seen that Toba Batak, for example, can relativizedirectlysubjects using one strategy, and 1Os, OBLs, and GENs using another. But DOscannot be directly relativized using any strategy. As we mentioned, however, in allcases in our samplein whichlanguagespresent gapped strategies, the NPs that occur in


28/38


the gap can always be promoted to positions that can be directly relativized viaoperations ike Passive. In these cases, then, we wouldlike some independentevidencethat it is psychologically easier, in those languages,to for instance promotea DO tosubject and then relativizeit ratherthan to relativizeit as a DO. We have no directpsychological evidence of the right sort. It is worth noting, however, that in thoselanguages that promote an NP to a higherpositionon the AH in orderto relativizeit,the promotion system (e.g. the ways of convertingNPs low on the AH to ones higheron the AH) is usually well developed, very commonly used, and has a wide syntacticdistribution, n distinctionfor instanceto Passive in English,which is, by comparison,a less usual, more markedform. In fact, in Hawkinsand Keenan(1974)it is reportedthat on certain types of repetition tasks, English-speakingchildren(10-12 years) dosignificantly ess well on RCs like the boy who was seen by Marythanthey do on oneslike the boy who Marysaw. Thus, promoting o relativizeis certainlynot universallyeasier thanrelativizingdirectly.Finally, note that this psychologicalinterpretation f the AH cannot fully accountfor HC1,that subjectsmust, in general,be relativizable. It only justifies the claim thatsubjects are easier to relativize than any other position on the AH, but it would allowin principlethat, in some language,no position on the AH be relativizable.2.1. Evidence in Support of the Psychological Validity of the Accessibility HierarchySeveral recent experimental studies do provide partial supportfor the claim that theAH does represent the psychologicalaccessibility to RC formation. Thus, in Legum(1975) it is shown that English-speaking hildren, aged 6-8, comprehend RCs formedon subjectsbetter thanones formed on DOs. Similarresults were achieved by Brown(1971) for childrenaged 3-5 years. Hatch (1971) shows that young childrenrespondmore rapidlyto RCs formed on subjectsthan on objects, and Cook (forthcoming)hasshown that children and adults recognize with fewer errors instances of head NPsbearing the subject relationto the subordinateverb in RCs than instances of the headbearing the object relation to the verb. And Valli et al. (1972), in a productiontest,show that French childrenat the level of 61 (approx. 12 years of age) produce RCsformedon subjectsmuchmore readilythan on objects, and RCs on possessor NPs aremuch rarerstill and in the majorityof instances incorrectlyformed.It should be notedthat the purpose of these studies was not specificallyto validatethe AH. Variablesother than the position on the AH of the NP relativizedwere ingeneral shown to be relevant(e.g. whether the RC was embedded or not, whetherthefunction of the RC in the matrixwas the same as that of the head in the RC (Sheldon(1974));see Legum(1975) for some discussion of these other parameters).On the other hand, Hawkins and Keenan (1974) report a study specificallydesignedto test the psychological validityof the AH. Children,aged 10-12, were givenrepetitiontests involving RCs formed on all the positions on the AH. Correctnessofrecall correlated significantly (


29/38

90 EDWARD L. KEENAN AND BERNARD COMRIEestablishedin that study was:

SU > DO > IO > OBL, OCOMP > GENThus, RCs formed on subjects were recalledwith fewer errorsthan those formedonDOs, which were better than ones formed on 1Os, which were better than thoseformed on either OBLs or OCOMPs, which were better than ones formed onpossessor NPs. Clearly the only position out of place on the recall hierarchyis theOCOMPposition, whichwas treatedby the childrenas the same in accessibilityas theOBL position. All RCs on OBLsleft the prepositionstranded,e.g. the boy whoJohnnytook the toy from. Apparentlystrandingthe comparativeparticle, e.g. the boy whoJohnnyis taller than, was interpretedas being similarto prepositionstranding.While it would certainly be premature,on the basis of these few studies, toconclude that the AH does establisha hierarchyof psychologicalaccessibility, it doesappearthat the available evidence points in that direction. If furtherresearchalongthese lines justifies the psychologicalinterpretationof the AH, then we will have anexplanation or two further acts fromour own study. Alternatively,these facts can beconsidered furthersupportfor the psychologicalinterpretationof the AH. The firstconcerns a pattern of judgments of relative accessibility of RCs from languagesinwhich most positions on the AH can be grammaticallyrelativized. The secondconcerns the distributionof personalpronounsretained n the RC.2.2.1. Intralanguage Relative Accessibility Judgments. One natural way to extend theHCs would be to interpret hem not only as a cross-languageorderingof grammatical-ity, as we have done, but to consider them as an acceptabilityorderingwithin eachlanguage.Thus one mightexpect that in general,withina given language,RCs formedon the high end of the AH would be judged more acceptablethan RCs formedon thelow end. And in the extreme cases this appears to be correct.RCs formed on subjects are always among the most acceptable in any givenlanguage, and those formed on objects of comparison,where possible at all, are oftenjudgedto be only marginallyacceptable.For instance, manyinformants n Englishareuncomfortable with the man who Mary is taller than. Similar judgments of relativeacceptabilityobtainedfor our Hebrew informants.Thus, even in languages in whichobjects of comparisoncan be relativized,therewill be a preference to express the RCas one formed from a semanticallyequivalentsentence in whichthe semanticobject ofcomparison s presentedas a subject.That is, the preferredway to express the idea ofthe man who Mary is taller than will be, in general, the man who is shorter than Mary.A somewhat more subtle performancepreference is illustratedby the genitiveposition on the AH. Although a majorityof languages in our sample possessed someway of relativizinggenitives, there was often some awkwardness n doing so, and notinfrequently, n specific cases, a preferredalternativewas offered-one that relativizes


30/38


a position higheron the AH. For example, (58) is perfectly grammaticaln French:(58) la femme dont le manteaua ete volethe womanwhose the coat has been stolen

'the womanwhose coat was stolen'But in practice people will tend to say:

(59) la femme qui s' est fait voler le manteauthe womanwho herselfis caused to-stealthe coat(59) is literally 'the woman who got her coat stolen', althoughthere is no necessaryimplicationof her having engineered the theft; relativization s on the subject, ratherthan the genitive. The translationsabove indicate that a similar alternativeto genitiverelativization exists in English. A like situation holds in Swedish, where the (b)alternative n each pair is preferred:

(60) a. kvinnan, vars kappablev stulenwoman whose coat was stolenb. kvinnan, som fick sin kappa stulenwoman who got her coat stolen(61) a. kvinnan, vars man arp'a jukhusetwoman whose husband s in hospitalb. kvinnan, som harsin man pa sjukhusetwoman who has her husband n hospitalIn Yoruba, inalienablepossessives/genitives can often be paraphrasedby otherconstructions, and even with simple sentences these paraphrasesare preferred:(62) a. Johnlu ts Qkunrin aa.John strikeleg man the'Johnstruck the man's leg.'b. Johnlu okunrinnaa1' se.John strike man the on leg'John struck the manon his leg.'

Under relativization,the preference is even more strongly marked, and our informantwas very hesitant aboutadmitting 63a) below:(63) a. ?*Mo ri okunrin i John lu esq ri.I see man thatJohn strike leg his'I see the manwhose leg John struck.'b. Mo ri Qkunrini Johnlu 1' ese.I see man thatJohn strike on leg'I see the man thatJohn struck on the leg.'


31/38


Bamgbose (1966, 158-159) claims that the relation between sentences like (63a) and(63b)is transformational.2.2.2. Pronoun Retention. A less obvious patterningdetermined n part by the HCsconcerns the distributionof personal pronouns in relativizedpositions. (We excludepronouns that are instances of verb agreement;see section 1.4.2.2). We have alreadynotedthat Semitic languages characteristically presentsuch pronouns, e.g. Hebrew:

(64) ha- isha she- David natan a et ha- seferthe woman that Davidgave to-her DO the book'the woman that David gave the book to'Further,it has been arguedin Keenan(1972a; 1975a) hat such RC-forming trategiespresent in surface structuremore of the logical structureof the RC thando languageslike English that do not present such pronouns. The reason, in brief, is that in thepronoun-retaining trategies the restricting clause in surface is a sentence-one thatexpresses exactly the restricting sentence of logical structure.That is, it is just thesentence that must be true of the referentof the RC. So a possible referent of (64)above mustbe a woman of whom the sentence David gave the book to her is true. Andthe translationof that sentence in Hebrew is preciselyDavid natan la et ha-sefer-thesurface restrictingclause. Note that the correspondingclause in English,David gavethe book to or to whom David gave the book, is not a surface structure sentence, andso is not immediatelyperceivedas the sortof linguisticentitythat is trueof objects.

It has been shown elsewhere (Keenan (1972a, and especially 1975a)) hat the RC-forming strategiesthat retainpronouns are applicableto a greaterrange of otherwisedifficult environments e.g. it is often, but not always, possible in these languages orelativize into coordinate NPs, other relative clauses, indirect questions, and evensentence complementsof NPs). The reasonis that the logically moreexplicit strategiesstill successfullyexpress the basic meaningof the RC in contexts where the meaning sotherwise difficultto perceive. Consequently,we are led to predict that, as we descendthe AH, languageswill exhibit a greater endencyto use pronoun-retainingRC-formingstrategies.Table 2 (p. 93) dramatically upportsthis prediction.Languages hat do notnormallyretainpronouns n any positionare not includedin the table;nor is the use ofpronouns as markersof verb agreement see 1.4.2.2).It should be clear from Table 2 that not only does the tendency to presentpronouns in positions relativized ncrease as we descend the AH, but also that once alanguagebeginsto retainpronouns t mustdo so for as long as relativization s possibleat all. This is a naturalconsequence of the hypothesis that pronounretentionwill beused in proportion o the difficulty of the position being relativized, though the criticalpoint of difficultyis differentfor different anguages.2.2.3. Explaining the Psychological Validity of the AH. To some extent, explainingthe HCs by interpreting he AH as a psychologicalhierarchymerely pushes back the


32/38


Table2Pattern of Pronoun Retention in Relative Clauses

Language Subj DO 10 Obl Gen OCompAoban (North-East) (+) + + + + +Arabic - + + + + +Batak - 0 + + +Chinese (Pekingese) - +/- + + + +Czech (colloquial) - +/- + + + +Fulani (Gombe) - - * + +Genoese - (+) + + +Gilbertese - + + + + +Greek (Modern) - - +(?) +(?) + +Hausa - - (+) + +Hebrew - + + + + +Japanese - - - - +/-Javanese - - - - +Kera - + + + +Korean - - - - + 0Malay - - - - + *Minang-Kabau - - - -/+ + *Persian - (+) + + + +Roviana - - - - + 0Shona - - - (+) +(?) *Slovenian - + + + + +Turkish - - - - + +Urhobo + + + + + +Welsh - - + + + +Yoruba - - * * +Zurich German - - + + + +Key: + means that personalpronounsare normallypresentin that positionwhen it is relativized,usingthatRC-forming trategywhich admits of pronounretention. (+) means optionalretention.+/- means that insome cases the pronoun s retainedandin othersit is not. - meansthatpronounsare usuallynot retained.*means that that NP position does not naturallyexist in that language.0 means that that position is notrelativizable,and a blankmeans thatwe lack the relevantdata.An entryof the formx(?) meansthat ourdataare uncertainbutx is ourbest guess.

problemof explanationone step. We would still want to know why it is psychologicallyeasier to relativizesubjectsthan objects, etc. We have two speculationsto makehere.2.2.3.1. A Recognition Strategy. Impressionistically,the initial portion of the AHappearsto coincide with the degree to which NPs are requiredto appear in simplesentences. Thus, lexical predicates almost always require a subject. Many require


33/38


DOs, some require 1Os, and a few (e.g. put) require OBLs. Furthermore,a few verbs(R. Stockwell, P. Schachter, personal communication)such as gnash (one's teeth),blink (one's eyes), and water (his eyes watered) appear to require that their argumentsbe possessed body parts, similarto the more idiomatic constructions ose one's nerve,blow one's cool, etc. Finally, no lexical predicate in English requires that it beconstrued in a comparativeconstruction.Perhaps, then, there is a kind of universalrecognitionstrategyof the sort: If anNP plays a role in another clause, interpret t as a subject unless there are indicationsto the contrary, otherwise try the DO slot, etc. Needless to say, such a generalrecognition strategy would require a more precise formation and much experimentalresearch before it could constitute a seriousexplanationof the AH.2.2.3.2. Independent Reference. Another explanation, which pertains only to therelative accessibility of subjects over otherNPs, was offeredin Keenan(1974).Thereitwas argued hatheadsof RCs sharea logicalpropertywith subjectsof sentencesbut donot share this propertywith nonsubjects. Thus, more of what we need to know tounderstand the meaning of a RC formed on subjects is already contained in themeaningof simple sentences thanis the case when the RC is formed on a nonsubject.In more transformational erms, RC formationon subjects distortsthe meaning ofthe underlyingsentence less than RCs formedon objects.The logical propertysharedby heads andsubjectsis thatof independentreference.Thus, in simple sentences we cannot generallymake the referenceof the subject NPdependenton that of some other NP in the sentence. For instance, if subject and objectare marked as coreferential, t must be the object which is marked(if anything s), byfor instance a reflexive pronoun. Thus, the referenceof object phrases can be madedependent on that of subjects, but not conversely. But it is in the inherent nature ofsubjects to be independentlyreferring. For a weakening,but not abandonment,of thisprinciplefor more complex cases, see Keenan(1974).)Similarly,to understand he meaningof an RC such as the girl thatJohn likes, wemustbe able to understandwhat set is designatedby the head NP independentlyof thereference of the NPs in the restrictingclause. For instance, the head NP can neverbecoreferentiallypronominalizedby an NP in the restrictingclause, even if, as in manylanguages, the restricting clause precedes the head. But of course NPs within therestrictingclause can be stipulatedas, for instance, being identicalin referenceto thehead, as in (1), repeated below:

(65) der in seinemiBuro arbeitendeMannithe in his study working man'the manwho is working n his study'Consequently,if we relativize on a nonsubject, the resultingstructurewill contain twonecessarilyindependentlyreferringexpressions, the head of the RC and the subject ofthe restrictingclause. This explains, for instance, why the pronoun n the man who he


34/38


hit cannot be understoodas being necessarily coreferentialwith the head.3But if werelativize on a subject, there is only one necessarily independentlyreferring xpression.In this sense, then, subject relatives are psychologically simpler than nonsubjectrelatives.3. FurtherProspectsIn this article, we have restrictedourselves by and large to the particulardata areathatoriginally led us to propose the Accessibility Hierarchy, namely, restrictions onrelative clause formation.We wish here to indicate brieflytwo other areas where recentwork has shown the possible relevanceof the same Accessibility Hierarchy.In Comrie (in press), it is shown that the AH is useful in the syntacticdescriptionof causative constructions,in particular ynthetic causativeformations,in a variety oflanguages. Summarizing he argumentpresentedin the cited work, we may say that thesyntactic position used to encode the causee of a causative construction(i.e. theindividualcaused to carry out some action)will be the highest positionon the AH thatis not already occupied. The following French examples illustratethe generaltrend ofthe data:

(66) J'ai fait courir Henriette.'I made Henrietterun.'In (66), the causee is presented as a DO (Henriette);note that the constructionalreadyhas a subject.(67) J'ai fait manger es gateaux a Henriette.'I made Henriette eat the cakes.'Here, the causee is presented as an IO (a'Henriette); the constructionalreadyhas asubject (je) and a DO (les gateaux).

(68) J'ai fait ecrire une lettre au directeurpar Henriette.'I made Henriette write a letter to the director.'In (68), the causee is an OBL (parHenriette); there is alreadya subject (e), a DO (unelettre), and an IO (au directeur).A second extension of the rangeof applicabilityof the AH is in determining ross-language restrictions on advancementprocesses (Keenan and Comrie (1972);Perlmut-ter and Postal (1974); Johnson (1974b); Keenan (1975b); Trithart (1975)). By anadvancementprocess we shall mean a productive syntactic process that convertssentences containing NPs on a low position on the AH into roughly synonymous

I Similar facts obtain in languages in which the RC precedes the head. Thus in Japanese and Basque,for instance, in the equivalent of (the) he hit man (= the man that he hit ), the subject pronoun of themain verb in the RC cannot be understood as being coreferential with the head. See Keenan (1974) forexamples. Thus the Crossover Principle (Postal (1971)) is not the correct explanation for these cases.


35/38


36/38


37/38


Keenan, E. L. (1974) The Functional Principle: Generalizing the Notion of Subject-of, in M.W. LaGaly, R. Fox, and A. Bruck, eds. (1974), 298-309.

Keenan, E. L. (1975a) Logical Expressive Power and Syntactic Variation in Natural Lan-guage, in E. L. Keenan, ed., Formal Semantics of Natural Language, CambridgeUniversity Press, Cambridge.Keenan, E. L. (1975b) Some Universals of Passive in Relational Grammar, in R. E.Grossman, L. J. San, and T. J. Vance, eds. (1975), 340-352.

Keenan, E. L. (1976a) Remarkable Subjects in Malagasy, in C. Li, ed. (1976).Keenan, E. L. (1976b) Towards a Universal Definition of Subject of, in C. Li, ed. (1976).Keenan, E. L. and B. Comrie (1972) Noun Phrase Accessibility and Universal Grammar,Winter Meeting, LSA.Lafitte, P. (1962) Grammaire Basque, Editions des Amis du Musee Basque et lkas ,Bayonne.LaGaly, M. W., R. Fox, and A. Bruck, eds. (1974) Papers from the Tenth Regional Meeting ofthe Chicago Linguistic Society, University of Chicago, Chicago, Illinois.Legum, S. (1975) Strategies in the Acquisition of Relative Clauses, Technical Note 2-75-10,Southwest Regional Laboratory, Los Alamitos, California.Li, C., ed. (1976) Subject and Topic, Academic Press, New York.Li, C. and S. Thompson (1974) Subject and Topic: A New Typology of Language, Winter

Meeting, LSA.Li, C. and S. Thompson (1976) Evidence Against Topicalization in Topic Prominent Lan-guages, in C. Li, ed. (1976).MacDonald, R. and D. Soenjono (1967) Indonesian Reference Grammar, Georgetown Univer-sity Press, Washington, D.C.Peranteau, P. M., J. N. Levi, and G. C. Phares, eds. (1972) The Chicago Which Hunt, ChicagoLinguistic Society, Chicago.Perlmutter, D. M. and P. M. Postal (1974) Linguistic Institute Lectures, unpublished ms.Postal, P. M. (1971) Crossover Phenomena, Holt, Rinehart and Winston, New York.Schachter, P. (1973) Focus and Relativization, Language 49, 19-46.Schachter, P. (1976) The Subject in Philippine Languages: Topic, Actor, Actor-Topic, or Noneof the Above, in C. Li, ed. (1976).Sheldon, A. (1974) The Role of Parallel Function in the Acquisition of Relative Clauses in

English, Journal of Verbal Learning and Verbal Behavior 13, 272-281.Silitonga, M. (1973) Some Rules Reordering Constituents and their Constraints in Batak,unpublished Doctoral dissertation, University of Illinois, Urbana, Illinois.Tagashira, Y. (1972) Relative Clauses in Korean, in P. M. Peranteau, J. N. Levi, and G. C.

Phares, eds. (1972).Tchekhoff, C. (1973) Some Verbal Patterns in Tongan, Journal of the Polynesian Society 82,281-292.Trithart, L. (1975) Relational Grammar and Chicewa Subjectivization Rules, in R. E.Grossman, L. J. San, and T. J. Vance, eds. (1975), 615-624.Valli, H., N. Hernandez, M. Archard-Baule, and M. F. Beretti (1972) Compte-rendu d'uneexperience realisee dans une classe de 6e 2, dont le but etait d'etudier les mecanismes dela production des relatives chez l'enfant, Institut de Didactique et Pedagogie, Universitede Provence.Woodbury, A. (1975) Ergativity of Grammatical Processes: A Study of Greenlandic Eskimo, M. A.

Essay, Department of Linguistics, University of Chicago, Chicago, Illinois.


38/38


KeenanDepartment of LinguisticsUCLALos Angeles, California 90024ComrieDepartment of LinguisticsSidgwick AvenueCambridge CB3 9DAEngland

keenan com rie

Documents