Relative Clauses in Hindi and Arabic:A Paninian Dependency Grammar Analysis

Mark Pedersen * †† School of Information Technology &

Electrical EngineeringUniversity of QueenslandSt. Lucia 4072, Australia

markp@itee.uq.edu.au

Domenyk Eades ‡‡ Department of English,

Faculty of Arts,Sultan Qaboos University,Muscat, PC 123 Omandomenyk@squ.edu.om

Samir K. Amin andLakshmi Prakash ∗

∗Faculty of Applied Science,Sohar University,

Sohar, PC 311 Omans.amin@soharuni.edu.om,

l.prakash@soharuni.edu.om

Abstract

We present a comparative analysis of relativeclauses in Hindi and Arabic in the traditionof the Paninian Grammar Framework (Bharatiet al., 1996b) which leads to deriving a com-mon logical form for equivalent sentences. Par-allels are drawn between the Hindi co-relativeconstruction and resumptive pronouns in Ara-bic. The analysis arises from the developmentof lexicalised dependency grammars for Hindiand Arabic that have application for machinetranslation.

1 Introduction

(Kruijff, 2002) notes that Dependency Gram-mar has its roots in Panini’s grammar of San-skrit (350/250 BC) and also the work of earlyArabic grammarians (Kitab al-Usul of Ibn al-Sarraj, d. 928). Among the recent activities inDependency Grammar, (Bharati et al., 1996b)have established a computational approach toIndian languages which they call the PaninianGrammar Framework (PGF).

Bharati et. al. (ibid) suggest that PGF is ex-tensible to other languages, including fixed wordorder languages such as English. In consideringa Machine Translation system for Hindi-Arabic,and given the availability of a PGF-style parserfor Hindi (Pedersen, 2001), we have sought toestablish the suitability of PGF for Arabic.

In the following sections we will briefly de-scribe the general PGF-inspired parsing frame-work, and then contrast the analysis of Hindiand Arabic relative clauses within this frame-work. In particular, we examine parallels be-tween the Hindi co-relative construction and re-sumptive pronouns in Arabic, and demonstratehow a common logical interpretation can begiven to syntactic variations of relative clausesin both languages.

2 Dependency Grammar in ThePaninian Grammar Framework

2.1 BackgroundThe modern formulation of Dependency Struc-ture is frequently attributed to (Tesniere, 1959).It is interesting to note that among contem-porary proponents of dependency structure,there are those, such as (Hudson, 1984), whomaintain a general principle of projectivity fortheir dependency structures, and devise addi-tional means of coping with discontinuity whenit arises. Others, such as (Mel’cuk, 1988),(Bharati et al., 1996b) and (Covington, 1990)allow non-projective dependency structures andrely upon a separate means of linearisation.Most recently, (Debusmann and Duchier, 2003)have presented a new formulation of depen-dency grammar which generalises multi-stratalapproaches to an n-dimensional formalism of in-teracting dependency graphs.

The Paninian Grammar Framework proposedby (Bharati et al., 1996b) is particularly aimedat treating heavily inflected free word orderlanguages such as Hindi and other Indian lan-guages. Like Hindi, Arabic is heavily inflectedwith overt case marking of nouns, noun-verbagreement, as well as incorporation of pronom-inals into verb forms. Although Arabic has aword order (VSO) that is more fixed comparedto Hindi (canonically SOV but with significantword order freedom), there is also significantword order variation found in nominal and top-icalised sentences, thus making alternate wordorders such as SVO quite common.

2.2 The Paninian Grammar FrameworkPGF has two levels of representation that medi-ate between an utterance and its meaning: thevibhakti level and the karaka level. Figure 1shows the relationship between the levels of rep-resentation in Paninian Grammar.

Let us take the surface level of of Figure 1to represent the level of tokenised and morpho-

Semantic Level

Karaka Level

Vibhakti Level

Surface Level

Figure 1: Levels of representation in PGF

logically analysed input to a syntactic-semanticparsing process. Vibhakti is a Sanskrit gram-matical term that encompasses postpositionalsand case endings for nouns, as well as inflec-tion and auxiliaries for verbs. The vibhakti levelgroups words together according to explicit caseendings and other inflectional markers. Vib-hakti word groups are necessarily contiguousand typically have a fixed word order, but vib-hakti groups can generally appear in any or-der without affecting the karaka relations whichconnect them.

The karaka level serves to relate the lexical el-ements of a sentence according to their syntacticfunctions, as derivable from their vibhakti. Interms of more familiar linguistic theory, we cangenerally equate the vibhakti level with mor-photactics and the karaka level with syntactico-semantic functions (e.g. theta relations). Thesemantic level shown in figure 1 is indicative of apurely semantic representation or logical form,the details of which are beyond the scope of thispaper.

Panini recognises six type of karaka relations(Kiparsky and Staal, 1969):

• karta (agent)

• karma (patient)

• karan (instrument)

• sampradan (recipient)

• apaadaan (point of departure, or cause)

• adhikaran (locality)

It is important to note that karaka rela-tions differ from typical theta relations in thatthey tend to have more fluid definitions ratherthan fixed syntactico-semantic roles. For in-stance, the karta relation, which may generally

be equated to the agent theta role, is more pre-cisely defined as “the most independent of allparticipants” (Bharati et al., 1996b, p. 187).The karta relation can equally be applied tothe subject of all of the following sentences, al-though karta should not be confused with thepurely syntactic role of subject:

(1) (a) Mohan opened the lock.

(b) The key opened the lock.

(c) The lock opened.

Karaka relations under consideration here arekarta, as described, and karma, the principlegoal of the karta (roughly, patient). Karakarelations do not map exactly on to the typi-cal semantic relation categories common in con-temporary Western linguistic theory, largely be-cause of the differing conceptions of semanticroles underpinning Paninian logical form1.

A further difference between karaka relationsand typical theta roles is that, traditionally, vib-hakti is the primary means of identifying whichkaraka relations may apply2. This character-istic lends itself well to heavily inflected lan-guages with free word order. A simple de-fault karaka chart (adapted from (Bharati et al.,1996b)) showing the mapping between vibhaktiand karaka is given in Table 1.

Karaka Vibhakti Presencekarta NOM (∅) mandatory

num & gen agr.karma ACC (ko or ∅) mandatory

Table 1: Default Karaka chart

PGF elegantly handles active-passive trans-formations and compound sentences through asystem of karaka transformations, which changethe default vibhakti required by a differentlyconjugated verb for a given karaka. If we wereto use the passive form of khaataa - khayaa , thevibhakti required for the karta karaka becomesne, as illustrated in (2).

1Paninian semantics makes use of the Indian systemsof philosophy and logic. Among these, Navya Nyayalogic was of particular significance. For more detail onNavya Nyaya, see (Ingalls, 1951) and (Matilal, 1968).

2Panini’s grammar of Sanskrit asserted that everykaraka relation in an utterance must have a phonetic re-alisation (Bharati et al., 1996b, p. 187), rather than viaan intermediate syntactic role or sentential position. Inthis sense, there is a tighter binding between karaka re-lations and the surface-level syntax than would normallybe seen in a typical theta role analysis.

Despite its orientation toward inflected lan-guages, some work has been done on gener-alising the notion of vibhakti to include lin-ear position where this has syntactic signifi-cance (Bharati et al., 1996a). The generalisa-tion of vibhakti to account for word order essen-tially introduces a facility analogous to topolog-ical fields (as found in Topological DependencyGrammar (Duchier and Debusmann, 2001) andDACHS (Broker, 1998)) to PGF. It is impor-tant to note however that these word order con-straints are treated in the same way as othervibhakti (morphological constraints), and donot use a separate representation. In this sense,PGF does not attempt to separate linear prece-dence from immediate dominance at a formallevel.

In the following sections we explore how bothmorphology and word order constraints in Ara-bic can be captured through a mapping of vib-hakti to karaka relations.

3 Relative Clauses in Hindi andArabic

To briefly summarise salient features of Hindiand Arabic:

• Hindi word order is relatively free

• Arabic word order is more fixed, but top-icalisation and nominal sentence forms ef-fectively license significant permutations inword order

• both have complex morphology, but

• most case marking in Hindi uses both in-flection and post positions

• whereas Arabic generally uses inflectiononly

3.1 DataIn comparing relative clauses, let us considerthe data shown in (2) and (3)3. The sentencesgiven are intended to represent the same seman-tic content, but give different emphases through

3In 3(a) and (c), the resumptive pronouns -hu and-ha are glossed as -3.M and -3.F, meaning third personpronoun masculine and third person pronoun feminine,respectively. Throughout the examples, gender markingon verbs indicates agreement with the relevant argument.

topicalisation4. It should be noted that someutterances, such as 2(b), would only be used inspecific context and don’t represent the normalspeech pattern. Nevertheless all are consideredto be grammatical by native speakers and fol-low in the same pattern as data presented by(Dwivedi, 1994).

3.2 PGF analysisThe immediately observable difference betweenthe Hindi and Arabic is that the Hindi data per-mits topicalisation primarily through changes inword order and a limited number of changes invibhakti. The syntactic structure of 2(b) and(c) are isomorphous modulo word order. In2(a), machli is internal to the relative clause,and hence vo is pronominal rather than demon-strative. This requires two changes at the vib-hakti level: khayi thi agrees in gender with theexplicitly present female object, rather than themale subject, because of the ergative construc-tion5; and because of its position as a syntac-tic object, the ergative marking on machli isdropped.

In contrast, topicalisation in Arabic requiresa variety of different syntactic constructions andassociated changes in vibhakti, since the wordorder is not as free. In 3(a) we find the equiv-alent Arabic sentence to 2(a) preposes the sub-ject of the main verb, which is a permissiblevariation on the standard VSO word order.

However in 3(b), the sentence must be recon-structed to allow Zayd to occupy a topic po-sition. We can neither extract Zayd from therelative clause, nor have the relative clause asthe topic, as it appears in 2(b). Therefore, Zay-dun ↩akala ↩al-samakah becomes the main clauseof the sentence.

For the final topicalisation, Arabic takes onthe nominal sentence form, in which there isno main verb. Instead, ↩al-d. ifda↪ acts as thepredicate, to which nested relative clauses areattached. Analyses for the Hindi and Arabicsamples are shown in 4(a)-(c) and 5(a)-(c) re-spectively.

For the purposes of this analysis, we bor-4For the purposes of this paper we use the term top-

icalisation broadly to refer to the phenomenon of lexical“movement” to a sentence initial position without distin-guishing between various types of such movement, suchas those discussed by (Gambhir, 1981), (Dwivedi, 1994)and (Butt and Holloway-King, 1997).

5In the other example sentences, the agreement re-verts to the default male gender, since the direct objecthas been extracted.

(2) (a) joREL

machlifish.F

Zayd neZayd-ERG

khayi thieat-PAST.F

voCO-REL

maindek kofrog.M-ACC

khayaa thaeat-PAST.M

The fish that was eaten by Zayd had eaten a frog.

(b) joREL

Zayd neZayd-ERG

khayaa thaeat-PAST.M

voCO-REL

machli nefish.F-ERG

maindek kofrog.M-ACC

khayaa thaeat-PAST.M

The fish that was eaten by Zayd had eaten a frog.

(c) maindek kofrog.M-ACC

voCO-REL

machli nefish.F-ERG

khayaa thaeat-PAST.M

joREL

Zayd neZayd-ERG

khayaa thaeat-PAST.M

The fish that was eaten by Zayd had eaten a frog.

(3) (a) ↩al-samakahDEF-fish.F

allatyREL.F

↩akala-haate.M-3.F

Zayd-unZayd-NOM

↩akalatate.F

↩al-d. ifda↪DEF-frog.M

The fish that was eaten by Zayd, had eaten a frog.

(b) Zayd-unZayd-NOM

↩akalaate.M

↩al-samakahDEF-fish.F

allatyREL.F

↩akalatate.F

↩al-d. ifda↪DEF-frog.M

Zayd ate the fish that had eaten the frog.

(c) ↩al-d. ifda↪DEF-frog.M

↩alladhyREL.M

↩akalat-huate.F-3.M

↩al-samakahDEf-fish.F

allatyREL.F

↩akala-haate.M-3.F

Zayd-unZayd-NOM

The frog had been eaten by the fish that Zayd ate.

row the term avachchedak (limiter) from NavyaNyaya logic to express the relationship betweenthe relative clause and the noun being modi-fied, which is the avachchinna - “the limited”.From a PGF perspective, the data is explainedas follows.

3.3 Discussion3.3.1 Hindi AnalysisIn Hindi, ergative marking of the subject is re-quired by the yaa form of the verb, as shown ear-lier, except in the case of pronominal vo which isnot explicitly marked with ne. The subject forboth clauses is thus clearly identified (either itis marked with ne or it is the pronominal vo). Ifword order were completely free, both the mainverb and the complement would be candidateheads for the ergatively marked verbs. For 2(a)and 2(b), the co-relative vo introduces a projec-tivity constraint which removes the ambiguity.In 2(a), the pronominal co-relative vo takes theplace of the object of the relative clause machli.In 2(b), the object of the relative clause is ab-sent (the typical constituency analysis wouldsay that it leaves a trace), but is marked by thedemonstrative co-relative vo in the main clause,establishing the connection to the relativiser jo.Both of these represent the left-adjoined formof the relative clause.

The relationship between the relativiser joand the co-relative vo is implicit, since the directrelationship between jo and the modified nounmerely marks the noun, in the case of 2(a), as

the avachchinna - the item limited by the rel-ative clause. In practical terms, the presenceof an argument marked as avachchinna is prop-agated up to the head of the relative clause,marking it as the avachchedak limiter, meaningthat the verb cannot take anything other thanthe co-relative vo as its head. This analysis is inkeeping with the requirement that jo must havea matching vo in the sentence.

The same analysis applies when the modi-fied noun is not explicitly present in the relativeclause. In this case, jo still signifies the presenceof an avachchinna argument, but it is pronom-inal, and the property of being avachchinna isconferred upon the modified noun by vo. Inthis way, there is a pleasing symmetry betweenvo and jo in that when they depend on a noun,they act as demonstratives, marking the nounas avachchinna, and when appearing indepen-dently, they act as pronouns in place of theavachchinna noun.

In 2(c), we have the right adjoined form of therelative clause. Additionally, the typical SOVword order of the main clause has been alteredto OSV through topicalising the object, main-dek. The co-relative vo continues to assert aprojectivity constraint on the relative clause.

A further constraint on right-adjoined rela-tive clauses is that they may not contain theexplicit subject. This constraint is enforced viathe karaka relation between the avachchedak-marked verb and the pronominal co-relative vo6

6The demonstrative co-relative vo marks an explicit

(4) (a)

jo

Rel

machli

Karma

Zayd ne

Karta

kaya thi

Avachchedak

vo

Karta

maindek ko

Karma

kaya tha

(b)

jo

Karma

Zayd ne

Karta

kaya thi

Avachchedak

vo

CoRel

machli

Karta

maindek ko

Karma

kaya tha

(c)

maindek ko

Karma

vo

CoRel

machli

Karta

kaya tha

jo

Karma

Zayd ne

Karta

kaya thi

Avachchedak

(5) (a)

'al-samakah

Karta

allaty

Rel

'akala

Avachchedak

-ha

Karma

Zaydun

Karta

'akalat

'al-difda'

Karma

(b)

Zaydun

Karta

'akala

'al-samakah

Karma

allaty

Rel

'akalat

Avachchedak

'al-difda'

Karta

(c)

'al-difda

alladhy

Rel

'akalat

Avachchedak

-hu

Karma

'al-samakah

Karta

allaty

Rel

'akala

Avachchedak

-ha

Karma

Zaydun

Karta

requiring jo to always precede vo7, thus ensur-ing that the pronominal vo never precedes itsreferent.

noun, while the pronominal vo does not, and hence wetreat these separate lexical entries.

7This is an example of a generalisation of the karakachart to include word order constraints.

3.3.2 Arabic AnalysisTo account for the Arabic data, we first must es-tablish the relevant vibhakti to karaka mapping.The syntactic role of inflection in Arabic is welldefined (cf. (Fischer, 2002)), and the relevantrules (for the examples under consideration) can

be summarised as follows:

• Nominative case: (-un or -u) marks thesubject

• Verbs agree in number and gender with thesubject

• Where a verbal argument is absent, apronominal affix (e.g. -ha) is attached tothe verb, which will agree in number andgender with the missing item.

• Relativisers (e.g. allaty) agree in genderwith the noun being modified.

Given that the karta relation does not neces-sarily require an explicit subject, in absence ofexplicit marking of the subject, the relationshipcan still be derived from gender agreement8.Furthermore, the subject will always precedethe object. In the sense that vibhakti can begeneralised to include word order (Bharati etal., 1996a), we may also include this conditionas part of the karaka chart (see Table 2). Asin Hindi, the relativisers impose projectivity ontheir dependents.

Karaka Vibhakti Presencekarta (NOM (-uN/-u) mandatory

or gen agr).and precedes object(if present)

karma ACC (-aN/-a) optional

Table 2: Default Karaka chart for Arabic

This set of rules is sufficient to account forthe grammaticality of the data given here, butdoes not account for the semantic relationshipbetween the three sentences. In this respect,the karaka relations have so far only illustratedtheir connection to the purely syntactic level ofanalysis. While we do not wish to regress tosome kind of transformational account of thesurface forms, it is desirable to illustrate thatthe karaka relations provide sufficient abstrac-tion to permit the formulation of a common se-mantic representation.

One way to approach this is to argue thatrelativisers ‘mediate’ the appropriate karaka re-lation from the complement to the noun be-ing modified. This is essentially a process ofunifying the embedded resumptive pronoun in

8This is also true in Hindi.

the complement with the relevant external argu-ment. To make this more explicit, we have sepa-rated the resumptive pronoun from the comple-ment in the PGF analysis and shown the karakarelation that exists between them.

Thus in 3(a), allaty mediates a karma relationbetween al-samakah and ↩akala-ha. In 3(c), thisconnection is repeated in the relative clause, and↩alladhy mediates a karma relationship betweenal-d. ifda↩ and ↩akalat-hu. In 3(b), there is no ex-plicit resumptive pronoun, since this feature isonly used for a missing object. Instead, genderand number agreement between al-samakah andakalat means that the karta relationship is ob-tained via allaty. This mediation of karaka rela-tions is illustrated by a typical feature structureunification diagram shown in (6).

Even though the feature structures differ interms of argument structure, they can be givenan equivalent semantic interpretation in termsof lambda abstraction:

• We can represent 6(a) as:ate(Zayd, λx[fish′(x) ∧ ate(x,frog)]

• Likewise, 6(b) is represented by:ate(λx[fish′(x) ∧ (Zayd, x)],frog)

• Both of these can be reduced to:λx[fish′(x) ∧ ate(Zayd, x) ∧ ate(x,frog)]

• 6(c) merely asserts an additional variable,

λy[ate(Zayd, λx[fish′(x)∧ate(x,frog′(y))])],

which can be factored out, since it is theidentity function of frog, thus leaving 6(c)with the same interpretation as 6(a).

The same mediation process can be applied tothe Hindi data, using the same argument. Here,the relative and (pronominal) co-relative guidethe unification. The karaka relation between the(pronominal) relativiser jo and the avachchedakverb is mapped onto the avachchinna nounmarked by vo or directly onto the pronominalco-relative vo, as the case may be. In this sense,the pronominal relative or co-relative operatesin the same was the resumptive pronouns inArabic.

After unification, all three Hindi sentencesshare the same feature structure, shown in 7,and hence further analysis to demonstrate theirsemantic equivalence is not needed.

(6) (a)

PRED akala〈KARTA〉〈KARMA〉

KARTA

PRED samakah

AVACHCH

PRED akala〈KARTA〉〈KARMA〉KARTA [Zayd]KARMA [samakah]

KARMA[PRED d. ifda↪

]

(b)

PRED akala〈KARTA〉〈KARMA〉KARMA

[PRED Zayd

]

KARMA

PRED samakah

AVACHCH

[PRED akala〈KARTA〉〈KARMA〉KARMA [samakah]

]

KARMA [d. ifda↪]

(c)

PRED d. ifda↪〈AVACHCH〉

AVACHCH

PRED akala〈KARTA〉〈KARMA〉

KARTA

PRED samakah

AVACHCH

PRED akala〈KARTA〉〈KARMA〉KARTA [Zayd]KARMA [samakah]

KARMA[PRED d. ifda↪

]

(7) (a)

PRED kaya〈KARTA〉〈KARMA〉

KARTA

PRED machli

AVACHCH

PRED kaya〈KARTA〉〈KARMA〉KARTA [Zayd]KARMA [machli]

KARMA[PRED maindek

]

4 Conclusions and Future Work

We have sketched an application of the PaninianGrammar Framework to Arabic, and outlinedan approach to deriving a common logical formfor equivalent sentences in Hindi and Arabic. Inparticular, the PGF analysis given here high-lights the similarity of resumptive pronoun affixin Arabic to the co-relative in Hindi.

This is a first step toward developing com-prehensive Paninian dependency grammars ofArabic and Hindi, with a view to applying thegrammars to machine translation. Further de-velopment of the grammar is necessary before it

is clear that PGF is capable of handling all therequirements of a MT system. In particular, thesuitability of PGF for generation needs to be ex-plored, particularly with respect to generatingappropriate word orders in the target language.

In terms of on-going development of PGF as atheory for computational linguistics, (Bharati etal., 1996b) and (Pedersen, 2001) have comparedPGF to Lexical Functional Grammar (Bresnanand Kaplan, 1982) and Lexicalised Tree Adjoin-ing Grammar (Joshi, 1987) with encouragingresults. However a systematic comparison ofPGF to more recent dependency grammar for-malisms, such as DACHS and TDG, has not yet

been done. Given the strong parallels with re-cent work in these formalisms (cf. (Kruijff andDuchier, 2003)), such an investigation is nowessential.

5 Acknowledgements

Our thanks go to Professor Rajeev Sangal andthe Language Technology Research Centre staffat the International Institute of InformationTechnology, Hyderabad, for their ongoing sup-port of our efforts in applying the PaninianGrammar Framework, Petr Pajas for assistancewith the TrEd diagramming tool used for lay-ing out the dependency diagrams, and to Pro-fessor Joachim Diederich and the workshop re-view panel for their helpful comments duringthe preparation of this paper.

References

A. Bharati, M. Bhatia, V. Chaitanya, andR. Sangal. 1996a. Paninian Grammar Frame-work Applied to English. Technical ReportTRCS-96-238, CSE, IIT Kanpur.

A. Bharati, V. Chaitanya, and R. Sangal.1996b. Natural Language Processing - APaninian Perspective. Prentice Hall of India,New Delhi.

J. Bresnan and R. Kaplan. 1982. Lexical-functional grammar: A formal system forgrammatical representation. In J. Bresnan,editor, The Mental Representation of Gram-matical Relations. MIT Press.

N. Broker. 1998. Separating surface order andsyntactic relations in a dependency grammar.In Proceedings of COLING-ACL ’98, pages174–180.

Miriam Butt and Tracy Holloway-King. 1997.Null elements in discourse structure. In K. V.Subbarao, editor, Papers from the NULLSSeminar. Motilal Banarasidas, Delhi.

M. Covington. 1990. Parsing Discontinu-ous Constituents in Dependency Grammar.Computational Linguistics, 16(4), December.

Ralph Debusmann and Denys Duchier. 2003. Ameta-grammatical framework for dependencygrammar. Technical report, Universitat desSaarlande, Saarbrucken, Germany.

Denys Duchier and Ralph Debusmann. 2001.Topological dependency trees: A constraint-based account of linear precedence. In Meet-ing of the Association for Computational Lin-guistics, pages 180–187.

V. Dwivedi. 1994. Topicalization in Hindi andthe correlative construction. In Miriam Butt,

Tracy Holloway King, and Gillian Ramchand,editors, Theoretical Perspectives on word or-der in South Asian languages. CSLI Publica-tions, Stanford, CA.

Wolfdietrich Fischer. 2002. A Grammar ofClassical Arabic. Yale University Press, Newhaven and London.

V. Gambhir. 1981. Syntactic Restrictions andDiscourse Functions of Word Order in Stan-dard Hindi. PhD Thesis, University of Penn-sylvania, Philadelphia.

R. Hudson. 1984. Word Grammar. Basil Black-well, 108 Cowley Rd, Oxford OX4 1JF, Eng-land.

D. Ingalls. 1951. Materials for the Studyof Navya-Nyaya Logic. Harvard UniversityPress, Cambridge.

A. K. Joshi. 1987. An Introduction to TreeAjoining Grammars. In A. Manaster-Ramer,editor, Mathematics of Language. John Ben-jamins, Amsterdam.

P. Kiparsky and J. F. Staal. 1969. Syntacticand Semantic Relations in Panini. Founda-tions of Language, 5:84–117.

Geert-Jan M. Kruijff and Denys Duchier. 2003.Information structure in topological depen-dency grammar. In EACL 2003, 11th Con-ference of the European Chapter of the Asso-ciation for Computational Linguistics, pages219–226.

Geert-Jan M. Kruijff. 2002. Formal and com-putational aspects of dependency grammar:History and development of dg. Technical re-port, ESSLI2002.

B.K. Matilal. 1968. The Navya-Nyaya Doctrineof Negation. Harvard University Press, Cam-bridge.

I. A. Mel’cuk. 1988. Dependency Syntax: The-ory and Practice. State University Press ofNew York.

M. Pedersen. 2001. Usability Evaluation ofGrammar Formalisms for Free Word OrderNatural Language Processing. PhD thesis,School of Computer Science and ElectricalEngineering, University of Queensland, Bris-bane, Australia.

L. Tesniere. 1959. Elements de syntaxe struc-turale. Klincksieck, Paris.