File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/03/j03-4002_metho.xml
Size: 92,249 bytes
Last Modified: 2025-10-06 14:08:06
<?xml version="1.0" standalone="yes"?> <Paper uid="J03-4002"> <Title>c(c) 2003 Association for Computational Linguistics Anaphora and Discourse Structure</Title> <Section position="3" start_page="550" end_page="554" type="metho"> <SectionTitle> 3 A reviewer has asked how much stretching is possible in discourse without losing its thread or having </SectionTitle> <Paragraph position="0"> to rephrase later material in light of the intervening material. One could ask a similar question about the apparently unbounded dependencies of sentence-level syntax, which inattentive speakers are prone to lose track of and &quot;fracture.&quot; Neither question seems answerable on theoretical grounds alone, with both demanding substantial amounts of empirical data from both written and spoken discourse. The point we are trying to make is simply that there is a difference in discourse between any amount of stretching and even the smallest amount of crossing.</Paragraph> <Paragraph position="1"> Webber et al. Anaphora and Discourse Structure whose defeasible inference is triggered by adjacency, we stipulate that discourse structure itself does not admit crossing structural dependencies. (In this sense, discourse structure may be truly simpler than sentence structure. To verify this, one might examine the discourse structure of languages such as Dutch that allow crossing dependencies in sentence-level syntax. Initial cursory examination does not give any evidence of crossing dependencies in Dutch discourse.) If we now consider the corresponding properties of discourse adverbials, we see that they do admit crossing of predicate-argument dependencies, as shown in examples (11)-(13).</Paragraph> <Paragraph position="2"> (11) a. John loves Barolo.</Paragraph> <Paragraph position="3"> b. So he ordered three cases of the '97.</Paragraph> <Paragraph position="4"> c. But he had to cancel the order d. because then he discovered he was broke.</Paragraph> <Paragraph position="5"> (12) a. High heels are fine for going to the theater.</Paragraph> <Paragraph position="6"> b. But wear comfortable shoes c. if instead you plan to go to the zoo.</Paragraph> <Paragraph position="7"> (13) a. Because Fred is ill b. you will have to stay home c. whereas otherwise the two of you could have gone to the zoo. Consider first the discourse adverbial then in clause (11d). For it to get its first argument from (11b) (i.e., the event that the discovery in (d) is &quot;after&quot;), it must cross the structural connection between clauses (c) and (d) associated with because). This crossing dependency is illustrated in Figure 5(i). Now consider the discourse adverbial instead) in clause (12c). For it to get its first argument from (12a) (i.e., going to the zoo is an alternative to going to the theater), it must cross the structural connection between clauses (12b) and (12c) associated with if. This crossing dependency is illustrated in Figure 5(ii). Example (13) is its mirror image: For the discourse adverbial otherwise in (13c) to get its first argument from (13a) (i.e., alternatives to the state/condition of Fred being ill), it must cross the structural connection associated with because. This is illustrated in Figure 5(iii).</Paragraph> <Paragraph position="8"> Crossing dependencies are not unusual in discourse when one considers anaphora (e.g., pronouns and definite NPs), as for example in Discourse structures for examples (11)-(13). Structural dependencies are indicated by solid lines and dependencies associate with discourse adverbials are indicated by broken lines. (explanation prime is the inverse of explanation--i.e., with its arguments in reverse order. Such relations are used to maintain the given linear order of clauses.)</Paragraph> <Paragraph position="10"> This suggests that in examples (11)-(13), the relationship between the discourse adverbial and its (initial) argument from the previous discourse might usefully be taken to be anaphoric as well.</Paragraph> <Section position="1" start_page="552" end_page="554" type="sub_section"> <SectionTitle> 2.2 Discourse Adverbials Do Behave like Anaphors </SectionTitle> <Paragraph position="0"> There is additional evidence to suggest that otherwise, then, and other discourse adverbials are anaphors. First, anaphors in the form of definite and demonstrative NPs can take implicit material as their referents. For example, in (16) Stack five blocks on top of one another. Now close your eyes and try knocking {the tower, this tower} over with your nose.</Paragraph> <Paragraph position="1"> both NPs refer to the structure which is the implicit result of the block stacking. (Further discussion of such examples can be found in Isard [1975]; Dale [1992]; and Webber and Baldwin [1992].) The same is true of discourse adverbials. In (17) Do you want an apple? Otherwise you can have a pear.</Paragraph> <Paragraph position="2"> the situation in which you can have a pear is one in which you don't want an apple-that is, one in which your answer to the question is &quot;no.&quot; But this answer isn't there structurally: It is only inferred. Although it appears natural to resolve an anaphor to an inferred entity, it would be much more difficult to establish such links through purely structural connections: To do so would require complex transformations that introduce invisible elements into discourse syntax with no deeper motivation. For example, in (17), we would need a rule that takes a discourse unit consisting solely of a yes/no question P? and replaces it with a complex segment consisting of P? and the clause it is possible that P, with the two related by something like elaboration. Then and only then could we account for the interpretation of the subsequent otherwise structurally, by a syntactic link to the covert material (i.e., to the possibility that P holds, which otherwise introduces an alterative to).</Paragraph> <Paragraph position="3"> Second, discourse adverbials have a wider range of options with respect to their initial argument than do structural connectives (i.e., coordinating and subordinating conjunctions). The latter are constrained to linking a discourse unit on the right frontier of the evolving discourse (i.e., the clause, sentence and larger discourse units to its immediate left). Discourse adverbials are not so constrained. To see this, consider the following example: 4 We are aware that crossing examples such as (11)--(13) are rare in naturally occurring discourse. We believe that this is because they are only possible when, as here, strong constraints from the discourse adverbial and from context prevent the adverbial from relating to the closest (leftmost) eventuality or an eventuality coerced from that one. But rarity doesn't necessarily mean ill-formedness or marginality, as readers can see for themselves if they use Google to search the Web for strings such as because then, if instead, and whereas otherwise, and consider (1) whether the hundreds, even thousands, of texts in which these strings occur are ill-formed, and (2) what then, instead, and otherwise are relating in these texts. One must look at rare events if one is studying complex linguistic phenomena in detail. Thus it is not the case that only common things in language are real or worth further study.</Paragraph> <Paragraph position="4"> Webber et al. Anaphora and Discourse Structure (18) If the light is red, stop. Otherwise you'll get a ticket.</Paragraph> <Paragraph position="5"> (If you do something other than stop, you'll get a ticket.) This can be paraphrased using the conjunction or: If the light is red, stop, or you'll get a ticket.</Paragraph> <Paragraph position="6"> Here or links its right argument to a unit on the right frontier of the evolving discourse-in this case, the clause stop on its immediate left. Now consider the related example (19) If the light is red, stop. Otherwise go straight on.</Paragraph> <Paragraph position="7"> (If the light is not red, go straight on.) This cannot be paraphrased with or,asin (20) If the light is red, stop, or go straight on.</Paragraph> <Paragraph position="8"> even though both stop and If the light is red, stop are on the right frontier of the evolving discourse structure. This is because otherwise is accessing something else, so that (20) means something quite different from either (18) or (19). What otherwise is accessing, which or cannot, is the interpretation of the condition alone.</Paragraph> <Paragraph position="9"> Thus discourse adverbials, like other anaphors, have access to material that is not available to structural connectives.</Paragraph> <Paragraph position="10"> Finally, discourse adverbials, like other anaphors, may require semantic representations in which their arguments are bound variables ranging over discourse entities. That is, whereas it might be possible to represent Although P, Q using a binary modal operator (21) although(p, q) where formulas p and q translate the sentences P and Q that although combines, we cannot represent P ...Nevertheless, Q this way. We need something more like (22) p [?] nevertheless(e, q) The motivation for the variable e in this representation is that discourse adverbials, like pronouns, can appear intrasententially in an analog of donkey sentences. Donkey sentences such as example (23) are a special kind of bound-variable reading: (23) Every farmer who owns a donkey feeds it rutabagas.</Paragraph> <Paragraph position="11"> In donkey sentences, anaphors are interpreted as covarying with their antecedents: The it that is being fed in (23) varies with the farmer who feeds it. These anaphors, however, appear in a structural and interpretive environment in which a direct syntactic relationship between anaphor and antecedent is normally impossible, so they cannot be a reflex of true binding in the syntax-semantics interface. Rather, donkey sentences show that discourse semantics has to provide variables to translate pronouns, 5 This was independently pointed out by several people when this work was presented at ESSLLI'01 in Helsinki in August 2001. The authors would like to thank Natalia Modjeska, Lauri Karttunen, Mark Steedman, Robin Cooper, and David Traum for bringing it to their attention.</Paragraph> <Paragraph position="12"> Computational Linguistics Volume 29, Number 4 and that discourse mechanisms must interpret these variables as bound--even though the pronouns appear &quot;free&quot; by syntactic criteria.</Paragraph> <Paragraph position="13"> Thus, it is significant that discourse adverbials can appear in their own version of donkey sentences, as in (24) a. Anyone who has developed innovative new software has then had to hire a lawyer to protect his/her interests. (i.e., after developing innovative new software) b. Several people who have developed innovative new software have nevertheless failed to profit from it. (i.e., despite having developed innovative new software) c. Every person selling &quot;The Big Issue&quot; might otherwise be asking for spare change. (i.e., if he/she weren't selling &quot;The Big Issue&quot;) The examples in (24) involve binding in the interpretation of discourse adverbials. In (24a), the temporal use of then locates each hiring event after the corresponding software development. Likewise in (24b), the adversative use of nevertheless signals each developer's unexpected failure to turn the corresponding profit. And in (24c), otherwise envisions each person's begging if that person weren't selling &quot;The Big Issue&quot;. Such bound interpretations require variables in the semantic representations and alternative values for them in some model--hence the representation given in (22). Indeed, it is clear that the binding here has to be the discourse kind, not the syntactic kind, for the same reason as in (23), although we cannot imagine anyone arguing otherwise, since discourse adverbials have always been treated as elements of discourse interpretation. So the variables must be the discourse variables usually used to translate other kinds of discourse anaphors.</Paragraph> <Paragraph position="14"> These arguments have been directed at the behavioral similarity between discourse adverbials and what we normally take to be discourse anaphors. But this isn't the only reason to recognize discourse adverbials as anaphors: In the next section, we suggest a framework for anaphora that is broad enough to include discourse adverbials as well as definite and demonstrative pronouns and NPs, along with other discourse phenomena that have been argued to be anaphoric, such as VP ellipsis (Hardt 1992, 1999; Kehler 2002), tense (Partee 1984; Webber 1988) and modality (Kibble 1995; Frank and Kamp 1997; Stone and Hardt 1999).</Paragraph> </Section> </Section> <Section position="4" start_page="554" end_page="583" type="metho"> <SectionTitle> 3. A Framework for Anaphora </SectionTitle> <Paragraph position="0"> Here we show how only a single extension to a general framework for discourse anaphora is needed to cover discourse adverbials. The general framework is presented in Section 3.1, and the extension in Section 3.2.</Paragraph> <Section position="1" start_page="554" end_page="556" type="sub_section"> <SectionTitle> 3.1 Discourse Referents and Anaphor Interpretation </SectionTitle> <Paragraph position="0"> The simplest discourse anaphors are coreferential: definite pronouns and definite NPs that denote one (or more) discourse referents in focus within the current discourse 6 Although rhetorical structure theory (RST) (Mann and Thompson 1998) was developed as an account of the relation between adjacent units within a text, Marcu's guide to RST annotation (Marcu 1999) has added an &quot;embedded&quot; version of each RST relation in order to handle examples such as (24). Although this importantly recognizes that material in an embedded clause can bear a semantic relation to its matrix clause, it does not contribute to understanding the nature of the phenomenon.</Paragraph> <Paragraph position="1"> Webber et al. Anaphora and Discourse Structure context. (Under coreference we include split reference, in which a plural anaphor such as the companies denotes all the separately mentioned companies in focus within the discourse context.) Much has been written about the factors affecting what discourse referents are taken to be in focus. For a recent review by Andrew Kehler, see chapter 18 of Jurafsky and Martin (2000). For the effect of different types of quantifiers on discourse referents and focus, see Kibble (1995).</Paragraph> <Paragraph position="2"> Somewhat more complex than coreference is indirect anaphora (Hellman and Fraurud 1996) (also called partial anaphora [Luperfoy 1992], textual ellipsis [Hahn, Markert, and Strube 1996], associative anaphora [Cosse 1996] bridging anaphora [Clark 1975; Clark and Marshall 1981; Not, Tovena, and Zancanaro 1999], and inferrables [Prince 1992]), in which the anaphor (usually a definite NP) denotes a discourse referent associated with one (or more) discourse referents in the current discourse context; for example, (25) Myra darted to a phone and picked up the receiver.</Paragraph> <Paragraph position="3"> Here the receiver denotes the receiver associated with (by virtue of being part of) the already-mentioned phone Myra darted to.</Paragraph> <Paragraph position="4"> Coreference and indirect anaphora can be uniformly modeled by saying that the discourse referent e a denoted by an anaphoric expression a is either equal to or associated with an existing discourse referent e</Paragraph> <Paragraph position="6"> ). But coreference and associative anaphora do not exhaust the space of constructs that derive all or part of their sense from the discourse context and are thus anaphoric. Consider &quot;other NPs&quot; (Bierner 2001a; Bierner and Webber 2000; Modjeska 2001, 2002), as in: (26) Sue grabbed one phone, as Tom darted to the other phone.</Paragraph> <Paragraph position="7"> Although &quot;other NPs&quot; are clearly anaphoric, should the referent of the other phone (e a )--the phone other than the one Sue grabbed (e r )--simply be considered a case of</Paragraph> <Paragraph position="9"> )? Here are two reasons why they should not.</Paragraph> <Paragraph position="10"> First, in all cases of associative anaphora discussed in the literature, possible associations have depended only on the antecedent e r and not on the anaphor. For example, only antecedents that have parts participate in whole-part associations (e.g., phone - receiver). Only antecedents with functional schemata participate in schema-based associations (e.g., lock - key). In (26), the relationship between e a , the referent of the other phone, and its antecedent, e r , depends in part on the anaphor, and not just on the antecedent--in particular, on the presence of the word other. Second, we also have examples such as (27) Sue lifted the receiver as Tom darted to the other phone. in which the referent of the other phone (e a ) is the phone other than the phone associated with the receiver that Sue lifted. Together, these two points argue for a third possibility, in which an anaphoric element can convey a specific function f a that is idiosyncratic to the anaphor, which may be applied to either e r or an associate of e r . The result of that application is e a . For want of a better name, we will call these lexically specified anaphors.</Paragraph> <Paragraph position="11"> Other lexically specified anaphors include noun phrases headed by other (example (28)), NPs with such but no postmodifying as phrase (example (29)), comparative 7 Modjeska (2001) discovered such examples in the British National Corpus. Computational Linguistics Volume 29, Number 4 NPs with no postmodifying than phrase (example (30)), and the pronoun elsewhere (example (31)) (Bierner 2001b) (28) Some dogs are constantly on the move. Others lie around until you call them.</Paragraph> <Paragraph position="12"> (29) I saw a 2kg lobster in the fish store yesterday. The fishmonger said it takes about five years to grow to such a size.</Paragraph> <Paragraph position="13"> (30) Terriers are very nervous. Larger dogs tend to have calmer dispositions. (31) I don't like sitting in this room. Can we move elsewhere? To summarize the situation with anaphors so far, we have coreference when e</Paragraph> <Paragraph position="15"/> </Section> <Section position="2" start_page="556" end_page="557" type="sub_section"> <SectionTitle> 3.2 Discourse Adverbials as Lexical Anaphors </SectionTitle> <Paragraph position="0"> There is nothing in this generalized approach to discourse anaphora that requires that the source of e r be an NP, or that the anaphor be a pronoun or NP. For example, the antecedent e r of a singular demonstrative pronoun (in English, this or that) is often an eventuality that derives from a clause, a sentence, or a larger unit in the recent discourse (Asher 1993; Byron 2002; Eckert and Strube 2000; Webber 1991). We will show that this is the case with discourse adverbials as well. The extension we make to the general framework presented above in order to include discourse adverbials as discourse anaphors is to allow more general functions f a to be associated with lexically specified anaphors. In particular, for the discourse adverbials considered in this article, the function associated with an adverbial maps its anaphoric argument--an eventuality derived from the current discourse context--to a function that applies to the interpretation of the adverbial's matrix clause (itself an eventuality). The result is a binary relation that holds between the two eventualities and is added to the discourse context. For example, in (32) John loves Barolo. So he ordered three cases of the '97. But he had to cancel the order because he then discovered he was broke.</Paragraph> <Paragraph position="1"> then, roughly speaking, contributes the fact that its matrix clause event (John's finding he was broke) is after the anaphorically derived event of his ordering the wine. Similarly, in (33) John didn't have enough money to buy a mango. Instead, he bought a guava.</Paragraph> <Paragraph position="2"> instead contributes the fact that its matrix clause event (buying a guava) is an alternative to the anaphorically derived event of buying a mango. The relation between the two sentences is something like result,asinSo instead, he bought a guava. 8 Words and phrases that function as discourse adverbials usually have other roles as well: For example, otherwise also serves as an adjectival modifier, as in I was otherwise occupied with grading exams. This overloading of closed-class lexico-syntactic items is not unusual in English, and any ambiguities that arise must be handled as part of the normal ambiguity resolution process. Webber et al. Anaphora and Discourse Structure Note that our only concern here is with the compositional and anaphoric mechanisms by which adverbials contribute meaning. For detailed analysis of the lexical semantics of adverbials (but no attention to mechanism), the reader is referred to Jayes and Rossari (1998a, 1998b, Lagerwerf (1998), Traugott (1995, 1997), and others. Formally, we represent the function that a discourse adverbial a contributes as a l-expression involving a binary relation R a that is idiosyncratic to a, one of whose arguments (represented here by the variable EV) is resolved anaphorically:</Paragraph> <Paragraph position="4"> ) has been supplied by the discourse context and the other (s) compositionally from syntax.</Paragraph> <Paragraph position="5"> Note that this is a formal model, meant to have no implications for how processing takes place. We have not tried at this stage to instantiate our view of how discourse adverbials are resolved in the context of (simultaneous) sentence-level and discourse-level processing. Our basic view is that resolution is initiated when the discourse adverbial (a) is encountered. As a's matrix clause S is incrementally parsed and interpreted, producing eventuality s, the resolution process polls the discourse context and either finds an appropriate eventuality e i (or creates one by a bridging inference, as illustrated in the next section) such that R</Paragraph> <Paragraph position="7"> ) makes sense with respect to the discourse so far. As is the case with resolving a discourse deictic (Asher 1993; Byron 2002; Eckert and Strube 2000; Webber 1991) this resolution process would use syntactic and semantic constraints that it accumulates as the incremental sentence-level parser/interpreter processes S. As with discourse deixis, this is best seen as a constraint satisfaction problem that involves finding or deriving an eventuality from the current discourse context that meets the constraints of the adverbial with respect to the eventuality interpretation of the matrix clause. (Examples of this are given throughout the rest of the article.)</Paragraph> </Section> <Section position="3" start_page="557" end_page="558" type="sub_section"> <SectionTitle> 3.3 A Logical Form for Eventualities </SectionTitle> <Paragraph position="0"> Before using this generalized view of anaphora to show what discourse adverbials contribute to discourse and how they interact with discourse relations that arise from adjacency or explicit discourse connectives, we briefly describe how we represent clausal interpretations in logical form (LF).</Paragraph> <Paragraph position="1"> Essentially, we follow Hobbs (1985) in using a rich ontology and a representation scheme that makes explicit all the individuals and abstract objects (i.e., propositions, facts/beliefs, and eventualities) (Asher 1993) involved in the LF interpretation of an utterance. We do so because we want to make intuitions about individuals, eventualities, lexical meaning, and anaphora as clear as possible. But certainly, other forms of representation are possible.</Paragraph> <Paragraph position="2"> In this LF representation scheme, each clause and each relation between clauses is indexed by the label of its associated abstract object. So, for example, the LF interpretation of the sentence (34) John left because Mary left.</Paragraph> <Paragraph position="3"> ) where the first argument of the asymmetric binary predicate because is the consequent and the second is the eventuality leading to this consequent. Thus when because occurs sentence-medially, as in the above example, the eventuality arguments are in the same order as their corresponding clauses occur in the text. When because occurs sentenceinitially (as in Because Mary left, John did), the interpretation of the second clause (John [left]) will appear as the first argument and the interpretation of the first clause (Mary left) will appear as the second.</Paragraph> <Paragraph position="4"> The set of available discourse referents includes both individuals like j and m, and also abstract objects like e and e . We then represent resolved anaphors by reusing these discourse referents. So, for example, the LF interpretation of the follow-on sentence null (35) This upset Sue.</Paragraph> <Paragraph position="5"> would be written e :upset(DPRO, s) [?] sue(s) where DPRO is the anaphoric variable contributed by the demonstrative pronoun this. Since the subject of upset could be either the eventuality of John's leaving or the fact that he left because Mary left, DPRO could be resolved to either e</Paragraph> <Paragraph position="7"> depending on whether one took Sue to have been upset by (1) John's leaving or (2) that he left because Mary left.</Paragraph> </Section> <Section position="4" start_page="558" end_page="562" type="sub_section"> <SectionTitle> 3.4 The Contribution of Discourse Adverbials to Discourse Semantics </SectionTitle> <Paragraph position="0"> Here we step through some examples of discourse adverbials and demonstrate how they make their semantic contribution to the discourse context. We start with example (32), repeated here as (36): (36) a. John loves Barolo.</Paragraph> <Paragraph position="1"> b. So he ordered three cases of the '97.</Paragraph> <Paragraph position="2"> c. But he had to cancel the order d. because he then discovered he was broke.</Paragraph> <Paragraph position="3"> 9 We are not claiming to give a detailed semantics of discourse connectives except insofar as they may affect how discourse adverbials are resolved. Thus, for example, we are not bothering to distinguish among different senses of because (epistemic vs. nonepistemic), while (temporal vs. concessive), since (temporal vs. causal), etc. Of course, these distinctions are important to discourse interpretation, but they are independent of and orthogonal to the points made in this article. Similarly, Asher (1993) argues that a simple ontology of eventualities is too coarse-grained, and that discourse representations need to distinguish different kinds of abstract objects, including actions, propositions, and facts as well as eventualities. Different discourse connectives will require different kinds of abstract objects as arguments. This distinction is also orthogonal to the points made in this article, because we can understand these abstract referents to be associates of the corresponding Hobbsian eventualities and leave the appropriate choice to the lexical semantics of discourse connectives. Byron (2002) advocates a similar approach to resolving discourse anaphora.</Paragraph> <Paragraph position="4"> Webber et al. Anaphora and Discourse Structure Using the above LF representation scheme and our notation from Section 3.2, namely, * a = the anaphoric expression (here, the discourse adverbial) * R a = the relation name linked with a * S = the matrix clause/sentence containing a * s = the interpretation of S as an abstract object and ignoring, for now, the conjunction because (discussed in section 4), the relevant elements of (36d) can be represented as: , and added to the discourse context.</Paragraph> <Paragraph position="5"> When then is understood temporally, as it is above, as opposed to logically, it requires a culminated eventuality from the discourse context as its first argument (which Vendler (1967) calls an achievement or an accomplishment). The ordering event in (36b) is such a Vendlerian accomplishment. In example (37), though, there is no culminated eventuality in the discourse context for then), to take as its first argument.</Paragraph> <Paragraph position="6"> (37) a. Go west on Lancaster Avenue.</Paragraph> <Paragraph position="7"> b. Then turn right on County Line.</Paragraph> <Paragraph position="8"> :turn-right(you, county line) and the unresolved interpretation of (37b) is thus [l x.after(x, EV)]e Computational Linguistics Volume 29, Number 4 As for resolving EV, in a well-known article, Moens and Steedman (1988) discuss several ways in which an eventuality of one type (e.g., a process) can be coerced into an eventuality of another type (e.g., an accomplishment, which Moens and Steedman call a culminated process). In this case, the matrix argument of then (the eventuality of turning right on County Line) can be used to coerce the process eventuality in (37b) into a culminated process of going west on Lancaster Avenue until County Line. We treat this coercion as a type of associative or bridging inference, as in the examples discussed in section 3.1. That is, ) That is, the eventuality of turning right onto County Line is after that of going west on Lancaster Avenue to County Line. This proposition would be indexed and added to the discourse context.</Paragraph> <Paragraph position="9"> It is important to stress here that the level of representation we are concerned with is essentially an LF for discourse. Any reasoning that might then have to be done on the content of LFs might then require making explicit the different modal and temporal contexts involved, their accessibility relations, the status of abstract objects as facts, propositions or eventualities, and so on. But as our goal here is primarily to capture the mechanism by means of which discourse adverbials are involved in discourse structure and discourse semantics, we will continue to assume for as long as possible that an LF representation will suffice.</Paragraph> <Paragraph position="10"> Now it may appear as if there is no difference between treating adverbials as anaphors and treating them as structural connectives, especially in cases like (37) in which the antecedent comes from the immediately left-adjacent context, and in which the only obvious semantic relation between the adjacent sentences appears to be the one expressed by the discourse adverbial. (Of course, there may also be a separate intentional relation between the two sentences [Moore and Pollack 1992], independent of the relation conveyed by the discourse adverbial.) One must distinguish, however, between whether a theory allows a distinction to be made and whether that distinction needs to be made in a particular case. It is clear that there are many examples in which the two approaches (i.e., a purely structural treatment of all connectives, versus one that treats adverbials as linking into the discourse context anaphorically) appear to make the same prediction. We have already, however, demonstrated cases in which a purely structural account makes the wrong prediction, and in the next section, we will demonstrate the additional power of an account that allows for two relations between an adverbial's matrix clause or sentence and the previous discourse: one arising from the anaphoric connection and the other inferred from adjacency or conveyed explicitly by a structural connective. Before closing this section, we want to step through examples (19)-(20), repeated here as examples (38)-(39).</Paragraph> <Paragraph position="11"> (38) If the light is red, stop. Otherwise you'll get a ticket.</Paragraph> <Paragraph position="12"> (39) If the light is red, stop. Otherwise go straight on.</Paragraph> <Paragraph position="13"> Webber et al. Anaphora and Discourse Structure Roughly speaking, otherwise conveys that the complement of its anaphorically derived argument serves as the condition under which the interpretation of its structural argument holds. (This complement must be with respect to some contextually relevant ), then we can represent the interpretation of otherwise as l x.if(VE, x), where complement(VE, EV) where variable EV is resolved anaphorically to an eventuality in the current discourse context that admits a complement. That is, otherwise requires a contextually relevant complement to its antecedent and asserts that if that complement holds, the argument to the l-expression will as well. The resulting l-expression applies to the interpretation of the matrix clause of otherwise, resulting in the conditional's being added to the ) As we showed in section 2.2, different ways of resolving the anaphoric argument lead to different interpretations. In (38), the anaphoric argument is resolved to e :stop(you), 10 Kruijff-Korbayov'a and Webber (2001a) demonstrate that the information structure of sentences in the previous discourse (theme-rheme partitioning, as well as focus within theme and within rheme [Steedman 2000a]) can influence what eventualities e r are available for resolving the anaphorically derived argument of otherwise. This then correctly predicts different interpretations for &quot;otherwise&quot; in (i) and (ii): (i) Q. How should I transport the dog? A. You should carry the dog. Otherwise you might get hurt.</Paragraph> <Paragraph position="14"> (ii) Q. What should I carry? A. You should carry the dog. Otherwise you might get hurt.</Paragraph> <Paragraph position="15"> In both (i) and (ii), the questions constrain the theme/rheme partition of the answer. Small capitals represent focus within the rheme. In (i), the otherwise clause will be interpreted as warning the hearer (H) that H might get hurt if he/she transports the dog in some way other than carrying it (e.g., H might get tangled up in its lead). In (ii), the otherwise clause warns H that he/she might get hurt if what she is carrying is not the dog (e.g., H might be walking past fanatical members of the Royal Kennel Club). Computational Linguistics Volume 29, Number 4 whereas in (39), it is resolved to e (If the light is not red, go straight on.) We have not been specific about how the anaphoric argument of otherwise (or of any other discourse adverbial) is resolved, other than having it treated as a constraint satisfaction problem. This is the subject of current and future work, exploring the empirical properties of resolution algorithms with data drawn from appropriately annotated corpora and from psycholinguistic studies of human discourse interpretation. To this end, Creswell et al. (2002) report on a preliminary annotation study of discourse adverbials and the location and type of their antecedents. This initial effort involves nine discourse adverbials--three each from the classes of concessive, result, and reinforcing (additive) conjuncts given in Quirk et al. (1972). Meanwhile, Venditti et al. (2002) present a preliminary report on the use of a constraint satisfaction model of interpretation, crucially combining anaphoric and structural reasoning about discourse relations, to predict subjects' on-line interpretation of discourses involving stressed pronouns. In addition, two proposals have recently been submitted to construct a larger and more extensively annotated corpus, covering more adverbials, based on what we have learned from this initial effort. This more extensive study would be an adequate basis for developing resolution algorithms.</Paragraph> </Section> <Section position="5" start_page="562" end_page="574" type="sub_section"> <SectionTitle> 3.5 Summary </SectionTitle> <Paragraph position="0"> In this section, we have presented a general framework for anaphora with the following features: * Anaphors can access one or more discourse referents or entities associated with them through bridging inferences. These are sufficient for interpreting anaphoric pronouns, definite NPs and demonstrative NPs, allowing entities to be evoked by NPs or by clauses. In the case of clauses, this may be on an as-needed basis, as in Eckert and Strube (2000).</Paragraph> <Paragraph position="1"> * A type of anaphor a that we call lexically specified can also contribute additional meaning through a function f a that is idiosyncratic to a, that can be applied to either an existing discourse referent or an entity associated with it through a bridging inference. In the case of the premodifier other, f a applied to its argument produces contextually 11 With respect to how many discourse adverbials there are, Quirk et al. (1972) discuss 60 conjunctions and discourse adverbials under the overall heading time relations and 123 under the overall heading conjuncts. Some entries appear under several headings, so that the total number of conjunctions and discourse adverbials they present is closer to 160. In another enumeration of discourse adverbials, Forbes and Webber (2002) start with all annotations of sentence-level adverbials in the Penn Treebank, then filter them systematically to determine which draw part of their meaning from the preceding discourse and how they do so. What we understand from both of these studies is that there are fewer than 200 adverbials to be considered, many of which are minor variations of one another (in contrast, by contrast, by way of contrast, in comparison, by comparison, by way of comparison that are unlikely to differ in their anaphoric properties, and some of which, such as contrariwise, hitherto, and to cap it all, will occur only rarely in a corpus of modern English.</Paragraph> <Paragraph position="2"> Webber et al. Anaphora and Discourse Structure relevant alternatives to that argument. In the case of the premodifier such, it yields a set of entities that are similar to its argument in a contextually relevant way.</Paragraph> <Paragraph position="3"> * Discourse adverbials are lexically specified anaphors whose meaning function f a is a l-expression involving a binary relation R a that is idiosyncratic to a, one of whose arguments is resolved anaphorically and the other is provided compositionally, when the l-expression is applied to a's matrix clause interpreted as an eventuality s.</Paragraph> <Paragraph position="4"> In the next section, we move on to consider how the presence of both a semantic relation associated with a discourse adverbial and a semantic relation associated with the adjacency of two clauses or a structural connective between them allows for interesting interactions between the two.</Paragraph> <Paragraph position="5"> 4. Patterns of Anaphoric Relations and Structural/Inferred Relations Prior to the current work, researchers have treated both explicit structural connectives (coordinating and subordinating conjunctions, and &quot;paired&quot; conjunctions) and discourse adverbials simply as evidence for a particular structural relation holding between adjacent units. For example, Kehler (2002) takes but as evidence of a contrast relation between adjacent units, in general as evidence of a generalization relation, in other words as evidence of an elaboration relation, therefore as evidence of a result relation, because as evidence of an explanation relation, and even though as evidence of a denial of preventer relation (Kehler 2002, Section 2.1). Here Kehler has probably correctly identified the type of relation that holds between elements, but not which elements it holds between.</Paragraph> <Paragraph position="6"> In one respect, we follow previous researchers, in that we accept that when clauses, sentences, or larger discourse units are placed adjacent to one another, listeners infer a relation between the two, and that the structural connective (coordinate or subordinate conjunction) gives evidence for the relation that is intended to hold between them. Because we take discourse adverbials to contribute meaning through an anaphoric connection with the previous discourse, however, this means that there may be two relations on offer and opens up the possibility that the relation contributed by the discourse adverbial can interact in more than one way with the relation conveyed by a structural connective or inferred through adjacency. Below we show that this prediction is correct.</Paragraph> <Paragraph position="7"> We start from the idea that, in the absence of an explicit structural connective, defeasible inference correlates with structural attachment of adjacent discourse segments in discourse structure, relating their interpretations. The most basic relation is that the following segment in some way describes the same object or eventuality as the one it abuts (elaboration). But evidence in the segments can lead (via defeasible inference) to a more specific relation, such as one of the resemblance relations (e.g., parallel, contrast, exemplification, generalisation), or cause-effect relations (result, explanation, violated expectation), or contiguity relations (narration) described in Hobbs (1990) and Kehler (2002). If nothing more specific can be inferred, the relation will remain simply elaboration. What explicit structural connectives can do is convey relations that are not easy to convey by defeasible inference (e.g., if, conveying condition, and or, conveying disjunction)or provide nondefeasible evidence for an inferrable relation (e.g., yet, so, and because). Discourse adverbials can interact with structural connectives, with adjacency-triggered defeasible inference, and with each other. To describe the ways in which we Computational Linguistics Volume 29, Number 4 have so far observed discourse adverbials to interact with relations conveyed structurally, we extend the notation used in the previous section: * a = discourse adverbial * R a = the name of the relation associated with a * S = the matrix clause/sentence of a * s = the logical form (LF) interpretation of S adding the following: * D = the discourse unit that is left-adjacent to S, to which a relationship holds by either inference or a structural connective * d = the LF interpretation of D * R = the name of the relation that holds with d Although d is one argument of R, we show below that the other argument of R may be one of at least two different abstract objects.</Paragraph> <Paragraph position="8"> Case 1: s separately serves as an argument to both R a and R. This is the case that holds in example (36) (repeated here): (36) a. John loves Barolo.</Paragraph> <Paragraph position="9"> b. So he ordered three cases of the '97.</Paragraph> <Paragraph position="10"> c. But he had to cancel the order d. because he then discovered he was broke.</Paragraph> <Paragraph position="11"> We have already seen that the interpretation of the clause in (36d) following because</Paragraph> <Paragraph position="13"> ) is an argument of R. In case 1, it is the interpretation of the adverbial's matrix clause s that serves as one argument to the discourse relation R.In contrast, in case 2, that argument is filled by the relation contributed by the discourse adverbial (itself an abstract object available for subsequent reference). In both cases, the other argument to R is d.</Paragraph> <Paragraph position="14"> One configuration in which case 2 holds is with the discourse adverbial otherwise. Recall from section 3.4 that the interpretation of otherwise involves a conditional relation between the complement of its anaphoric argument and the interpretation s of its matrix clause: [lx.if(VE,x)] s [?] if(VE,s), where complement(VE,EV) With variable EV resolved to an eventuality in the discourse context, it is the resulting relation (viewed as an abstract object) that serves as one argument to R, with d serving as the other. We can see this most clearly by considering variants of examples (38) and (39) that contain an explicit connective between the clauses. In (38), the conjunction because is made explicit (example (40)), and in (39), the connective is simply and or but (example (41)).</Paragraph> <Paragraph position="15"> (40) If the light is red, stop, because otherwise you'll get a ticket.</Paragraph> <Paragraph position="16"> Computational Linguistics Volume 29, Number 4 red) you should stop is in contrast to the situation that if the light is not red, you should go straight on.</Paragraph> <Paragraph position="17"> As is clear from the original pair of examples (38) and (39), interpretations can arise through adjacency-triggered inference that are similar to those that arise with an explicit connective. In either case, the above treatment demonstrates that there is no need for a separate otherwise relation, as proposed in rhetorical structure theory (Mann and Thompson 1988). We are not, however, entirely clear at this point when case 1 holds and when case 2 does. A more careful analysis is clearly required. a is parasitic on R. Case 3 appears to apply with discourse adverbials such as for example and for instance. The interpretation of such adverbials appears to be parasitic on the relation associated with a structural connective or discourse adverbial to their left, or on an inferred relation triggered by adjacency. The way to understand this is to first consider intraclausal for example, where it follows the verb, as in (42) Q. What does this box contain? A. It contains, for example, some hematite.</Paragraph> <Paragraph position="18"> The interpretation of for example here involves abstracting the meaning of its matrix structure with respect to the material to its right, then making an assertion with respect to this abstraction. That is, if the LF contributed by the matrix clause of (42A) is, roughly, i. contain(box1,hematite1) then the LF resulting from the addition of for example can be written either with set notation (as in (ii)), taking an entity to exemplify a set, or with l-notation (as in (iii)), taking an entity to exemplify a property: ii. exemplify(hematite1, {X |contain(box1,X)}) iii. exemplify(hematite1, lX . contain(box1,X)) Both express the fact that hematite is an example of what is contained in the box. Since one can derive (i) logically from either (ii) or (iii), one might choose to retain only (ii) or (iii) and derive (i) if and when it is needed. In the remainder of the article, we use the l notation given in (iii). Note that from the perspective of compositional semantics, for example resembles a quantifier, in that the scope of its interpretation is not isomorphic to its syntactic position. Thus producing an interpretation for for example will require techniques similar to those that have long been used in interpreting quantifiers (Woods, 1978; Barwise and Cooper 1981). We take this up again in section 5.</Paragraph> <Paragraph position="19"> If we look at the comparable situation in discourse, such as (43)-(44), where for example occurs to the right of a discourse connective, it can also be seen as abstracting 13 A much finer-grained treatment of the semantics of otherwise in terms of context-update potential is given in Kruijff-Korbayov'a and Webber (2001b). Here we are just concerned with its interaction with structural connectives and adjacency-triggered relations.</Paragraph> <Paragraph position="20"> 14 The material to the right of for example can be any kind of constituent, including such strange ones as John gave, for example, a flower to a nurse.</Paragraph> <Paragraph position="21"> Here, a flower to a nurse would be an example of the set of object-recipient pairs within John's givings. Such nonstandard constituents are also found with coordination, which was one motivation for combinatory categorial grammar (Steedman 1996). This just illustrates another case in which such nonstandard constituents are needed.</Paragraph> <Paragraph position="22"> Webber et al. Anaphora and Discourse Structure the interpretation of its discourse-level matrix structure, with respect to the material to its right: (43) John just broke his arm. So, for example, he can't cycle to work now. (44) You shouldn't trust John because, for example, he never returns what he borrows.</Paragraph> <Paragraph position="23"> In (43), the connective so leads to result(s,d) being added to the discourse, where s is the interpretation of John can't cycle to work now, and d is the interpretation of John just broke his arm. For example then abstracts this relation with respect to the material to its right (i.e., s), thereby contributing exemplify(s, lX.result(X, d)) That is, John can't cycle to work is an example of what results from John's breaking his arm. Similarly, because in (44) leads to explanation(s,d) being added to the discourse, where s is the interpretation of he never returns what he borrows, d is the interpretation of you shouldn't trust John, and for example adds exemplify(s, lX.explanation(X,d)) that is, that s is an example of the reasons for not trusting John.</Paragraph> <Paragraph position="24"> For example interacts with discourse adverbials in the same way: (45) Shall we go to the Lincoln Memorial? Then, for example, we can go to the White House.</Paragraph> <Paragraph position="25"> (46) As a money manager and a grass-roots environmentalist, I was very dis- null appointed to read in the premiere issue of Garbage that The Wall Street Journal uses 220,000 metric tons of newsprint each year, but that only 1.4% of it comes from recycled paper. By contrast, the Los Angeles Times, for example, uses 83% recycled paper. [WSJ, from Penn Treebank /02/wsj-0269] In example (45), the resolved discourse adverbial then leads to after(s,d) being added to the discourse context, where s is the interpretation of we can go to the White House, d is the interpretation of we shall go to the Lincoln Memorial, and for example adds exemplify(s, lX.after(X,d)) that is, that s is an example of the events that [can] follow going to the Lincoln Memorial. (As already noted, we are being fairly fast and loose regarding tense and modality, in the interests of focusing on the types of interactions.) Computational Linguistics Volume 29, Number 4 In example (46), the resolved discourse anaphor by contrast contributes contrast(s,d), where s is the interpretation of the Los Angeles Times's using 83% recycled paper and d is the intepretation of only 1.4% of it [newsprint used by the WSJ] comes from recycled paper. For example then contributes exemplify(s, lX.contrast(X,d)) that is, that s is one example of contrasts with the WSJ's minimal use of recycled paper.</Paragraph> <Paragraph position="26"> What occurs with discourse connectives and adverbials can also occur with relations added through adjacency-triggered defeasible inference, as in (47) You shouldn't trust John. For example, he never returns what he borrows. explanation(d,s) exemplify(s, lX.explanation(d,X)) Here, as in (44), the relation provided by adjacency-triggered inference is R = explanation, which is then used by for example.</Paragraph> <Paragraph position="27"> But what about the many cases in which only exemplify seems present, as in (48) In some respects they [hypertext books] are clearly superior to normal books, for example they have database cross-referencing facilities ordinary volumes lack. (British National Corpus, CBX 1087) (49) He [James Bellows] and his successor, Mary Anne Dolan, restored respect for the editorial product, and though in recent years the paper had been limping along on limited resources, its accomplishments were notable. For example, the Herald consistently beat its much-larger rival on disclosures about Los Angeles Mayor Tom Bradley's financial dealings.</Paragraph> <Paragraph position="28"> There are at least two explanations: One is that for example simply provides direct nondefeasible evidence for exemplify, which is the only relation that holds. The other explanation follows the same pattern as the examples given above, but with no further relation than elaboration(s,d). That is, we understand in (48) that having database cross-referencing facilities elaborates the respects in which hypertext books are superior to normal books, whereas in (49), we understand that the Herald's [newspaper] consistently beating its much-larger rival elaborates the claim that its accomplishments were notable. This elaboration relation is then abstracted (in response to for example) to produce: exemplify(s, lX.elaboration(X, d)) that is, that this is one example of many possible elaborations. Because this is more specific than elaboration and seems to mean the same as exemplify(s,d), one might simply take it to be the only relation that holds. Given that so many naturally occuring instances of for example occur with elaboration, it is probably useful to persist with the above shorthand. But it shouldn't obscure the regular pattern that appears to hold. Before going on to case 4, we should comment on an ambiguity associated with for example. When for example occurs after an NP, a PP, or a clause that can be interpreted as a general concept or a set, it can contribute a relation between the general concept or set and an instance, rather than being parasitic on another relation. For example, in: Webber et al. Anaphora and Discourse Structure (50) In the case of the managed funds they will be denominated in a leading currency, for example US dollar, ...(BNC CBX 1590) for example relates the general concept denoted by a leading currency to a specific instance, U.S. dollars. (In British English, the BNC shows that most such examples occur with such as--i.e., in the construction such as for example. This paraphrase does not work with the predicate-abstracting for example that is of primary concern here, as in example (42).) But for example occurring after an NP, a PP, or a clause can, alternatively, contribute a more subtle parasitic relationship to the previous clause, as in (51) All the children are ill, so Andrew, for example, can't help out in the shop. This differs from both (43) and (50). That is, one cannot paraphrase (51) as (52) as in (43) where for example follows so: (52) All the children are ill, so for example Andrew can't help out in the shop. Example (52) simply specifies an example consequence of all the children being ill, as does (53) All the children are ill, so for example one of us has to be at home at all times.</Paragraph> <Paragraph position="29"> In contrast, (51) specifies an example consequence for Andrew, as one of the children. Support for this comes from the fact that in (52), Andrew doesn't have to be one of the children: He could be their nanny or child minder, now stuck with dealing with a lot of sick kids. But (51) is not felicitous if Andrew is not one of the children. We suspect here the involvement of information structure (Steedman 2000a): Whereas the interpretation conveyed by for example is parasitic on the adjacency relation (result in example (51)), its position after the NP Andrew in (51) may indicate a contrastive theme with respect to the previous clause, according to which Andrew in contrast to the other children suffers this particular consequence. But more work needs to be done on this to gain a full understanding of what is going on.</Paragraph> <Paragraph position="30"> is a defeasible rule that incorporates R. Case 4 occurs with discourse adverbials that carry the same presupposition as the discourse connectives although and the concessive sense of while (Lagerwerf 1998). Case 4 shares one feature with case 1, in that the discourse relation R conveyed by a structural connective or inferred from adjacency holds between s (the interpretation of the adverbial's matrix clause) and d (the interpretation of the left-adjacent discourse unit). Where it differs is that the result is then incorporated into the presupposition of the discourse adverbial. This presupposition, according to Lagerwerf (1998), has the nature of a presupposed (or conventionally implicated) defeasible rule that fails to hold in the current situation. He gives as an example (54) Although Greta Garbo was called the yardstick of beauty, she never married. null This asserts both that Greta Garbo was called the yardstick of beauty and that she never married. The first implies that Greta Garbo was beautiful. The example also Computational Linguistics Volume 29, Number 4 presupposes that, in general, if a woman is beautiful, she will marry. If such a presupposition can be accommodated, it will simply be added to the discourse context. If not, the hearer will find the utterance confusing or possibly even insulting. We argue here that the same thing happens with the discourse adverbials nevertheless and though. The difference is that, with discourse adverbials, the antecedent to the rule derives anaphorically from the previous discourse, whereas the consequent derives from the adverbial's matrix clause. (With the conjunctions although and concessive while, both arguments are provided structurally.) We first illustrate case 4 with two examples in which nevertheless occurs in the main clause of a sentence containing a preposed subordinate clause. The subordinate conjunction helps clarify the relation between the clauses that forms the basis for the presupposed defeasible rule. After these, we give a further example in which the relation between the adjacent clauses comes through inference.</Paragraph> <Paragraph position="31"> (55) While John is discussing politics, he is nevertheless thinking about his fish. In (55), the conjunction while conveys a temporal relation R between the two clauses it connects: we will write informally as</Paragraph> <Paragraph position="33"> Normally, whatever one does during the time one is discussing politics, it is not thinking about one's fish.</Paragraph> <Paragraph position="34"> This rule uses Asher and Morreau's (1991) defeasible implication operator (>) and abstracts over the individual (John), which seems appropriate for the general statement conveyed by the present tense of the utterance.</Paragraph> <Paragraph position="35"> Similarly, in (56) Even after John has had three glasses of wine, he is nevertheless able to solve difficult math problems.</Paragraph> <Paragraph position="36"> the conjunction after contributes a relation between the two clauses it connects: Webber et al. Anaphora and Discourse Structure Again, we have abstracted over the individual, as the presupposed defeasible rule associated with the present-tense sentence appears to be more general than a statement about a particular individual.</Paragraph> <Paragraph position="37"> On the other hand, in the following example illustrating a presupposed defeasible rule and a discourse relation associated with adjacency, it seems possible for the presupposed defeasible rule to be about John himself: (57) John is discussing politics. Nevertheless, he is thinking about his fish. Here the discourse relation between the two clauses, each of which denotes a specific Normally what occurs during John's discussing politics is not John's thinking about his fish.</Paragraph> <Paragraph position="38"> Lagerwerf (1998) does not discuss how specific or general will be the presupposed defeasible rule that is accommodated or what factors affect the choice. Kruijff-Korbayov'a and Webber (2001a) also punt on the question, when considering the effect of information structure on what presupposed defeasible rule is associated with although. Again, this seems to be a topic for future work.</Paragraph> <Paragraph position="39"> Summary We have indicated four ways in which we have found the relation associated with a discourse adverbial to interact with a relation R triggered by adjacency or conveyed by structural connectives or, in some cases, by another relational anaphor: is a defeasible rule that incorporates R.</Paragraph> <Paragraph position="40"> We do not know whether this list is exhaustive or whether a discourse adverbial always behaves the same way vis-`a-vis other relations. Moreover, in the process of setting down the four cases we discuss, we have identified several problems that we have not addressed, on which further work is needed. Still, we hope that we have convinced the reader of our main thesis: that by recognizing discourse adverbials as doing something different from simply signaling the discourse relation between adjacent discourse units and by considering their contribution as relations in their own right, one can begin to characterize different ways in which anaphoric and structural relations may themselves interact.</Paragraph> <Paragraph position="41"> 15 We speculate that the reason examples such as (55) and (56) sound more natural with the focus particle even applied to the subordinate clause is that even conveys an even greater likelihood that the defeasible rule holds, so nevertheless emphasizes its failure to do so. Computational Linguistics Volume 29, Number 4 5. Lexicalized Grammar for Discourse Syntax and Semantics The question we consider in this section is how the treatment we have presented of discourse adverbials and structural connectives can be incorporated into a general approach to discourse interpretation. There are three possible ways.</Paragraph> <Paragraph position="42"> The first possibility is simply to incorporate our treatment of adverbials and connectives into a sentence-level grammar, since such grammars already cover the syntax of sentence-level conjunction (both coordinate and subordinate) and the syntax of adverbials of all types. The problem with this approach is that sentence-level grammars, whether phrasal or lexicalized, stop at explicit sentence-level conjunction and do not provide any mechanism for forming the meaning of multiclausal units that cross sentence-level punctuation. Moreover, as we have already shown in section 3, the interpretation of discourse adverbials can interact with the implicit relation between adjacent sentences, as well as with an explicitly signaled relation, so that a syntax and compositional semantics that stops at the sentence will not provide all the structures and associated semantics needed to build the structures and interpretations of interest. The second possibility is to have a completely different approach to discourse-level syntax and semantics than to sentence-level syntax and semantics, combining (for example) a definite clause grammar with rhetorical structure theory. But as we and others have already noted, this requires discourse semantics reaching further and further into sentence-level syntax and semantics to handle relations between main and embedded clauses, and between embedded clauses themselves, as in example (58). (58) If they're drunk and they're meant to be on parade and you go to their room and they're lying in a pool of piss, then you lock them up for a day.</Paragraph> <Paragraph position="43"> (The Independent, June 17, 1997) Thus it becomes harder and harder to distinguish the scope of discourse-level syntax and semantics from that at the sentence-level.</Paragraph> <Paragraph position="44"> The third possibility is to recognize the overlapping scope and similar mechanisms and simply extend a sentence-level grammar and its associated semantic mechanisms to discourse. The additional responsibilities of the grammer would be to account for the formation of larger units of discourse from smaller units; the projection of the interpretation of smaller discourse units onto the interpretation of the larger discourse units they participate in; and the effect of discourse unit interpretation on the evolving discourse model. There are two styles of grammar one could use for this: (1) a phrase structure grammar (PSG) extended to discourse, as in Figure 6, or (2) a lexicalized grammar that extends to discourse, a sentence-level lexicalized grammar such as tree-adjoining grammar (Joshi, 1987; XTAG-Group 2001) or combinatory categorial grammar (CCG) (Steedman 1996, 2000b).</Paragraph> <Paragraph position="45"> Whereas Polanyi and van den Berg (1996) extend a PSG to discourse, we argue for extending a lexicalized grammar, even though TAG and CCG are weakly context-sensitive (CS) and the power needed for a discourse grammar with no crossing dependencies is only context-free (section 2.1). Our argument is based on our desire to use a discourse grammar in natural language generation (NLG). It is well-known that context-free PSGs (CF PSGs) set up a complex search space for NLG. A discourse grammar specified in terms of phrase structure rules such as those shown in Figure 6 doesn't provide sufficient guidance when reversed for use in generating discourse. For example, one might end up having to guess randomly how many sentences and connectives one had, in what order, before being able to fill in the sentences and connectives with any content. More generally, trying to generate exactly a given semantics Webber et al. Anaphora and Discourse Structure Seg := SPunct Seg |Seg SPunct |SPunct | on the one hand Seg on the other hand Seg | not only Seg but also Seg</Paragraph> <Paragraph position="47"> PS rules for a discourse grammar.</Paragraph> <Paragraph position="48"> when semantics underspecifies syntactic dependency (as discourse semantics must, on our account) is known to be intractable (Koller and Striegnitz 2002). An effective solution is to generate semantics and syntax simultaneously, which is straightforward with a lexicalized grammar (Stone et al. 2001).</Paragraph> <Paragraph position="49"> Given the importance of various types of inference in discourse understanding, there is a second argument for using a lexicalized discourse grammar that derives from the role of implicature in discourse. Gricean reasoning about implicatures requires a hearer be able to infer the meaningful alternatives that a speaker had in composing a sentence. With lexicalization, these alternatives can be given by a grammar, allowing the hearer, for example, to ask sensible questions like &quot;Why did the speaker say 'instead' here instead of nothing at all?&quot; and draw implicatures from this. A CF PSG, on the other hand, might suggest questions like &quot;Why did the speaker say two sentences rather than one here?&quot; which seem empirically not to lead to any real implicatures. (On the contrast between choices, which seem to lead to implicatures, and mere alternative linguistic formulations, which do not seem to, see, for example, Dale and Reiter [1995] and Levison [2000].) In several previous papers (Webber, Knott, and Joshi, 2001; Webber et al., 1999a, 1999b), we described how our approach fits into the framework of tree-adjoining grammar. This led to the initial version of a discourse parser (Forbes et al. 2001) in which the same parser that builds trees for individual clauses using clause-level LTAG trees then combines them using discourse-level LTAG trees. Here we simply outline the grammar, called DLTAG (section 5.1), then show how it supports the approach to structural and anaphoric discourse connectives presented earlier (section 5.2).</Paragraph> <Paragraph position="50"> (Of course, one still needs to account for how speakers realize their intentions through text and how what is achieved through a single unit of text contributes to what a speaker hopes to achieve through any larger unit in which it is embedded.</Paragraph> <Paragraph position="51"> Preliminary accounts are given in Grosz and Sidner [1990] and Moser and Moore [1996]. Given the complex relation between individual sentences and speaker intentions, however, it is unlikely that the relation between multisentence discourse and speaker intentions can be modeled in a straightforward way similar to the basically monotonic compositional process that we have discussed in this article for discourse semantics.) Initial trees for a subordinate conjunction: (a) postposed; (b) preposed. D c stands for discourse clause, |indicates a substitution site, and subconj stands for the particular subordinate conjunction that anchors the tree.</Paragraph> </Section> <Section position="6" start_page="574" end_page="577" type="sub_section"> <SectionTitle> 5.1 DLTAG and Discourse Syntax </SectionTitle> <Paragraph position="0"> A lexicalized TAG begins with the notion of a lexical anchor, which can have one or more associated tree structures. For example, the verb likes anchors one tree corresponding to John likes apples, another corresponding to the topicalized Apples John likes, a third corresponding to the passive Apples are liked by John, and others as well. That is, there is a tree for each minimal syntactic construction in which likes can appear, all sharing the same predicate-argument structure. This syntactic/semantic encapsulation is possible because of the extended domain of locality of LTAG.</Paragraph> <Paragraph position="1"> A lexicalized TAG contains two kinds of elementary trees: initial trees that reflect basic functor-argument dependencies and auxiliary trees that introduce recursion and allow elementary trees to be modified and/or elaborated. Unlike the wide variety of trees needed at the clause level, we have found that extending a lexicalized TAG to discourse requires only a few elementary tree structures, possibly because clause-level syntax exploits structural variation in ways that discourse doesn't.</Paragraph> <Paragraph position="2"> 5.1.1 Initial Trees. DLTAG has initial trees associated with subordinate conjunctions, with parallel constructions, and with some coordinate conjuctions. We describe each in turn.</Paragraph> <Paragraph position="3"> In the large LTAG developed by the XTAG project (XTAG-Group 2001) subordinate clauses are seen as adjuncts to sentences or verb phrases (i.e., as auxiliary trees) because they are outside the domain of locality of the verb. In DLTAG, however, it is predicates on clausal arguments (such as coordinate and subordinate conjunctions) that define the domain of locality. Thus, at this level, these predicates anchor initial trees into which clauses substitute as arguments. Figure 7 shows the initial trees for (a) postposed subordinate clauses and (b) preposed subordinate clauses.</Paragraph> <Paragraph position="4"> At both leaves and root is a discourse clause (D c ): a clause or a structure composed of discourse clauses.</Paragraph> <Paragraph position="5"> One reason for taking something to be an initial tree is that its local dependencies can be stretched long distance. At the sentence level, the dependency between apples and likes in Apples John likes is localized in all the trees for likes. This dependency can be stretched long distance, as in Apples, Bill thinks John may like. In discourse, as we noted in section 2, local dependencies can be stretched long distance as well, as in (59) a. Although John is generous, he's hard to find.</Paragraph> <Paragraph position="6"> 16 Although in an earlier paper (Webber and Joshi 1998), we discuss reasons for taking the lexical anchors of the initial trees in Figures 7 and 8 to be feature structures, following the analysis in Knott (1996) and Knott and Mellish (1996), here we just take them to be specific lexical items. Webber et al. Anaphora and Discourse Structure An initial tree for parallel constructions. This particular tree is for a contrastive construction anchored by on the one hand and on the other hand.</Paragraph> <Paragraph position="7"> b. Although John is generous--for example, he gives money to anyone who asks him for it--he's hard to find.</Paragraph> <Paragraph position="8"> (60) a. On the one hand, John is generous. On the other hand, he's hard to find.</Paragraph> <Paragraph position="9"> b. On the one hand, John is generous. For example, suppose you needed some money: You'd only have to ask him for it. On the other hand, he's hard to find.</Paragraph> <Paragraph position="10"> Thus DLTAG also contains initial trees for parallel constructions as in (60). Such an initial tree is shown in Figure 8. Like some initial trees in XTAG (XTAG-Group 2001), such trees can have a pair of anchors. Since there are different ways in which discourse units can be parallel, we assume a different initial tree for contrast (on the one hand...on the other (hand)...), disjunction (either...or...), addition (not only...but also...), and concession (admittedly...but...).</Paragraph> <Paragraph position="11"> Finally, there are initial trees for structural connectives between adjacent sentences or clauses that convey a particular relation between the connected units. One clear example is so, conveying result. Its initial tree is shown in Figure 9. We will have a better sense of what other connectives to treat as structural as a result of annotation efforts of the sort described in Creswell et al. (2002).</Paragraph> <Paragraph position="12"> 5.1.2 Auxiliary Trees. DLTAG uses auxiliary trees in two ways: (1) for discourse units that continue a description in some way, and (2) for discourse adverbials. Again we describe each in turn.</Paragraph> <Paragraph position="13"> 17 For example, one might also have initial trees for marked uses of and and or that have a specific meaning beyond simple conjunction or disjunction, as in (61) a. Throw another spit ball and you'll regret it.</Paragraph> <Paragraph position="14"> b. Eat your spinach or you won't get dessert.</Paragraph> <Paragraph position="15"> These differ from the more frequent, simple coordinate uses of and and or in that the second conjunct in these marked cases bears a discourse relation to the first conjunct (result in both (61a) and (61b)). With simple coordinate uses of and and or, all conjuncts (disjuncts) bear the same relation to the same immediately left-adjacent discourse unit. For example, in (62), each conjunct is a separate explanation for not trusting John, wheras in (63), each disjunct conveys an alternative result of John's good fortune: (62) You shouldn't trust John. He never returns what he borrows, and he bad-mouths his associates behind their backs.</Paragraph> <Paragraph position="16"> (63) John just won the lottery. So he will quit his job, or he will at least stop working overtime. For simple coordinate uses of and and or, we have auxiliary trees (section 5.1.2). Auxiliary trees for basic elaboration. These particular trees are anchored by (a) the punctuation mark &quot;period&quot; and (b) and. The symbol [?] indicates the foot node of the auxiliary tree, which has the same label as its root. (c) Auxiliary tree for the discourse adverbial then. TAG derivation of example (64).</Paragraph> <Paragraph position="17"> First, auxiliary trees anchored by punctuation (e.g., period, comma, semicolon.) (Figure 10a) or by simple coordination (Figure 10b) are used to provide further description of a situation or of one or more entities (objects, events, situations, states, etc.) within the situation.</Paragraph> <Paragraph position="18"> The additional information is conveyed by the discourse clause that fills its substitution site. Such auxiliary trees are used in the derivation of simple discourses such as (64) a. John went to the zoo.</Paragraph> <Paragraph position="19"> b. He took his cell phone with him.</Paragraph> <Paragraph position="20"> Figure 11 shows the DLTAG derivation of example (64), starting from LTAG derivations of the individual sentences.</Paragraph> <Paragraph position="21"> To the left of the horizontal arrow are the elementary trees to be combined: T1 stands for the LTAG tree for clause (64a), T2 for clause 18 The latter use of an auxiliary tree is related to dominant topic chaining in Scha and Polanyi (1988) and entity chains in Knott et al. (2001).</Paragraph> <Paragraph position="22"> 19 We comment on left-to-right incremental construction of DLTAG structures in parallel with sentence-level LTAG structures at the end of Section 5.2.</Paragraph> <Paragraph position="23"> Webber et al. Anaphora and Discourse Structure (64b), and b:punct1 for the auxiliary tree assocated with the period after (64a). In the derivation, the foot node of b:punct1 is adjoined to the root of T1 and its substitution site filled by T2, resulting in the tree to the right of the horizontal arrow. (A standard way of indicating TAG derivations is shown under the horizontal arrow, where broken lines indicate adjunction and solid lines, substitution. Each line is labeled with the address of the argument at which the operation occurs. t1 is the derivation tree for T1 and t2, the derivation tree for T2.) The other auxiliary trees used in the lexicalized discourse grammar are those for discourse adverbials, which are simply auxiliary trees in a sentence-level LTAG (XTAG-Group 2001), but with an interpretation that projects up to the discourse level. An example is shown in Figure 10c. Adjoining such an adverbial to a clausal/sentential structure contributes to how information conveyed by that structure relates to the previous discourse.</Paragraph> <Paragraph position="24"> There is some lexical ambiguity in this grammar, but no more than serious consideration of adverbials and conjunctions demands. First, as already noted, discourse adverbials have other uses that may not be anaphoric (65a-b) and may not be clausal (65a-c): (65) a. John ate an apple instead of a pear.</Paragraph> <Paragraph position="25"> b. In contrast with Sue, Fred was tired.</Paragraph> <Paragraph position="26"> c. Mary was otherwise occupied.</Paragraph> <Paragraph position="27"> Second, many of the adverbials found in second position in parallel constructions (e.g., on the other hand, at the same time, nevertheless) can also serve as simple adverbial discourse connectives on their own. In the first case, they will be one of the two anchors of an initial tree (Figure 8), and in the second, they will anchor a simple auxiliary tree (Figure 10c). These lexical ambiguities correlate with structural ambiguity.</Paragraph> </Section> <Section position="7" start_page="577" end_page="583" type="sub_section"> <SectionTitle> 5.2 Example Derivations </SectionTitle> <Paragraph position="0"> It should be clear by now that our approach aims to explain discourse semantics in terms of a product of the same three interpretive mechanisms that operate within clause-level semantics: * compositional rules on syntactic structure (here, discourse structure) * anaphor resolution * inference triggered by adjacency and structural connection For the compositional part of semantics in DLTAG (in particular, computing interpretations on derivation trees), we follow Joshi and Vijay-Shanker (2001). Roughly, they compute interpretations on the derivation tree using a bottom-up procedure. At each level, function application is used to assemble the interpretation of the tree from the interpretation of its root node and its subtrees. Where multiple subtrees have function types, the interpretation procedure is potentially nondeterministic: The resulting ambiguities in interpretation may be admitted as genuine, or they may be eliminated by a lexical specification. Multicomponent TAG tree sets are used to provide an appropriate compositional treatment for quantifiers, which we borrow for interpreting for example (examples (66c-d)).</Paragraph> <Paragraph position="1"> In showing how DLTAG and an interpretative process on its derivations operate, we must, of necessity, gloss over how inference triggered by adjacency or associated with a structural connective provides the intended relation between adjacent discourse Computational Linguistics Volume 29, Number 4 units: It may be a matter simply of statistical inference, as in Marcu and Echihabi (2002), or of more complex inference, as in Hobbs et al. (1993). As we noted, our view is that there are three mechanisms at work in discourse semantics, just as there are in clause-level semantics: Inference isn't the only process involved. Thus the focus of our presentation here is on how compositional rules and anaphor resolution (which itself often appears to require inference) operate together with inference to yield discourse semantics.</Paragraph> <Paragraph position="2"> We start with previous examples (44) (here (66c)) and (47) (here (66d)) and two somewhat simpler variants (66a-b): (66) a. You shouldn't trust John because he never returns what he borrows. b. You shouldn't trust John. He never returns what he borrows.</Paragraph> <Paragraph position="3"> c. You shouldn't trust John because, for example, he never returns what he borrows.</Paragraph> <Paragraph position="4"> d. You shouldn't trust John. For example, he never returns what he bor- null rows.</Paragraph> <Paragraph position="5"> This allows us to show how (66a-b) and (66c-d) receive similar interpretations, despite having somewhat different derivations, and how the discourse adverbial for example contributes both syntactically and semantically to those interpretations. We let T1 stand for the LTAG parse tree for you shouldn't trust John, t1, for its derivation tree, and interp(T1), for the eventuality associated with its interpretation. Similarly, we let T2 stand for the LTAG parse tree for he never returns what he borrows, t2, for its derivation tree, and interp(T2), for the eventuality associated with its interpretation.</Paragraph> <Paragraph position="6"> Example (66a) involves an initial tree (a:because-mid) anchored by because (Figure 12). Its derived tree comes from T1 substituting at the left-hand substitution site of a:because-mid (index 1) and T2 at its right-hand substitution site (index 3). Compositional interpretation of the resulting derivation tree yields explanation(interp (T2),interp(T1)). (A more precise interpretation would distinguish between the direct and epistemic causality senses of because, but the derivation would proceed in the same way.) In contrast with (66a), example (66b) employs an auxiliary tree (b:punct1) anchored by a period (Figure 13). Its derived tree comes from T2 substituting at the right-hand substitution site (index 3) of b:punct1, and b:punct1 adjoining at the root of T1 (index 0). Compositional interpretation of the derivation tree yields merely that T2 continues the description of the situation associated with T1, that is, elaboration(interp(T2),interp(T1)). Further inference triggered by adjacency and structural connection leads to a cona:because_mid null a:because_mid Derivation of example (66a). The derivation tree is shown below the arrow, and the derived tree, to its right. (Node labels D c have been omitted for simplicity.) Webber et al. Anaphora and Discourse Structure Derivation of example (66c).</Paragraph> <Paragraph position="7"> clusion of causality between them, that is, explanation(interp(T2),interp(T1)), but this conclusion is defeasible because it can be denied without a contradiction: for example, (67) You shouldn't trust John. He never returns what he borrows. But that's not why you shouldn't trust him.</Paragraph> <Paragraph position="8"> Example (66c) differs from (66a) in containing for example in its second clause. As noted earlier, for example resembles a quantifier with respect to its semantics, as its interpretation takes wider scope than would be explained by its syntactic position. We handle this in the same way that quantifiers are handled in Joshi and Vijay-Shanker (2001) by associating with for example a two-element TAG tree set (Figure 14). Both trees in the tree set participate in the derivation: The auxiliary tree b:for ex1 adjoins at the root of T2, whereas the auxiliary tree b:for ex2 adjoins at the root of the higher discourse unit. Since we saw from example (66a) that the interpretation of this higher discourse unit is explanation(interp(T2),interp(T1)), the interpretation associated with the adjoined b:for ex2 node both embeds and abstracts this interpretation, yielding exemplification(interp(T2), lX.explanation(X,interp(T1)) That is, John's never returning what he borrows is one instance of a set of explanations. Similarly, example (66d) differs from (66b) in containing for example in its second sentence (Figure 15). As in example (66b), an inferred relation is triggered between the interpretations of T2 and T1, namely, explanation(interp(T2),interp(T1)). Then, as a result of b:for ex1 adjoining at T2 and b:for ex2 adjoining at the root of the higher Derivation of example (66d).</Paragraph> <Paragraph position="9"> discourse unit, for example again contributes the interpretation exemplification(interp(T2), lX.explanation(X,interp(T1)) Thus (66c) and (66d) differ only in the derivation of the interpretation that for example then abstracts over.</Paragraph> <Paragraph position="10"> The next example we will walk through is example (11) (repeated here as example (68)): (68) John loves Barolo. So he ordered three cases of the '97. But he had to cancel the order because then he discovered he was broke.</Paragraph> <Paragraph position="11"> As shown in Figure 16, this example involves two initial trees (a:so, a:because mid) for the structural connectives so and because; an auxiliary tree for the structural connective but (b:but), since but functions as a simple conjunction to continue the description of the situation under discussion; an auxiliary tree (b:then) for the discourse adverbial then; and initial trees for the four individual clauses T1-T4. As can be seen from the derivation tree, T1 and T2 substitute into a:so as its first and third arguments, and b:but root-adjoins to the result. The substitution argument of b:but is filled by a:because mid, with T3 and T4 substituted in as its first and third arguments, and b:then is rootadjoined to T4. The interpretation contributed by then, after its anaphoric argument is resolved to interp(T2), is i4: after(interp(T4), interp(T2)) The interpretations derived compositionally from the structural connectives so, because, and but are i1: result(interp(T2), interp(T1)) i2: explanation(interp(T4), interp(S3)) i3: elaboration(i2,i1) Further inference may then refine elaboration to contrast, based on how but is being used.</Paragraph> <Paragraph position="12"> Finally, we want to point out one more way in which texts that seem to be close paraphrases get their interpretations in different ways. Consider the two texts in example (69): Webber et al. Anaphora and Discourse Structure Derivation of example (68).</Paragraph> <Paragraph position="13"> (69) a. You should eliminate part 2 before part 3 because part 2 is more susceptible to damage.</Paragraph> <Paragraph position="14"> b. You should eliminate part 2 before part 3. This is because part 2 is more susceptible to damage.</Paragraph> <Paragraph position="15"> Example (69b) is a simpler version of an example in Moser and Moore (1995), in which This is because is treated as an unanalyzed cue phrase, no different from because in (69a). We show here that this isn't necessary: One can analyze (69b) using compositional semantics and anaphor resolution and achieve the same results.</Paragraph> <Paragraph position="16"> First consider (69a). Given the interpretations of its two component clauses, its overall interpretation follows in the same way as (66a), shown in Figure 12. Now consider (69b) and the derivation shown in Figure 17. Here the initial tree a:because mid Computational Linguistics Volume 29, Number 4 has its two arguments filled by T2, the TAG analysis of this is and TB, the TAG analysis of part 2 is more susceptible to damage. The overall derived tree for (69b) comes from b:punct1 root-adjoining to T1 (the TAG analysis of You should eliminate part 2 before part 3), with the subsitution site of b:punct1 filled by the a:because mid derivation. The compositional interpretation of the derivation tree yields the interpretation of the a:because mid tree (i1) as an elaboration of the interpretation of T1: i1: explanation(interp(TB),interp(T2)) i2: elaboration(i1,interp(T1)) But this is not all. The pronoun this in T2 is resolved anaphorically to the nearest consistent eventuality (Eckert and Strube 2000; Byron 2002) which in this case is interp(T1). Taking this as the interpretation of T2 and substituting, we get i1: explanation(interp(TB),interp(T1)) i2: elaboration(i1,interp(T1)) Notice that i1 is also the interpretation of (69a). To this, i2 adds the somewhat redundant information that i1 serves to elaborate the advice in T1. Thus (69a) and (69b) receive similar interpretations but by different means. This treatment has the added advantage that one does not have to treat This is not because as a separate cue phrase. Rather, negation simply produces i1: !explanation(interp(TB),interp(T1)) i2: elaboration(i1,interp(T1)) That is, T1 is elaborated by a denial of a (possible) explanation. Presumably, the text would go on to provide the actual explanation.</Paragraph> <Paragraph position="17"> Finally, we want to comment on the Holy Grail of discourse parsing: a realistic model that is computed in parallel with incremental sentence-level parsing. Neither the analyses given in this section nor the discourse parsing described in Forbes et al. (2001) is done in a left-to-right incremental fashion, in parallel with incremental left-to-right sentence-level parsing.</Paragraph> <Paragraph position="18"> What would an integrated incremental method of sentence-discourse processing require? At minimum, we believe it would involve: * A left-to-right parser that would simultaneously compute increments to sentence-level syntactic structure, sentence-level semantics, discourse-level syntactic structure, and discourse-level semantics.</Paragraph> <Paragraph position="19"> Increments to the latter two would occur only at clause boundaries and with discourse adverbials and structural connectives.</Paragraph> <Paragraph position="20"> * An incremental anaphor resolution mechanism, similar to that in Strube (1998), but extended both to deictic pronouns, as in Eckert and Strube (2000) and Byron (2002) and to the anaphoric argument of discourse adverbials.</Paragraph> <Paragraph position="21"> * Incremental computation of discourse structure in terms of elaboration relations and further nondefeasible reasoning to more specific relations, where possible.</Paragraph> <Paragraph position="22"> A left-to-right parser that simultaneously produces sentence-level syntactic and semantic analyses already exists for combinatory categorial grammar (Steedman 1996, Webber et al. Anaphora and Discourse Structure 2000b; Hockenmaier, Bierner, and Baldridge, forthcoming), and it would seem straight-forward to extend such a parser to computing discourse-level syntax and semantics as well. Similarly, it seems straightforward to produce an incremental version of any of the current generation of anaphor resolution mechanisms, extended to deictic pronouns, although current approaches attempt to resolve this and that only with the interpretation of a single clause, not with that of any larger discourse unit. As these approaches are also not very accurate as yet, incremental anaphor resolution awaits improvements to anaphor resolution in general. Moreover, as we better understand the specific anaphoric properties of discourse adverbials through empirical analysis such as Creswell et al. (2000), such anaphor resolution mechanisms can be extended to include them as well.</Paragraph> <Paragraph position="23"> As for building discourse structure incrementally in parallel with syntactic structure, there is no working prototype yet that will do what is needed. But we have no doubt that as psycholinguistics and computation together develop a better understanding of incremental semantic processing, researchers' desire for a working prototype will eventually result in the development of one.</Paragraph> </Section> </Section> class="xml-element"></Paper>