File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/06/w06-1509_metho.xml
Size: 22,711 bytes
Last Modified: 2025-10-06 14:10:39
<?xml version="1.0" standalone="yes"?> <Paper uid="W06-1509"> <Title>Binding of Anaphors in LTAG</Title> <Section position="4" start_page="0" end_page="65" type="metho"> <SectionTitle> 2 Basic Anaphor Binding </SectionTitle> <Paragraph position="0"> In traditional accounts, binding is defined representationally: an antecedent binds an anaphor iff they are are coindexed and in a certain structural relationship. In an LTAG, binding cannot be viewed in this way as the notion of coindexation is foreign to the formalism. An LTAG analysis can therefore not be a mere translation of a previous account.</Paragraph> <Paragraph position="1"> Although the phenomenon is very complex, the basic properties of binding are quite well understood. Binding of an anaphor by an antecedent consists of coreference and agreement between the two items. Furthermore, it is well known that binding of English anaphors is an asymmetrical, local, structural relationship. The asymmetry of binding can be easily observed in examples (1) versus (3). Locality is reflected by the fact that (1) is grammatical, but not (4).</Paragraph> <Paragraph position="2"> (4) * Johni knows that Mary likes himselfi.</Paragraph> <Paragraph position="3"> Finally, the binding relationship is known to be structural because the positions of binder and anaphor play a crucial role. This is discussed in more detail below.</Paragraph> <Section position="1" start_page="65" end_page="65" type="sub_section"> <SectionTitle> 2.1 Lexical Entry </SectionTitle> <Paragraph position="0"> The domain of locality that LTAG provides enables us to encode a local structural relationship, such as the one between the anaphor and its antecedent, very directly. We understand binding as a lexical requirement of the anaphor: that it must be bound. Thus, we propose the lexical entry in Figure 1 for reflexives. It is a multicomponent set whose second component is the anaphor. The first component is a degenerate auxiliary tree which adjoins into the elementary tree of the antecedent.</Paragraph> <Paragraph position="1"> In LTAG, elementary trees encode both syntax and semantics. Thus, the two components of binding, coreference and agreement, are simultaneously guaranteed by the coindexations between the feature structures of binder and anaphor. Furthermore, since the derivation must be tree-local, locality is also ensured. A c-command constraint between the two components accounts for the asymmetry between the antecedent and anaphor as shown in examples (1) and (3). This constraint is checked when the two components are composed into an elementary tree (by tree-locality).</Paragraph> </Section> <Section position="2" start_page="65" end_page="65" type="sub_section"> <SectionTitle> 2.2 Example Derivation </SectionTitle> <Paragraph position="0"> Consider (5), where himself has two possible antecedents, John and Bill. Our analysis derives both readings, given a standard tree inventory as in Figure 2.</Paragraph> <Paragraph position="1"> (5) Johni showed Billj himselfi/j.</Paragraph> <Paragraph position="2"> Sentence (5) is syntactically ambiguous under this analysis, since two different derivations lead to distinct readings. This seems to reflect our intuitions about this sentence well, although it contrasts with the traditional vew of BT, where the coindexation between binder and anaphor is part of the syntactic structure for the sentence, and thus no ambiguity arises.</Paragraph> </Section> <Section position="3" start_page="65" end_page="65" type="sub_section"> <SectionTitle> 2.3 Flexible Composition </SectionTitle> <Paragraph position="0"> Tree-locality requires the first component of himself to adjoin into a higher NP substitution node.</Paragraph> <Paragraph position="1"> However, adjunction into substitution nodes is generally disallowed. Adjunction of the first component of himself into the root node of the John-tree tj or the Bill-tree tb is, however, not tree-local. Therefore, we employ flexible composition (Joshi et al., 2003) to compose tj with the first component of th (t1h), yielding a derived multicomponent set. Composition of th with ts is then treelocal. This yields the reading where John is the antecedent of himself.</Paragraph> <Paragraph position="2"> Alternatively, tb composes with th first, which derives the other reading. The two derivation trees representing these readings are shown in Figure 3.</Paragraph> </Section> <Section position="4" start_page="65" end_page="65" type="sub_section"> <SectionTitle> 2.4 Advantages </SectionTitle> <Paragraph position="0"> The different binding options (e.g., in double-object sentences) follow directly from the derivation and do not have to be hardcoded. Furthermore, the reflexive itself is responsible for agreement and coreference with its antecedent.</Paragraph> </Section> <Section position="5" start_page="65" end_page="65" type="sub_section"> <SectionTitle> 2.5 Alternative Analysis </SectionTitle> <Paragraph position="0"> There is at least one obvious alternative analysis for BT in LTAG. In this case, features are employed instead of a multicomponent set to derive the binding relationship. Features on each verbal elementary tree would encode whether an argument is an anaphor, and if so, what it is bound to. Just like in our analysis introduced above, a certain locality necessary for binding can be ensured under this approach. However, this approach is very stipulative. It is merely an accident that agreement and coreference go hand in hand: Two separate feature equations have to ensure agreement between the binder and anaphor, and coreference between them. Furthermore, a number of verbal trees is added; and the reflexive itself becomes syntactically and semanticially vacuous.</Paragraph> </Section> </Section> <Section position="5" start_page="65" end_page="65" type="metho"> <SectionTitle> 3 Reciprocals </SectionTitle> <Paragraph position="0"> Another advantage of the proposed account is that it allows an analogous treatment of reciprocals like each other in (6).</Paragraph> <Paragraph position="1"> (6) [John and Mary]i like each otheri. This is desirable given that the syntactic behavior of reciprocals resembles reflexives. Semantically, though, reciprocals are very complex (Dimitriadis, 2000). The meaning of &quot;each other&quot; roughly corresponds to its parts, &quot;each&quot;, and &quot;other&quot;. That is, &quot;John and Mary love each other&quot; means something like &quot;Of John and Mary, each loves the other&quot;.1 These properties are neatly accounted for with our analysis of each other that is syntactically analogous to himself, but contributes additional operators in the semantics2. The proposed lexical entry is spelled out in Figure 4.</Paragraph> <Paragraph position="2"> The fact that each other contributes two distinct quantifiers corresponds directly to its syntactic analysis as a two-part multicomponent set.</Paragraph> </Section> <Section position="6" start_page="65" end_page="69" type="metho"> <SectionTitle> 4 Nonlocal Antecedents </SectionTitle> <Paragraph position="0"> The discussion of anaphoric pronoun binding discussed in the previous section demonstrated how certain locality (7) and configurational restrictions (8) on anaphoric pronouns follow from TAG's constrained mechanisms of structural composition coupled with a multicomponent analysis of reflex1It is sometimes claimed that &quot;long-distance&quot; reciprocals require non-local adjunction of &quot;each&quot;: (i) The boxers thought they would defeat each other.</Paragraph> <Paragraph position="2"> The LTAG analysis proposed here does not allow this. This may constitute independent evidence for Dimitriadis' (2000) analysis of reciprocals in which &quot;each&quot; is not as high as it seems in these kinds of examples.</Paragraph> <Paragraph position="3"> discussion. We assume for simplicity that each other corresponds to each+the other, as reflected in the lexical entry. 3subsetsqequalA= &quot;is an atomic part of&quot;.</Paragraph> <Paragraph position="4"> In the absence of a complete analysis of plural semantics in LTAG, we assume here that plural noun phrases like &quot;John and Mary&quot; or &quot;the boys&quot; contribute at least a group (G) variable. This variable is used by certain collective predicates, for example in &quot;The boys surrounded the castle.&quot; It corresponds to the plural individual contributed by the NP.</Paragraph> <Paragraph position="5"> The semantics given here predicts strongly distributive &quot;each other&quot;. Some adjustment is needed to account for lesser forms of distributivity.</Paragraph> <Paragraph position="6"> ives and reciprocals.</Paragraph> <Paragraph position="7"> (7) a. Johni likes himselfi.</Paragraph> <Paragraph position="8"> b. *Johni thinks that Mary believes that Kate likes himselfi.</Paragraph> <Paragraph position="9"> (8) a. John believes Maryi to like herselfi.</Paragraph> <Paragraph position="10"> b. *John believes herselfi to like Maryi.</Paragraph> <Paragraph position="11"> A significant problem with this analysis as stands, however, is that it works too well, denying the grammaticality of certain raising (9) and ECM constructions (10) and constructions in which the anaphor is embedded within a subject (11). Under current assumptions, the antecedent-anaphor dependency must be established within an elementary tree (by adjunction of a single multi-component set). However, for both of these constructions the anaphor and its antecedent lie in different elementary trees. In (9) the auxiliary tree of the raising verb seems contains no local argument for the degenerate NP* component to combine with. In (10) himself occurs as an argument of like while its antecedent occurs in another elementary tree, believe. In each case, generating the examples requires that we relax some of our current assumptions.</Paragraph> <Paragraph position="12"> (9) Johni seems to himselfi to be a decent guy.</Paragraph> <Paragraph position="13"> (10) Johni believes himselfi to be a decent guy.</Paragraph> <Paragraph position="14"> (11) Johni thought that the pictures of himselfi were wonderful.</Paragraph> <Section position="1" start_page="67" end_page="67" type="sub_section"> <SectionTitle> 4.1 Raising </SectionTitle> <Paragraph position="0"> We see from (9) that anaphors can occur as experiencer arguments of raising verbs providing they are c-commanded by a possible antecedent.</Paragraph> <Paragraph position="1"> Though predicted to be ungrammatical under the current proposal, (9) can be generated if we relax the requirement that the two parts of the multicomponent set of the anaphor attach to the same elementary tree. This relaxation could take the form of simply allowing non-local adjunction for specific classes of multicomponent sets, those with a degenerate components. Alternately, we retain the restriction to tree-local MC-TAG but achieve nonlocality through more extensive use of flexible composition, already adopted for independent reasons. null Under a flexible composition analysis (Figure 6), the John-tree composes with the degenerate NP* member of the reflexive set as before. This yields a derived multicomponent set consisting of one derived part, John, and one underived part, himself. The seems-tree then composes with the himself component of the reflexive set, yielding a derived set (Figure 5) containing the components John and seems to himself. Finally, this derived multicomponent set combines with the like-tree, the John component substituting into the open NP slot and the seems to himself component adjoining at VP.</Paragraph> </Section> <Section position="2" start_page="67" end_page="68" type="sub_section"> <SectionTitle> 4.2 ECM </SectionTitle> <Paragraph position="0"> In ECM constructions such as (10) the anaphor appears as the subject of the embedded to be a decent guy-tree while its antecedent appears as subject of the matrix believes-tree. A derivation for this sentence under our account is shown in Figure 7. As before, the John-tree first composes with the degenerate NP* component of the reflexive tree, followed by the the substitution of the himself-tree into the to be a decent guy-tree, yielding the derived multicomponent set containing John and believes himself, which locally composes with the to be a decent guy-tree.</Paragraph> </Section> <Section position="3" start_page="68" end_page="69" type="sub_section"> <SectionTitle> 4.3 Subject Embedding </SectionTitle> <Paragraph position="0"> Anaphors contained within embedded subjects4 (12) cause the binding domain to be minimally expanded. Again, it is transparent that these cases can be derived successfully from the lexical entry in Figure 1 and repeated use of flexible composition. null (12) a. The meni knew that pictures of each otheri were on sale.</Paragraph> <Paragraph position="1"> b. The meni felt that the pictures of themselvesi were horrid.</Paragraph> <Paragraph position="2"> c. The meni knew that each otheri's pictures were on sale.</Paragraph> <Paragraph position="3"> be a decent guy.&quot;</Paragraph> </Section> <Section position="4" start_page="69" end_page="69" type="sub_section"> <SectionTitle> 4.4 Constraints on Flexible Composition </SectionTitle> <Paragraph position="0"> The use of flexible composition with tree-local MC-TAG is very powerful, thus able to account for the non-local binding in (9), (10), and (12).</Paragraph> <Paragraph position="1"> However, it is too powerful if unconstrained as it will also generate (13). It is necessary to constrain the derivation such that in the derived tree no sub-ject node intervenes between the antecedent and anaphor (Chomsky's Subject Intervention Constraint). This is obtained by strengthening the link between NP and himself in the lexical entry s.t. when the two trees connected by the link are adjoined, a requirement that NP* c-command himself and no subject NP intervenes between the two (c-commanding himself and c-commanded by NP* ) is checked. This constraint formalizes the descriptive account given in the linguistic literature. Note that a similar account may be active in other places in the grammar as well, due to the pervasiveness of left-edge phenomena (see section 5.4).</Paragraph> <Paragraph position="2"> Computationally, this constraint can be checked as soon as the multicomponent set which contains it attaches into another elementary tree. C-command as well as subject intervention cannot be disturbed by later operations on the outer tree, if they are valid at the time of composition.</Paragraph> <Paragraph position="3"> (13) * Johni believes me to like himselfi.</Paragraph> </Section> </Section> <Section position="7" start_page="69" end_page="71" type="metho"> <SectionTitle> 5 Further Issues </SectionTitle> <Paragraph position="0"/> <Section position="1" start_page="69" end_page="69" type="sub_section"> <SectionTitle> 5.1 Exempt Pronouns </SectionTitle> <Paragraph position="0"> As it currently stands, the proposal follows heavily in the footsteps of traditional configurational approaches to BT. As such, it mirrors the more traditional BT of Chomsky in it's inability to license such examples as (17b), where the antecedent does not c-command the anaphor and (14) and (15), where binding is possible despite presence of an intervening subject along the c-command path.</Paragraph> <Paragraph position="1"> (14) a. I spoke to [John and Bill]i about each otheri.</Paragraph> <Paragraph position="2"> b. Pictures of myselfi frighten mei.</Paragraph> <Paragraph position="3"> c. Johni's greatest problem is a fear of himselfi.</Paragraph> <Paragraph position="4"> (15) [John and Mary]i are letting the honey drip on each otheri's feet.</Paragraph> <Paragraph position="5"> (16) Clones of each other annoy the children. The examples in (14) can be accommodated by having the prepositions appearing before the arguments be surface case markers rather than real prepositions (as suggested in (Jackendoff, 1972)). Even so, (15) and (16) remain and seem to present an intractable problem for an LTAG account, as well as traditional accounts of English binding phenomena. This may in fact be the case and prove firm support for claims by numerous authors (Pollard and Sag, 1994; Postal, 1971; Kuroda, 1965) that at least part of the data subsumed under BT (the &quot;exempt pronouns&quot;) is governed by pragmatic constraints such as point-of-view rather than purely structural constraints. In fact, the LTAG analysis proposed here is a relatively clean structural account of English binding data. The (un)availability of a derivation for certain examples may thus point to their classification into &quot;exempt&quot; and regular anaphora. These considerations are left for further work.</Paragraph> </Section> <Section position="2" start_page="69" end_page="70" type="sub_section"> <SectionTitle> 5.2 Extraction </SectionTitle> <Paragraph position="0"> A potential problem for the proposed analysis is presented by extraction phenomena, as in wh-movement or topicalization. Extraction of a phrase containing an anaphor, whether topicalization or (17) or wh-movement (18), does not induce a Condition A violation. The current proposal predicts the grammaticality of (17a) and (18a) given that in each case the reflexive is locally c-commanded by its antecedent. However, in (17b) and (18b) the reflexive fails to be c-commanded by its antecedent, hence these examples are predicted to be ungrammatical although they are clearly acceptable. null (17) a. Johni saw himselfi.</Paragraph> <Paragraph position="1"> b. Himselfi John saw ti.</Paragraph> <Paragraph position="2"> (18) a. Johni liked the color pictures of himselfi.</Paragraph> <Paragraph position="3"> b. [Which pictures of himselfi] did Johni like e? A classical solution to these facts involves reconstruction of the Aprime-moved element to its original site for the purposes of binding. Clearly, syntactic reconstruction is untenable in LTAG. However, it is possible to emulate it through an entry akin to that in Figure 8, which is capable of deriving the topicalization examples in (17). The first component is the extracted reflexive (Aprime-moved constituents are marked by extractionfeatures (XTAG Group, 2001)), the second component is the binder, and the third component is the position that the reflexive has been extracted from. The requirement that the antecedent locally c-command the trace of movement has the effect of emulating reconstruction.</Paragraph> <Paragraph position="4"> Note, furthermore, that even if some manner of reconstruction operation were to be implemented in LTAG, we are faced with the reality of cases such as (19), which demonstrate that extraction of an element can alter its set of possible binders. GB accounts (van Riemsdijk and Williams, 1986; Clark, 1982) have explained the opposition in (19) by allowing partial reconstruction to an intermediate trace from which the matrix subject is an accessible binder of the anaphor. The LTAG analysis of wh-movement, though, neither exploits intermediate traces nor allows transformational movement over domains larger than an elementary tree, meaning that such intermediate traces are simply unavailable to us.</Paragraph> <Paragraph position="5"> (19) a. *Marshai thought that I painted a picture of herselfi.</Paragraph> <Paragraph position="6"> b. [Which pictures of herselfi] did Marshai think that I painted e? Instead, we suggest that Spec,IP subjects of clauses are able to bind into Spec,CP of the same clause as proposed by Reinhart (1991) and Frank and Kroch (1995). Rather than being a disadvantage, though, this seems to be a strength, predicting as it does that (20) is bad where reconstruction to a posited intermediate trace would predict acceptability. null (20) *[Which pictures of himselfi] did Mary think that Johni believed that Sally wanted? Future work should attempt to determine the correct form of this lexical entry as well as whether or not it is possible to collapse it with the previously proposed Figure 8.</Paragraph> </Section> <Section position="3" start_page="70" end_page="70" type="sub_section"> <SectionTitle> 5.3 Conditions B,C </SectionTitle> <Paragraph position="0"> It is often assumed that the analyses for anaphors and regular pronouns should be related, because of a certain complementarity in distribution: While anaphors must be locally bound, pronouns must be locally free. In English, however, this complementarity is not absolute (cf. 21-22). Furthermore, a negative locality constraint seems to be discouraged by the LTAG framework. This suggests that the analysis of pronouns is independent of our account of anaphors. We leave pronouns, as well as r-expressions (Mary, the man) for further work.</Paragraph> <Paragraph position="1"> (21) Johni pulled the blanket over himi / himselfi.</Paragraph> <Paragraph position="2"> (22) a. Theyi saw each otheri's friends.</Paragraph> <Paragraph position="3"> b. Theyi saw theiri friends.</Paragraph> </Section> <Section position="4" start_page="70" end_page="70" type="sub_section"> <SectionTitle> 5.4 Importance of the Left Edge </SectionTitle> <Paragraph position="0"> Examination of language exposes the left edge to be special with regards to certain phenomena. In Binding Theory, this is revealed in the guise of a Subject Intervention Constraint. Case assignment represents a similar case. We see that verbs can assign accusative case to objects, and subjects of the next lowest clause (ECM), but no further. Ideally, a new analysis of left-edge effects would clarify the relationship between the two components of the lexical entry proposed above.</Paragraph> </Section> <Section position="5" start_page="70" end_page="71" type="sub_section"> <SectionTitle> 5.5 Inherent Reflexives </SectionTitle> <Paragraph position="0"> English has a small number of inherently reflexive verbs, such as behave: (23) John behaves himself.5 Note that this verb requires its object to be a reflexive pronoun which is coreferent with its subject: null 5We would like to thank one anonymous reviewer for bringing this example to our attention.</Paragraph> <Paragraph position="1"> (24) * John behaves her.</Paragraph> <Paragraph position="2"> We conclude from this that behave has a specific lexical constraint, namely that its object should be [+ reflexive]. Since there can be no other binder for this reflexive pronoun, it must be bound by the subject of the sentence.</Paragraph> </Section> </Section> class="xml-element"></Paper>