<?xml version="1.0" standalone="yes"?>
<Paper uid="C02-1116">
  <Title>Linking syntactic and semantic arguments in a dependency-based formalism</Title>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2 Linguistic Data
</SectionTitle>
    <Paragraph position="0"> Insights from corpus studies (e.g. the NEGRA treebank for German (Skut et al., 1998), or the material annotated in the Framenet (Baker et al., 1998) project on the basis of The Bank of English (Cobuild, 2001) show that the syntactic patterns specific verbs occur with vary stongly. Not only do we observe different patterns for different verbs, but also alternative patterns with the same verbs. (1) to (6) illustrate the well-known dative shift (Levin, 1993):45 phenomenon, which occurs with a restricted class of verbs only. While the distinction between (2) and (4) can be explained in terms of lexical semantics, even semantically closely related verbs as English give and deliver can differ in their syntactic behaviour, as the contrast between (1) and  (5) shows.</Paragraph>
    <Paragraph position="1"> (1) [The postman] gave [him] [a package].</Paragraph>
    <Paragraph position="2"> (2) [The postman] gave [a package] [to him].</Paragraph>
    <Paragraph position="3"> (3) [The postman] charged [him] [5 Euros].</Paragraph>
    <Paragraph position="4"> (4) *[The postman] charged [5 Euros] [to him].</Paragraph>
    <Paragraph position="5"> (5) *[The postman] delivered [him] [a package].</Paragraph>
    <Paragraph position="6"> (6) [The postman] delivered [a package] [to him]. In contrast to (Davis, 1998):3:562 for instance, we do not assume a difference in meaning between (1) and (2).3 Therefore, in order to compute a se- null mantics from this data without spurious ambiguity, we must be able to express the semantic generalisation that to him and him realize the same semantic argument. It is useful to employ thematic roles in the lexicon and in grammatical descriptions for this purpose.4 See e.g. (Tarvainen, 1987), (Helbig, 1989), (Sgall et al., 1986) or the Framenet project (Baker et al., 1998) for different application-oriented sets of thematic roles (or &amp;quot;frame elements&amp;quot;). For discussion see e.g. (Helbig, 1995) or (Kruijff, 2001), for criticism see (Dowty, 1991). We can also use thematic roles to structure verbs into an ontology, as e.g. attempted by (Helbig, 1995), (Davis, 1998) or (Baker et al., 1998) in order to make semantic predictions for syntactic valency patterns. For instance, it is a regularity in English that verbs of charging do not show the dative shift (Levin, 1993), while verbs of change of possession sometimes do.</Paragraph>
    <Paragraph position="7"> Now consider the set of German examples in (7) to (11), which all roughly express the proposition Peter robs her of her money. All of the patterns are attested in the NEGRA corpus (Skut et al., 1998), but (10) cannot be found.</Paragraph>
    <Paragraph position="8">  The data illustrates that it can be a lexical prop-erty of verbs to allow or prohibit omission of their  cause someone trouble, pain, etc. with different lexical entries, following established lexicographic practice. While the lexical entry for meaning (b) will exhibit the syntactic pattern illustrated by (1) only (*To give headache to someone), the entry for meaning (a) exhibits both the patterns in (1) and (2). 4Note that we do not commit ourselves to a specific set of thematic roles in this paper.</Paragraph>
    <Paragraph position="9"> complements (Levin, 1993):33, (Helbig, 1995):99.</Paragraph>
    <Paragraph position="10"> Therefore, we will analyse syntactic arguments in terms of optionality and obligatoriness. Note that this distinction is not predictable from the thematic roles realized by the syntactic elements (e.g. by distinguishing inner and outer roles in (Sgall et al., 1986) and (Tarvainen, 1987)) nor by the syntactic form or even function of the syntactic elements. Neither is the distinction between obligatory and optional elements the same as the complement/adjunct distinction.</Paragraph>
    <Paragraph position="11"> We analyse (1) to (6) as alternative realizations of a thematic role, because one semantic argument (the PATIENT) can either be realized as indirect object or PP, while the THEME is always realized as a direct object NP. Compare this data to alternations as in (12) and (13). Here, additionally, one syntactic function (direct object) is open for either of two thematic roles (Levin, 1993).</Paragraph>
    <Paragraph position="12">  (12) [He] cleared [the dirt] [from the table].</Paragraph>
    <Paragraph position="13"> (13) [He] cleared [the table] [of the dirt].</Paragraph>
    <Paragraph position="14">  We will show in Section 4 how we can account for the data illustrated in this section in a lexicalized dependency grammar formalism and show that the linguistic view taken above helps to reduce redundancy in the lexicon.</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3 Alternative approaches
</SectionTitle>
    <Paragraph position="0"> The approach taken in this paper formalizes notions that have only been informally employed in dependency grammar. (Helbig, 1995):167 defines valency frames on a formal syntactic and functional syntactic level, a thematic role level and a logical level in his 6-level-model, but only informally and for the purpose of learners' dictionaries. There is a long tradition in German lexicography which has produced a number of valency dictionaries (e.g. (Helbig and Schenkel, 1975), (Fischer, 1997), (Engel and Schumacher, 1976)). The syntactic analyses in these dictionaries are compatible with our model, but they do not provide a thematic role level.</Paragraph>
    <Paragraph position="1"> (Melcuk, 1988) characterizes valency frames in a similar fashion (94), but uses informal additions in natural language to constrain the possible patterns. Also (Melcuk, 1988) assumes different levels of representation. A shortcoming of the syntactic level in (Melcuk, 1988) is, though, that his syntactic classes are dependent on the specific lexical item, and therefore problematic to define. The approach we will take resembles LFG (Kaplan and Bresnan, 1982) (Bresnan, 2001) in that it assumes syntactic relations.</Paragraph>
    <Paragraph position="2"> (Davis, 1998) has recently proposed a linking theory in the formal framework of HPSG. He separates syntax and semantics by postulating thematic roles under the CONTENT feature of his HPSG architecture (Pollard and Sag, 1994), and syntactic characterizations of the arguments under CAT-EGORYjARG-ST and CATEGORYjSUBCAT. He has separate hierarchies of syntactic patterns (intrans, trans, ditrans, 5:32) and semantic classes (subtypes of RELATION, 5:72). These hierarchies interact by a set of linking constraints and yield a hierarchy of predicators (5:41), which specifies possible linkings of thematic roles to syntactic arguments. While (Helbig, 1995) obviously employs a large role set, (Davis, 1998) has only 6 roles, and moves thematic roles further down into semantics than we assume by postulating them on an event level, which &amp;quot;effectively amounts to a limited amount of semantic decomposition&amp;quot; (5:39). The shortcoming of the model is that the syntactic patterns assumed are very sparse indeed with only three transitivity classes.</Paragraph>
    <Paragraph position="3"> Due to this, semantic predictions can be made only for NP-complements, while PPs must be treated by a separate mechanism (&amp;quot;content sharing account&amp;quot;). Thus, there is no specific prediction for the prepositional complement in English dative shift constructions. The advantage of Davis's model, in contrast, is the lexical inheritance architecture which is a formal means to capture generalizations.</Paragraph>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
4 Formalization
</SectionTitle>
    <Paragraph position="0"> We formalize our idea of linking and valency frames as an extension of the new lexicalized, dependency-based grammar formalism of Topological Dependency Grammar (TDG) (Duchier and Debusmann, 2001), (Debusmann, 2001). So far, TDG is only concerned with syntax: every TDG analysis consists of an unordered dependency tree (ID tree) and an ordered and projective topology tree (LP tree).</Paragraph>
    <Paragraph position="1"> We only describe a subset of the full TDG grammar formalism (e.g. completely ignoring any issues concerning word order) and extend it with the notion of a thematic graph (TH graph). We call the version of TDG described in this paper TDGTH.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
4.1 Thematic graphs
</SectionTitle>
      <Paragraph position="0"> notion by an example. (14) is an ID tree analysis for the sentence Peter will live in Taipei: We show the corresponding TH graph in (15). Here, Peter is the patient of will live and in Taipei the locative. Note that we collapsed the auxiliary will and its verbal complement live into a single node, and also the PP in Taipei: Peter will live in Taipei th loc (15)</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
4.2 The lexicon
</SectionTitle>
      <Paragraph position="0"> This section deals with the TDGTH-lexicon. We assume a finite set of syntactic roles R and a finite set of thematic roles T . We write r for a syntactic role in R and th for a thematic role in T . P =f!;?g is the set of optionality flags pi. A = fvalID;valTH;linkg is the set of lexical features a, and E the set of lexical entries e, having the following signature:5</Paragraph>
      <Paragraph position="2"> E is a lattice of TDGTH-lexicon entries; lexical entries either correspond to words or to lexical types which can be inherited (see below).</Paragraph>
      <Paragraph position="3"> The value of feature valID is a set of pairs (r;pi) of syntactic roles and an optionality flag modeling the concept of syntactic valency. The value of valTH a set of pairs (th;pi) of thematic roles and an optionality flag (thematic valency). For convenience, we write rpi for (r;pi), and thpi for (th;pi). The value of link is a set of pairs (th;r) of thematic and syntactic roles, expressing the mapping between them. We call a pair in this set a linking.</Paragraph>
      <Paragraph position="5"> As an example, (16) is a lexical entry for finite eat: eat has an obligatory subject (subj) and an optional direct object (objd) in its syntactic valency. Its thematic valency contains an obligatory AGENT and an optional THEME. The link-feature defines two linkings: one links the AGENT to the subject and the THEME to the direct object.</Paragraph>
    </Section>
    <Section position="3" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
4.3 Lexical inheritance
</SectionTitle>
      <Paragraph position="0"> We introduce a mechanism of lexical inheritance.</Paragraph>
      <Paragraph position="1"> We write e = e1 u:::uen for 'lexical entry e inherits from lexical entries e1;:::;en', and define inheritance as the set union of the individual features' values:</Paragraph>
      <Paragraph position="3"> We can now use lexical inheritance to model our notion of valency frames. We introduce the notion of a linking type as a lexical entry that does not specify any other lexical attributes besides valID, valTH and link. Such linking types specify a partial valency frame from which we can build complete valency frames by inheritance. For instance, consider the following two linking types:</Paragraph>
      <Paragraph position="5"> The linking type l ag subj maps the agent to the subject, and l th objd the theme to the direct object.</Paragraph>
      <Paragraph position="6"> Out of the two, we can construct our lexical entry for eat by lexical inheritance: eat = l ag subj u l th objd (19) which amounts precisely to the lexical entry displayed in (16) above. We call the values of the three features valID, valTH and link in a lexical entry obtained by inheriting from linking types a valency frame.</Paragraph>
    </Section>
    <Section position="4" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
4.4 Role hierarchies
</SectionTitle>
      <Paragraph position="0"> We arrange the set R of syntactic roles in a role hierarchy modeled as a meet semi-lattice. Here is an example cut-out of the role hierarchy: dativeshift ppdirectional ppspatial obji ppto ppinto ppin ppunder (20) We write r v r0 for r is a specialization of r0 (i.e. r is below r0 in the hierarchy).</Paragraph>
      <Paragraph position="1"> We employ the role hierarchy to model alternative realizations in the sense of section 2: e.g. using the hierarchy above, dativeshift can be realized as either obji or ppto but not by either ppdirectional, ppinto or ppin. Note that certain roles (ppto, ppinto, etc.) will be realized by only two lexical entries, viz. the prepositions to and into respectively, while other roles like subj, obji or objd can be realized by a large set of lexical entries.</Paragraph>
      <Paragraph position="2"> In the same fashion, we arrange the set T of thematic roles in a role hierarchy, but in this article we keep this hierarchy completely flat.</Paragraph>
      <Paragraph position="3"> Lexical entry constraint. To forbid that different thematic roles are mapped to the same syntactic role realization, we add a condition for well-formed lexical entries: for every lexical entry e, the value of its link-feature, link(e) =f(th1;r1);:::(thn;rn)g must not include two syntactic roles ri, r j (1 i6= j n) such that ri v r j.</Paragraph>
    </Section>
    <Section position="5" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
4.5 TDGTH analyses
</SectionTitle>
      <Paragraph position="0"> We can now define a TDGTH-analysis as a tuple (V;EID;e; ;ETH). It consists of the finite set V of nodes w, the finite set EID of ID edges EID V V R , where R is the set of syntactic roles, and the lexical assignment function e : V ! E assigning lexical entries to nodes. We write w1 r!IDw2 for an ID edge from node w1 to node w2 labeled with syntactic role r.</Paragraph>
      <Paragraph position="1"> Collapsing nodes. As in the example ID tree and TH graph-analyses in (14) and (15) above, we would like to be able to collapse sets of nodes in the ID tree into single nodes in the TH graph. We introduce a collapsing principle into the grammar formalism, according to which the node of an auxiliary verb, a preposition or a determiner will collapse with its daughter.</Paragraph>
      <Paragraph position="2"> To capture this idea, we posit an equivalence relation , and write V= for the set of equivalence classes over V . An equivalence class directly corresponds to a &amp;quot;collapsed node&amp;quot; in the TH graph. ETH (V= ) (V= ) T is the finite set of of TH edges, and we write w for the equivalence class containing w. w1 th!THw2 is a TH edge from node w1 to node w2 labeled with th.</Paragraph>
      <Paragraph position="3"> When we collapse nodes, we also collapse their lexical entries: the value of the lexical feature a of a collapsed node w1 = fw1;:::;wng is the set union of the values assigned to the individual nodes:</Paragraph>
      <Paragraph position="5"> In the example TH graph in (15) above, the two nodes will and live are collapsed into the equivalence class will live. We assume that will is mapped to the following lexical entry:  Here, we use the independent definition of the valID, valTH and link features in order to express that function words like auxiliaries or prepositions realize syntactic arguments which are semantically dependent on different lexical items. This also allows for an elegant treatment of semantically void syntactic arguments as fake-reflexives and non-referential it.6 Infinitive live has this lexical entry:  according to its well-formedness conditions. In particular, the value of the valID feature is not relevant to the TH graph. constraint restricts the number of outgoing edges of each node in the ID tree: if (r;!) 2 valID(w), then w must have precisely one outgoing edge labeled with r0 v r, at most one if (r;?) 2 valID(w), and none otherwise. Thus, (r;!) stands for an obligatory rcomplement of w and (r;?) for an optional one. The thematic valency constraint is defined analogously.</Paragraph>
      <Paragraph position="6"> Linking constraint. An edge in the TH graph is only licensed if it satisfies the lexicalized linking constraint. It states that an edge w1 th!THw2 in the TH graph is licensed only if there is a linking (th0;r0) 2 link(w1) and an edge w01 r!IDw02 in the ID tree such that w10 = w1, w20 = w2, th v th0 and r v r0.</Paragraph>
      <Paragraph position="7"> Consider the example ID tree and TH graph analyses in (14) and (15) above. The edge will live loc!THin Taipei is mandated by the thematic valency of will live, but it must also be licensed by the linking principle: indeed there is a linking (loc;ppspatial) in link(will live) and an ID edge live ppin!IDin such that live = will live and in = in Taipei, and loc v loc and ppin v ppspatial.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>