<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-1502">
  <Title>Relative Clauses in Hindi and Arabic: A Paninian Dependency Grammar Analysis</Title>
  <Section position="3" start_page="0" end_page="8" type="metho">
    <SectionTitle>
2 Dependency Grammar in The
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
Paninian Grammar Framework
2.1 Background
</SectionTitle>
      <Paragraph position="0"> The modern formulation of Dependency Structure is frequently attributed to (Tesni`ere, 1959). It is interesting to note that among contemporary proponents of dependency structure, there are those, such as (Hudson, 1984), who maintain a general principle of projectivity for their dependency structures, and devise additional means of coping with discontinuity when it arises. Others, such as (Mel'Vcuk, 1988), (Bharati et al., 1996b) and (Covington, 1990) allow non-projective dependency structures and rely upon a separate means of linearisation.</Paragraph>
      <Paragraph position="1"> Most recently, (Debusmann and Duchier, 2003) have presented a new formulation of dependency grammar which generalises multi-stratal approaches to an n-dimensional formalism of interacting dependency graphs.</Paragraph>
      <Paragraph position="2"> The Paninian Grammar Framework proposed by (Bharati et al., 1996b) is particularly aimed at treating heavily inflected free word order languages such as Hindi and other Indian languages. Like Hindi, Arabic is heavily inflected with overt case marking of nouns, noun-verb agreement, as well as incorporation of pronominals into verb forms. Although Arabic has a word order (VSO) that is more fixed compared to Hindi (canonically SOV but with significant word order freedom), there is also significant word order variation found in nominal and topicalised sentences, thus making alternate word orders such as SVO quite common.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="2" type="sub_section">
      <SectionTitle>
2.2 The Paninian Grammar Framework
</SectionTitle>
      <Paragraph position="0"> PGF has two levels of representation that mediate between an utterance and its meaning: the vibhakti level and the karaka level. Figure 1 shows the relationship between the levels of representation in Paninian Grammar.</Paragraph>
      <Paragraph position="1"> Let us take the surface level of of Figure 1 to represent the level of tokenised and morpho- null logically analysed input to a syntactic-semantic parsing process. Vibhakti is a Sanskrit grammatical term that encompasses postpositionals and case endings for nouns, as well as inflection and auxiliaries for verbs. The vibhakti level groups words together according to explicit case endings and other inflectional markers. Vibhakti word groups are necessarily contiguous and typically have a fixed word order, but vibhakti groups can generally appear in any order without affecting the karaka relations which connect them.</Paragraph>
      <Paragraph position="2"> The karaka level serves to relate the lexical elements of a sentence according to their syntactic functions, as derivable from their vibhakti. In terms of more familiar linguistic theory, we can generally equate the vibhakti level with morphotactics and the karaka level with syntactico-semantic functions (e.g. theta relations). The semantic level shown in figure 1 is indicative of a purely semantic representation or logical form, the details of which are beyond the scope of this paper.</Paragraph>
      <Paragraph position="3">  It is important to note that karaka relations differ from typical theta relations in that they tend to have more fluid definitions rather than fixed syntactico-semantic roles. For instance, the karta relation, which may generally be equated to the agent theta role, is more precisely defined as &amp;quot;the most independent of all participants&amp;quot; (Bharati et al., 1996b, p. 187). The karta relation can equally be applied to the subject of all of the following sentences, although karta should not be confused with the purely syntactic role of subject:  (1) (a) Mohan opened the lock.</Paragraph>
      <Paragraph position="4"> (b) The key opened the lock.</Paragraph>
      <Paragraph position="5"> (c) The lock opened.</Paragraph>
      <Paragraph position="6">  Karaka relations under consideration here are karta, as described, and karma, the principle goal of the karta (roughly, patient). Karaka relations do not map exactly on to the typical semantic relation categories common in contemporary Western linguistic theory, largely because of the differing conceptions of semantic roles underpinning Paninian logical form  .</Paragraph>
      <Paragraph position="7"> A further difference between karaka relations and typical theta roles is that, traditionally, vibhakti is the primary means of identifying which karaka relations may apply  . This characteristic lends itself well to heavily inflected languages with free word order. A simple default karaka chart (adapted from (Bharati et al., 1996b)) showing the mapping between vibhakti and karaka is given in Table 1.</Paragraph>
      <Paragraph position="8">  PGF elegantly handles active-passive transformations and compound sentences through a system of karaka transformations, which change the default vibhakti required by a differently conjugated verb for a given karaka. If we were to use the passive form of khaataa - khayaa ,the vibhakti required for the karta karaka becomes ne, as illustrated in (2).</Paragraph>
      <Paragraph position="9">  Paninian semantics makes use of the Indian systems of philosophy and logic. Among these, Navya Nyaya logic was of particular significance. For more detail on Navya Nyaya, see (Ingalls, 1951) and (Matilal, 1968).  Panini's grammar of Sanskrit asserted that every karaka relation in an utterance must have a phonetic realisation (Bharati et al., 1996b, p. 187), rather than via an intermediate syntactic role or sentential position. In this sense, there is a tighter binding between karaka relations and the surface-level syntax than would normally be seen in a typical theta role analysis.</Paragraph>
      <Paragraph position="10"> Despite its orientation toward inflected languages, some work has been done on generalising the notion of vibhakti to include linear position where this has syntactic significance (Bharati et al., 1996a). The generalisation of vibhakti to account for word order essentially introduces a facility analogous to topological fields (as found in Topological Dependency Grammar (Duchier and Debusmann, 2001) and DACHS (Br&amp;quot;oker, 1998)) to PGF. It is important to note however that these word order constraints are treated in the same way as other vibhakti (morphological constraints), and do not use a separate representation. In this sense, PGF does not attempt to separate linear precedence from immediate dominance at a formal level.</Paragraph>
      <Paragraph position="11"> In the following sections we explore how both morphology and word order constraints in Arabic can be captured through a mapping of vibhakti to karaka relations.</Paragraph>
    </Section>
    <Section position="3" start_page="2" end_page="4" type="sub_section">
      <SectionTitle>
3.1 Data
</SectionTitle>
      <Paragraph position="0"> In comparing relative clauses, let us consider the data shown in (2) and (3)  . The sentences given are intended to represent the same semantic content, but give different emphases through  In 3(a) and (c), the resumptive pronouns -hu and -ha are glossed as -3.M and -3.F, meaning third person pronoun masculine and third person pronoun feminine, respectively. Throughout the examples, gender marking on verbs indicates agreement with the relevant argument. topicalisation  . It should be noted that some utterances, such as 2(b), would only be used in specific context and don't represent the normal speech pattern. Nevertheless all are considered to be grammatical by native speakers and follow in the same pattern as data presented by (Dwivedi, 1994).</Paragraph>
    </Section>
    <Section position="4" start_page="4" end_page="5" type="sub_section">
      <SectionTitle>
3.2 PGF analysis
</SectionTitle>
      <Paragraph position="0"> The immediately observable difference between the Hindi and Arabic is that the Hindi data permits topicalisation primarily through changes in word order and a limited number of changes in vibhakti. The syntactic structure of 2(b) and (c) are isomorphous modulo word order. In 2(a), machli is internal to the relative clause, and hence vo is pronominal rather than demonstrative. This requires two changes at the vibhakti level: khayi thi agrees in gender with the explicitly present female object, rather than the male subject, because of the ergative construction null  ; and because of its position as a syntactic object, the ergative marking on machli is dropped.</Paragraph>
      <Paragraph position="1"> In contrast, topicalisation in Arabic requires a variety of different syntactic constructions and associated changes in vibhakti, since the word order is not as free. In 3(a) we find the equivalent Arabic sentence to 2(a) preposes the sub-ject of the main verb, which is a permissible variation on the standard VSO word order.</Paragraph>
      <Paragraph position="2"> However in 3(b), the sentence must be reconstructed to allow Zayd to occupy a topic position. We can neither extract Zayd from the relative clause, nor have the relative clause as the topic, as it appears in 2(b). Therefore, Zaydun arrowhookrightakala arrowhookrightal-samakah becomes the main clause of the sentence.</Paragraph>
      <Paragraph position="3"> For the final topicalisation, Arabic takes on the nominal sentence form, in which there is no main verb. Instead, arrowhookrightal-d .</Paragraph>
      <Paragraph position="4"> ifdaarrowhookleft acts as the predicate, to which nested relative clauses are attached. Analyses for the Hindi and Arabic samples are shown in 4(a)-(c) and 5(a)-(c) respectively. null For the purposes of this analysis, we bor- null For the purposes of this paper we use the term topicalisation broadly to refer to the phenomenon of lexical &amp;quot;movement&amp;quot; to a sentence initial position without distinguishing between various types of such movement, such as those discussed by (Gambhir, 1981), (Dwivedi, 1994) and (Butt and Holloway-King, 1997).</Paragraph>
      <Paragraph position="5">  In the other example sentences, the agreement reverts to the default male gender, since the direct object has been extracted.</Paragraph>
      <Paragraph position="6">  The frog had been eaten by the fish that Zayd ate. row the term avachchedak (limiter) from Navya Nyaya logic to express the relationship between the relative clause and the noun being modified, which is the avachchinna - &amp;quot;the limited&amp;quot;. From a PGF perspective, the data is explained as follows.</Paragraph>
    </Section>
    <Section position="5" start_page="5" end_page="8" type="sub_section">
      <SectionTitle>
3.3 Discussion
</SectionTitle>
      <Paragraph position="0"> In Hindi, ergative marking of the subject is required by the yaa form of the verb, as shown earlier, except in the case of pronominal vo which is not explicitly marked with ne. The subject for both clauses is thus clearly identified (either it is marked with ne or it is the pronominal vo). If word order were completely free, both the main verb and the complement would be candidate heads for the ergatively marked verbs. For 2(a) and 2(b), the co-relative vo introduces a projectivity constraint which removes the ambiguity.</Paragraph>
      <Paragraph position="1"> In 2(a), the pronominal co-relative vo takes the place of the object of the relative clause machli.</Paragraph>
      <Paragraph position="2"> In 2(b), the object of the relative clause is absent (the typical constituency analysis would say that it leaves a trace), but is marked by the demonstrative co-relative vo in the main clause, establishing the connection to the relativiser jo.</Paragraph>
      <Paragraph position="3"> Both of these represent the left-adjoined form of the relative clause.</Paragraph>
      <Paragraph position="4"> The relationship between the relativiser jo and the co-relative vo is implicit, since the direct relationship between jo and the modified noun merely marks the noun, in the case of 2(a), as the avachchinna - the item limited by the relative clause. In practical terms, the presence of an argument marked as avachchinna is propagated up to the head of the relative clause, marking it as the avachchedak limiter, meaning that the verb cannot take anything other than the co-relative vo as its head. This analysis is in keeping with the requirement that jo must have amatchingvo in the sentence.</Paragraph>
      <Paragraph position="5"> The same analysis applies when the modified noun is not explicitly present in the relative clause. In this case, jo still signifies the presence of an avachchinna argument, but it is pronominal, and the property of being avachchinna is conferred upon the modified noun by vo.In this way, there is a pleasing symmetry between vo and jo in that when they depend on a noun, they act as demonstratives, marking the noun as avachchinna, and when appearing independently, they act as pronouns in place of the avachchinna noun.</Paragraph>
      <Paragraph position="6"> In 2(c), we have the right adjoined form of the relative clause. Additionally, the typical SOV word order of the main clause has been altered to OSV through topicalising the object, maindek. The co-relative vo continues to assert a projectivity constraint on the relative clause.</Paragraph>
      <Paragraph position="7"> A further constraint on right-adjoined relative clauses is that they may not contain the explicit subject. This constraint is enforced via the karaka relation between the avachchedakmarked verb and the pronominal co-relative vo  , thus ensuring that the pronominal vo never precedes its referent.</Paragraph>
      <Paragraph position="8"> noun, while the pronominal vo does not, and hence we treat these separate lexical entries.</Paragraph>
      <Paragraph position="9">  This is an example of a generalisation of the karaka chart to include word order constraints.</Paragraph>
      <Paragraph position="10">  To account for the Arabic data, we first must establish the relevant vibhakti to karaka mapping. The syntactic role of inflection in Arabic is well defined (cf. (Fischer, 2002)), and the relevant rules (for the examples under consideration) can be summarised as follows: * Nominative case: (-un or -u)marksthe subject * Verbs agree in number and gender with the subject * Where a verbal argument is absent, a pronominal affix (e.g. -h-a) is attached to the verb, which will agree in number and gender with the missing item.</Paragraph>
      <Paragraph position="11"> * Relativisers (e.g. allaty) agree in gender with the noun being modified.</Paragraph>
      <Paragraph position="12"> Given that the karta relation does not necessarily require an explicit subject, in absence of explicit marking of the subject, the relationship can still be derived from gender agreement  .</Paragraph>
      <Paragraph position="13"> Furthermore, the subject will always precede the object. In the sense that vibhakti can be generalised to include word order (Bharati et al., 1996a), we may also include this condition as part of the karaka chart (see Table 2). As in Hindi, the relativisers impose projectivity on their dependents.</Paragraph>
      <Paragraph position="14">  This set of rules is sufficient to account for the grammaticality of the data given here, but does not account for the semantic relationship between the three sentences. In this respect, the karaka relations have so far only illustrated their connection to the purely syntactic level of analysis. While we do not wish to regress to some kind of transformational account of the surface forms, it is desirable to illustrate that the karaka relations provide sufficient abstraction to permit the formulation of a common semantic representation.</Paragraph>
      <Paragraph position="15"> One way to approach this is to argue that relativisers 'mediate' the appropriate karaka relation from the complement to the noun being modified. This is essentially a process of unifying the embedded resumptive pronoun in  This is also true in Hindi.</Paragraph>
      <Paragraph position="16"> the complement with the relevant external argument. Tomakethismoreexplicit,wehaveseparated the resumptive pronoun from the complement in the PGF analysis and shown the karaka relation that exists between them.</Paragraph>
      <Paragraph position="17"> Thus in 3(a), allaty mediates a karma relation between al-samakah and arrowhookrightakala-h-a. In 3(c), this connection is repeated in the relative clause, and arrowhookrightalladhy mediates a karma relationship between al-d .</Paragraph>
      <Paragraph position="18"> ifdaarrowhookright and arrowhookrightakalat-hu. In 3(b), there is no explicit resumptive pronoun, since this feature is only used for a missing object. Instead, gender and number agreement between al-samakah and akalat means that the karta relationship is obtained via allaty. This mediation of karaka relations is illustrated by a typical feature structure unification diagram shown in (6).</Paragraph>
      <Paragraph position="19"> Even though the feature structures differ in terms of argument structure, they can be given an equivalent semantic interpretation in terms of lambda abstraction:  which can be factored out, since it is the identity function of frog, thus leaving 6(c) with the same interpretation as 6(a).</Paragraph>
      <Paragraph position="20"> The same mediation process can be applied to the Hindi data, using the same argument. Here, the relative and (pronominal) co-relative guide the unification. The karaka relation between the (pronominal) relativiser jo and the avachchedak verb is mapped onto the avachchinna noun marked by vo or directly onto the pronominal co-relative vo, as the case may be. In this sense, the pronominal relative or co-relative operates in the same was the resumptive pronouns in Arabic.</Paragraph>
      <Paragraph position="21"> After unification, all three Hindi sentences share the same feature structure, shown in 7, and hence further analysis to demonstrate their semantic equivalence is not needed.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>