Dialogue & Discourse 7(4) (2016) 1-35 doi: 10.5087/dad.2016.401 Reports in Discourse Julie Hunter juliehunter@gmail.com IRIT, Université Paul Sabatier Toulouse, France & GLiF, Universitat Pompeu Fabra Barcelona, Spain Editor: Jonathan Ginzburg Submitted 08/2014; Accepted 06/2016; Published online 06/2016 Abstract Attitude or speech reports in English with a non-parenthetical syntax sometimes give rise to inter- pretations in which the embedded clause, e.g., John was out of town in the report Jill said that John was out of town, seems to convey the main point of the utterance while the attribution predicate, e.g., Jill said that, merely plays an evidential or source-providing role (Urmson, 1952). Simons (2007) posits that parenthetical readings arise from the interaction between the report and the preceding discourse context, rather than from the syntax or semantics of the reports involved. To my knowl- edge, however, no account of these discourse interactions has been developed in formal semantics. Research on parenthetical reports within frameworks of rhetorical structure has yielded hypotheses about the discourse interactions of parenthetical reports, but these hypotheses are not semantically sound. The goal of this paper is to unify and extend work in semantics and discourse structure to develop a formal, discourse-based account of parenthetical reports that does not suffer the pitfalls faced by current proposals in rhetorical frameworks. 1 Keywords: Speech reports, Parenthetical reports, Evidential reports, Discourse structure, Dis- course Connectives 1. Introduction In the absence of further context, a speaker2 who utters (1) would naturally be understood as using (1b) to propose an explanation for why John did not come to her party. (1) a. John didn’t come to my party. b. Jill said he was out of town. The content of Jill said in (1b) does not participate, at least not directly, in this explanation—while it is possible to imagine a scenario in which John didn’t come to the party because Jill said he was out of town, this reading of (1b) would require a more elaborate discourse than (1). It is rather the 1. I would like to thank Márta Abrusán, Pascal Amsili, Nicholas Asher, Laurence Danlos, Eric Kow, Mandy Simons, participants of the CREST International Workshop on Formal and Computational Semantics at Kyoto, and participants at the Nijmegen workshop Backgrounded Reports, including organizers Corien Bary and Emar Maier, for helpful discussions on the issues discussed in this paper. I also thank my three anonymous Dialogue & Discourse reviewers for their thorough and constructive comments. This work has been supported by the French agency Agence Nationale de la Recherche (ANR-12-CORD-0004) and the European Research Council (grant 269427). 2. Throughout the paper, I will use speaker to refer to the agent of either a spoken or written utterance/sentence token. ©2016 Julie Hunter This is an open-access article distributed under the terms of a Creative Commons Attribution License (http ://creativecommons.org/licenses/by/3.0/). Hunter embedded clause, he was out of town that is understood as a (possible) explanation of John’s absence. Informally, we can say that the embedded clause seems to contribute the main point of (1b) while the attribution predicate, Jill said, merely serves to provide a source or evidence for the embedded content. Following Hooper (1975), Simons (2007) and Urmson (1952), I will call the use of say in (1b) a parenthetical use and I will call a report in which the embedding verb is used parenthetically a parenthetical report. Now compare (1) with (2), which involves a standard, non-parenthetical use of the same report. (2) a. John is mad at Jill. b. Jill said he was out of town, c. so I didn’t invite him to my party. The report in (2b) is identical to that in (1b), yet the report does not make the same contribution to (2) as it does to (1). In (2b), the attribution predicate is integral to the explanation of (2a): the speaker is not suggesting that John is mad at Jill because he was out of town, but that he is mad at Jill because she said something (which led to his being excluded from the speaker’s party). Thus, while the embedded clause alone seems to make the main discourse contribution of (1b), the attribution predicate of (2b), or perhaps the report as a whole, crucially participates in the report’s main contribution to (2). The discursive difference between (1b) and (2b) brings with it a difference in semantic entail- ments. To the extent that a speaker of (1) is committed to the possibility that John didn’t come to the party because he was out of town, she must also be committed to the possibility that John was out of town (at the relevant time). That is, she must be committed to at least the possibility that the embedded content of (1b) is true. There is no such requirement of commitment to the embedded content of (2b): (2b) can be used to explain (2a) even in a context in which it is common knowledge that John was not out of town. If we assume that parenthetical reports like (1b) have the same syntactic structure as their non- parenthetical counterparts—which I, following Simons (2007), will—then we must conclude that the discursive and semantic differences between (1b) and (2b) do not arise from the syntactic and semantic features of the report alone, but rather from the interaction between these features and other discourse moves. Accordingly, I will introduce the more specific term discourse parenthetical reports to talk about parenthetical reports akin to (1b). This will help to distinguish these reports from those whose parenthetical readings are marked syntactically (see §5 for further discussion of syntactic parentheticals).3 The goal of this paper is to develop a formal, discourse-based model of discourse parenthetical reports. While parenthetical reports have been discussed at length in Hooper (1975), Rooryck (2001), Simons (2007), and Urmson (1952) and it has been noted that parenthetical readings arise from the discourse function of parenthetical reports, there is as of yet no formal model of their discourse function or of how this function affects the semantic entailments of reports in different discourse contexts. Work in formal semantics has tended to bring observations about parenthetical readings back to bear on outstanding problems in formal semantics. Urmson, for example, was concerned with the implications of parenthetical readings for the semantics of attitude reports. Simons uses parenthetical reports to argue that the presuppositional behavior of factive verbs is not determined by their lexical semantics: given the right discourse context, even the content in the scope of a factive verb can carry the main point of the report—i.e., the report can have a parenthetical reading—and so the 3. I would like to thank an anonymous reviewer for suggesting the term discourse parenthetical. 2 Reports in Discourse embedded content can fail to be presupposed. This paper takes for granted the syntax and semantics of reports, and in particular, the assumption that reports with the surface structure of (1b)/(2b) have the same semantics and syntax regardless of whether they are interpreted parenthetically or not. The aim is rather to formalize the notion of discourse function at work in parenthetical readings and to explain how it accounts for the different entailments that arise from parenthetical and non- parenthetical readings. Simons provides the most developed discussion of the discourse function of parenthetical reports that I am aware of in the formal semantics and pragmatics literature, but as her focus is on the behavior of different embedding verbs, she limits her discussion to how reports behave in question/answer sequences: (3) A: Why didn’t John come to my party? B: Jill said (thinks, suspects, imagines, supposes, heard...) that he’s out of town. In these sequences, Simons proposes that the discourse function of the embedded clause of a discourse parenthetical report, e.g. (3B), is to answer the question posed by the preceding move, e.g. (3A). She says: “whatever proposition communicated by the response constitutes an answer (complete or partial) to the question is the main point of the response” (p. 1036). As Simons explicitly acknowl- edges (p. 1035), however, this criterion provides only a limited picture of the discourse function of parenthetical reports. Theories of rhetorical structure such as Rhetorical Structure Theory (RST; Mann and Thompson (1988)) and Segmented Discourse Representation Theory (SDRT; Asher and Lascarides (2003)) have more developed notions of discourse structure and discourse function. The discourse function of an utterance u is given by the discourse relation that connects the content c of u to the surrounding discourse: if c is related to some other content c′ via an Explanation relation, then c’s discourse function is to explain the eventuality described by c′; if c is related to c′ via a Narration relation, then its discourse function is to push the narrative forward by describing the next (discourse relevant) eventuality that occurs after that described by c′, and so on. The structure of the discourse is then determined by the collection of relation instances in the discourse. This notion of discourse function has been applied to discourse parenthetical reports in multiple efforts to annotate newspaper texts for rhetorical structure (Dinesh et al., 2005; Hardt, 2013; Hunter et al., 2006). The parenthetical/non-parenthetical distinction generally becomes relevant when an- notators, who are in many cases untrained in linguistics, decide that only the embedded content of a report is relevant to the discourse relation that connects the report to the preceding discourse. (4), an attested example with the form of (1), provides an illustration. (4) London serves increasingly as a conduit for program trading of U.S. stocks. Market professionals said London has several attractions. First, the trading is done over the counter ... Second, it can be used to unwind positions before U.S. trading begins... (PDTB, File 0097). (4) is originally from The Wall Street Journal, but figures in the corpus for the Penn Discourse Tree Bank (PDTB; Prasad et al. (2007b)). In this example, PDTB annotators chose the content of the boldface text as the proposed explanation for the claim expressed by the sentence in italics; the attribution predicate of the report, Market professionals said, was excluded from the explanans. In other words, annotators took the report in (4) to be discourse parenthetical. 3 Hunter While discourse parenthetical reports might not be discussed as such in these annotation efforts, the annotation schemes that they have developed in response to discourse parenthetical readings in effect yield the following definition: a report r is discourse parenthetical just in case it is the embedded clause of r, not the attribution predicate, that enters into a discourse relation with some element from the discourse preceding r.4 Simons’ proposed constraint for question/answer pairs can be seen as a particular case of this more general notion: a report r is discourse parenthetical just in case the embedded clause of r alone provides the second argument for an instance of the relation Question/Answer Pair, where q provides the first argument, for some question q in the incoming discourse.5 Of course, Simons’ constraint could be generalized in different ways. An alternative approach might involve defining discourse function and structure within a Question Under Discussion account (Ginzburg, 2012; Roberts, 2012; Simons et al., 2010), though I will not develop such an approach here. The model that I develop builds on the more general notion of discourse function offered by rhetorical theories, but rectifies certain semantic problems inherent in the annotation-driven solutions that have been offered. In particular, as I explain in §3, extant proposals do not account for the fact, well-known from semantic discussions of parenthetical reports, that a speaker who utters a discourse parenthetical report need not in general be fully committed to the embedded content of that report. Nor do they account for the fact that a speaker must nevertheless be at least tentatively committed to the embedded content. Finally, these accounts prevent the attribution predicate and the embedded clause of a report from being simultaneously relevant to the discourse outside of the report, at least when they play distinct rhetorical functions. These accounts in effect require annotators to treat either one or the other clause as the “main point” of the report, but that requirement is too strong given the data on reports in discourse. In a nutshell, my proposal is that a discourse parenthetical report is one in which the embedded clause is rhetorically connected to the discourse preceding the report, as in extant accounts, but unlike extant accounts, the embedded clause is discursively and semantically subordinate to the attribution predicate. As a result of these rhetorical connections, discourse parenthetical reports introduce modal discourse relations between the embedded clause and the preceding discourse, and the attribution predicate can be rhetorically related to the preceding discourse, although it must be related by a different discourse relation. It should be noted that the account that I offer is designed to model the contribution of discourse parenthetical reports to discourse structure, but is not intended as a complete semantic account of parenthetical reports. It is meant to complement, rather than replace, studies on other factors that affect the interpretation of speech or attitude reports such as the lexical semantics of embedding verbs, e.g. Simons (2007), or the influence of certain kinds of world knowledge on our judgements about the reliability of a discourse parenthetical report, e.g. de Marneffe et al. (2012). It should also be a useful supplement to annotation-based work on rhetorical structure. Determining the impact of speech and attitude reports on discourse structure and interpretation is of utmost importance. For tasks such as automated discourse parsing, text summarization, and the automatic recognition of textual entailment, for example, we need to be able to draw reliable inferences from a discourse as 4. Constraints on how a bit of content can attach to the incoming discourse context are given by independent principles that vary from theory to theory. All that is important here is that the embedded clause is being attached to a representation of the discourse prior to the report, rather than to its own attribution predicate. 5. Different frameworks might have different names for this relation; I’ve opted here for the terminology of SDRT, but that is not important. However, given the semantics of Question/Answer Pair (QAP) in SDRT (Asher and Lascarides, 2003), it is important to note that this criterion will be equivalent to Simons’ because for a discourse unit to serve as the second argument to an instance of QAP, it must provide a complete or partial answer to the first argument. 4 Reports in Discourse a whole. An important piece in this puzzle is figuring out how speakers make use of other peoples’ speech acts and attitudes to perform their own speech acts and convey their own attitudes. I begin in §2 by providing more detail on rhetorical theories and reviewing three proposals for the annotation of discourse parenthetical reports within different rhetorical frameworks. While these accounts differ in various ways, they all share the core idea that discourse parenthetical reports should be modelled by attaching the embedded clause to the preceding discourse. In §3, I show that these accounts as they stand come into conflict with semantic facts about discourse parenthetical reports and with independent principles of rhetorical structure. In §4 I show how the notion of discourse function adopted in rhetorical theories can be developed into a more more general, consistent account of discourse parenthetical reports. §5 takes a look at how syntactic parentheticals fit into the picture developed in §4. §6 concludes the discussion. 2. Rhetorical theories Extant rhetorical theories, including RST, SDRT, D-LTAG (Webber, 2004), and other annotation methods, including those for the PDTB and the Copenhagen Discourse Tree Bank (Buch-Kromann and Korzen, 2010), differ in critical ways, and in the more theoretical discussion below, we will need to focus on a single theory. Nevertheless, the import of the current study should be of interest for research on rhetorical structure more generally. While the three accounts that I describe below are worked out using very different annotation methodologies, the heart of the proposals is the same, and as such, parts of the discussion to follow will be relevant to all of them. Where we will need a particular theory is in working out the details of my concerns and developing a positive proposal. Building a representation of the rhetorical structure of a given discourse requires performing three tasks: segmenting the discourse into minimal units, attaching each discourse unit to some other unit in the structure, and labelling each discourse attachment with a rhetorical relation. When it comes to the annotation of discourse parenthetical reports, the approaches outlined in the PDTB (Dinesh et al., 2005), the Copenhagen Dependency Treebank (CDT) (Buch-Kromann et al., 2011), and SDRT (Hunter et al., 2006; Reese et al., 2007) all follow a similar recipe for accomplishing these tasks. First, a report r that is intuitively discourse parenthetical is segmented such that the embedded clause contributes its own discourse unit u. Second, u is attached to another unit u′ in the discourse structure s built from the set of utterances prior to that of r. Third, the relation between u′ and u is labelled with the discourse relation that intuitively would have related u′ and u had u not been the argument of an attribution predicate. Finally, if desired, the discourse function of the attribution predicate can be modelled by attaching it to u with a special discourse relation that indicates its evidential (or emotive, etc.) discourse function. The PDTB, for example, would indicate the discourse contribution of (1) as follows: (1) a. John didn’t come to my party. b. Implicit = because Jill said he was out of town. Following the conventions of the PDTB manual (Prasad et al., 2007b), content that contributes to the first argument of a discourse connective is placed in italics, while content that contributes to the second argument is placed in boldface. The connective, when explicit, is underlined; implicit connectives are marked as in (1b). We can see from the representation of (1) (which echoes the actual PDTB annotation of the found example (4)) that the embedded clause forms its own segment and it alone serves as the argument that ties the report to the discourse context prior to (1b) (Prasad et al., 5 Hunter 2007a).6 This segmentation choice contrasts with the segmentation for non-parenthetical reports, in which the entire report, attribution predicate plus embedded clause, is treated as a single discourse segment. The PDTB does not take a stand on the discourse contribution of the attribution predicate in discourse parenthetical reports; information about the sources and spans of attributions is stored alongside the annotation of a discourse in the PDTB but there is no connective posited to capture the discourse function of this information. In the CDT, the annotation of syntactic structure and discourse structure is done using a single dependency graph so that there is a high level of uniformity between discourse and syntactic struc- ture (Buch-Kromann and Korzen, 2010). One of the rare violations of this uniformity is allowed for discourse parenthetical reports, in which only the embedded clause is treated as a discourse ar- gument. Instances of relations involving discourse parenthetical reports (and other syntax/discourse mismatches) are marked with a ‘*’: an asterisk to the left of a discourse connective signals that the discourse/syntax mismatch can be found in the first argument of the relation; an asterisk to the right signals that the mismatch lies in the second argument (Hardt, 2013). (1), for example, would be annotated along the following lines: Explanation∗(1a, 1b). Like the PDTB, the CDT does not take a stand on the discourse contribution of the attribution predicate in a discourse parenthetical report. A third proposal was offered by Hunter et al. (2006) for SDRT as part of the project DiSCoR7. Hunter et al., in contrast to the PDTB group, always treat the attribution predicate and embedded clause of a report as separate segments, regardless of whether the report has a discourse parenthetical reading or a non-parenthetical one. The different readings of reports are distinguished by the fact that the embedded clause of a discourse parenthetical report is the clause that is attached to the incoming discourse structure, as for the PDTB and CDT. In addition, discourse parenthetical and non-parenthetical reports are distinguished by different discourse relations that connect the attribution predicate of the report to the embedded clause: Attribution is used for non-parenthetical reports and Source, for parenthetical reports.8 In Attribution, the embedded clause is discourse subordinate (see Asher and Lascarides (2003)) to the attribution predicate, echoing the syntactic structure of the report. Source, however, reverses the arguments of Attribution so that the attribution predicate is discourse subordinate to the syntactically embedded clause. Furthermore, Source and Attribution differ in their entailments: when two arguments are related by Source, the text entails the content of both arguments, whereas when two contents are related by Attribution, the text only entails that the agent of the attribution stand in the relation indicated by the embedding verb to the content of 6. The PDTB group argue that the relation of attribution is one that holds between an individual and an abstract object. As discourse relations are taken to relate abstract objects only, Attribution is not treated as a discourse relation. So the difference between discourse parenthetical reports and non-parenthetical reports is that in the latter, the whole report is treated as a single unit that can figure in an argument for a discourse relation whereas in the former, only the embedded content contributes to an argument. In their words: “a discourse relation may hold either between the attributions (and the agents of attributions) themselves or only between the abstract object arguments of the attribution...” (p. 20) 7. DiSCoR, Discourse Structure and Co-reference Resolution, was an NSF projet designed to study the relation between co-reference resolution and discourse structure. Most of the texts annotated for discourse structure were texts from the Message Understanding Conference (MUC) 6, that were already annotated for co-reference. 8. The relation Source was originally introduced by Hunter et al. (2006) under the name Evidence. The name was changed so that Evidence could be used for a different evidential relation. I have chosen to use the name Source here because it is consistent with later SDRT annotations, e.g. Reese et al. (2007), that incorporated Hunter et al.’s approach. 6 Reports in Discourse the embedded clause.9 An example of Source is provided in (1H), following the segmentation of (1) below: (1) a. [John didn’t come to my party.]α b. [Jill said]β [he was out of town.]γ (1H) Explanation(α, γ), Source(γ, β) Two further accounts of speech reports that deserve mention here, though I will not discuss them in the rest of the paper, are those of Carlson and Marcu (2001) and Redeker and Egg (2006), both developed in Rhetorical Structure Theory (Mann and Thompson, 1988). Carlson and Marcu suggest that all reports be annotated with a relation that they call Attribution, which is structurally similar to Hunter et al.’s Source in that the attribution predicate in a report is treated as the satellite and the reported content serves as the nucleus. Redeker and Egg (2006) criticizes Carlson and Marcu’s approach and offers an account that reverses the arguments of Attribution so that the reported content is the satellite and the attribution predicate, the nucleus. The reason why I will not pursue either of these accounts in this paper is that neither makes a distinction between discourse parenthetical reports and non-parenthetical reports.10 The attribution predicate is simply treated as the satellite of Attribution in Carlson and Marcu (2001) and as the nucleus in Redeker and Egg (2006). What we’re interested in for the purposes of this paper is accounts that advocate a solution specifically for discourse parenthetical readings of reports. The treatment of discourse parenthetical reports in the PDTB, the CDT, and Hunter et al. (2006) all have in common the idea that discourse parenthetical reports are best modelled by attaching the embedded clause directly to the incoming discourse and that this attachment pattern distinguishes them from non-parenthetical reports, in which it is the attribution predicate that is attached to the incoming discourse. Accordingly, I classify these accounts as attachment solutions. However, the claim that discourse parenthetical reports show different attachment patterns in discourse does not entail that discourse parenthetical reports and non-parenthetical reports should be distinguished syntactically, and I will take it for granted that the difference is not a syntactic one—except, of course, when the parenthetical verb appears in a syntactic parenthetical as in Mary will be late, John said (see §5 for a discussion of syntactic parentheticals). I will not defend this position here, because my focus is on the discourse contribution of parenthetical reports, regardless of their syntactic structure.11 The point is that claims about the discourse structure of a chunk of discourse do not automatically entail claims about the syntactic structure of the constituents in that chunk. In fact, of the frameworks introduced here, the one that assumes the closest tie between syntactic structure and discourse structure is that for the CDT, and even the CDT treats discourse parenthetical reports as involving a syntax/discourse mismatch. In other words, they assume that the discourse contribution of a parenthetical report does not mirror its syntactic form. 9. For Attribution(α, β) or Source(β, α), the content of α will entail: x said (thought,...) p for some agent x, and β will specify the content of p in the sense that p will denote a subset of the worlds in the proposition denoted by Kβ , i.e. the DRS that results from the processing of β. 10. However, see Matthiessen and Thompson (1988) for an interesting discussion about the fact that whether a given clause is ultimately considered to be a main clause or a subordinate clause depends on features of the discourse in which that clause is used. 11. See Simons (2007) for arguments that reports in which an embedding verb appears in the main clause, as in John said (that) Mary will be late, have the same syntactic structure regardless of whether the embedding verb is used parenthetically or not. 7 Hunter 3. Conflicting criteria In the ensuing discussion, I will adopt SDRT as my theoretical framework. To fully model the rhetorical contribution of reports, we need to take a stand on the discourse function of both the embedded clause and the attribution predicate, and we need a theory that offers independent principles to guide our theoretical choices on this matter. SDRT satisfies this criterion, but not all rhetorical frameworks do. The PDTB annotation method, for instance, is designed to be theory-neutral, and so cannot provide a theoretical framework by design. Moreover, part of its theory-neutral approach is to annotate isolated pairs of discourse arguments and the connectives that relate them; there is no goal to describe the discourse contribution of every discourse unit. As a result, the PDTB can remain agnostic about the role of the attribution predicate in discourse parenthetical reports. SDRT is also largely motivated by semantic and pragmatic concerns and is the only rhetorical theory to provide a semantics for its discourse relations and to deliver fully interpretable logical forms for discourse. Many of the issues that we confront with discourse parenthetical reports are semantic and pragmatic, having to do with tracking the entailments of reports in a discourse or exploring aspects of their behavior that cannot be traced back to their syntactic structure. Thus a theory like SDRT is preferable to the framework of the CDT, which is wedded to a strong correspondence between syntactic structure and discourse structure, and even to RST, which does not emphasize the semantic interpretation of its discourse structures. I will also limit the range of report verbs in my study. There are many factors that influence the interpretation of reports aside from rhetorical structure: lexical semantics, world knowledge, perhaps focus, and so on. To study the interaction of rhetorical structure and reports, we need to minimize the influence of these other factors as much as possible. Most of my discussion will therefore be centered around embedding verbs like say and other speech report verbs that are non-factive and so do not indicate a particular level of author commitment to the content in their syntactic scope. Such speech report verbs are common in the corpora that I am pulling from12 and easily give rise to both discourse parenthetical and non-parenthetical readings. Also common are third person reports, and my discussion will focus largely on these as well, as it is with third person reports that questions about author commitment and entailments of reported content really become tricky. With these caveats in place, I turn now to an argument that the attachment-based solution outlined in section 2 leads to conflicts with semantic facts about discourse parenthetical reports and with independent principles of rhetorical theories. 3.1 Hedged commitments A speaker will often use a discourse parenthetical report to weaken or “hedge” her commitment to the embedded content of the report (Simons, 2007). In (1), for example, if the speaker were sure that John had been out of town, then the simplest solution would be to say so directly; the fact that she does not suggests that she is not entirely certain that he was out of town. Of course, we can imagine scenarios in which the use of a discourse parenthetical report is compatible with full speaker commitment to the embedded content, but what’s important is that a discourse parenthetical report does not itself require such commitment. This fact about discourse parenthetical reports comes into conflict with the constraint of veridi- cality imposed by many discourse relations. A relation R is veridical just in case the truth of an 12. The majority of the verbs in the corpus annotated for DisCoR were evidential verbs, e.g. say, as opposed to, for example, verbs indicating emotions, e.g. regret. 8 Reports in Discourse instance of R entails the truth of that instance’s arguments. Boolean conjunction, for example, is veridical: a formula of the form p ∧ q can only be true in a model M if both p and q are true in M . Conditional relations, by contrast, are not veridical: a formula of the form p→ q can be true even if both of its arguments are false. SDRT incorporates these basic logical connections into the seman- tics of its discourse relations—an arguably reasonable move for any semantic theory of rhetorical structure. The relation Continuation, for instance, conjoins two discourse units, and so each instance of Continuation will entail both of its arguments. Contrast also has conjunction as the foundation of its semantics, as does Narration; these relations are therefore veridical as well. Explanation is yet another veridical relation: a discourse unit u cannot truly explain another discourse unit u′ unless u′ and u both describe eventualities that held in the world of evaluation. To capture this dependence, a discourse formula of the form Explanation(u′, u) can only be true in a modelM in SDRT if u and u′ are true in M . When a speaker performs a speech act that presents a discourse unit u as standing in a relation R to another unit u′, the content of this speech act, R(u′, u), is added to the logical form for the discourse (for anyR, u, u′). IfR is veridical, then the speaker takes on a commitment to the content of bothu andu′, i.e. the content of her discourse entails bothu andu′.13 Note that this does not ensure that either u or u′ will be true in the relevant model—speakers are not infallible—it only ensures the speaker’s commitment to their truth. Here is where the conflict between veridicality and hedged commitments arises: attachment solutions attempt to model the discourse function of discourse parenthetical reports by attaching the embedded content of a report to the incoming discourse with the same relation that would have been used had the embedded content not been embedded. Where the relation is veridical, this entails author commitment to the embedded content—commitment that the speaker may not be ready to take on. What’s more, the relation at issue always will be veridical. If a discourse parenthetical report seems to provide an argument to a non-veridical relation, such as Alternation or Conditional, the result is that the attribution predicate will be understood as scoping over the discourse relation, as in (5): (5) If [John finishes his housework]α, then [Linda said]β [he’ll come to the party.]γ If Linda has reasons for thinking that John will come to the party that have nothing to do with him finishing his housework, then the choice of antecedent in (5) is unmotivated. The natural interpretation is therefore one in which Linda is committed to the conditional as a whole. Hunter et al. (2006) avoid the conflict between veridicality and hedged commitments by requiring that Source be veridical. This means that Source can only be used in cases in which annotators (interpreters) judge that the speaker is committed to the embedded content. Recall Hunter et al.’s analysis of (1) repeated here: (1) a. [John didn’t come to my party.]α b. [Jill said]β [he was out of town.]γ (1H) Explanation(α, γ), Source(γ, β) In fact, while I provided this example to illustrate the structural features of Source, this annotation would have only been allowed by Hunter et al. if the context provided reason to accept Jill’s report as totally reliable. 13. This picture of commitment follows original SDRT, but is complicated by issues such as embedded commitments and disagreements about commitments, topics which are handled in more recent work on SDRT. See Venant et al. (2014) and Venant and Asher (2016). 9 Hunter In the corpora used by the PDTB, the CDT, and DiSCoR, which consist mainly of newspaper articles, treating Source as a veridical relation often yields the right results. This is because in many of the articles, the content is so uncontentious and the sources so reliable that there is little reason to question the embedded content of the reports or the author’s commitment to it. (6b) is discussed by Dinesh et al. (2005) as an example of Contrast with a discourse parenthetical report (their example (12)): (6) a. At the same time, the New Brunswick, N.J., company said negotiations about pricing and volumes of product had collapsed between it and its exclusive distributor in the U.S., National Medical Care Inc. ... Yesterday, the [Delmed] spokeswoman said sales of Delmed products through the exclusive arrangement with National Medical accounted for 87% of Delmed’s 1988 sales of $21.1 million. b. The current distribution arrangement ends in March 1990, although Delmed said it will continue to provide some supplies of the peritoneal dialysis products to National Medical, the spokeswoman said. (6a) provides two of the three sentences that precede (6b) in the PDTB (file 0970). (6b) is presented as being factual: Delmed and the spokeswoman should be reliable sources given their relation to Delmed, there is nothing contentious about the content that they have reported (the content that Dinesh et al. mark as contributing to Argument 2 of although), and the author gives no signal in the rest of the article that s/he is not fully committed to this content. In examples like (6b), the take-home message seems unchanged if we remove information about the writer’s sources. (7) illustrates the same point, but with an Elaboration or Instance relation. While this example is not from the PDTB, I will use their conventions for marking the arguments for ease of exposition. (7) a. Another firm, uReveal, thinks that it’s cracked the code. Charles “Bucky” Clarkson, uReveal’s chairman and CEO, said that software such as his makes it easier to to parse all those government reports and organize the data so that analysts can get more out of it, and more quickly. He also claims that the software is so simple to use that (gasp!) even liberal arts majors can use it. Jokes about “soft majors” aside, the idea of data analysis tools easy for anyone to use is compelling because it frees up data scientists to do more specialized work. b. It can also bring in specialist knowledge from people who aren’t data scientists. Clarkson said, for example, that deploying these kinds of data analysis tools in hospitals have allowed doctors to spot trends that they would have otherwise missed, by analyzing their observation notes in conjunction with other electronic medical records. One doesn’t have to look far, however, even in the realm of newspaper articles, to see that discourse parenthetical reports are used widely in contexts in which speaker commitment to the embedded content is not ensured. A paradigm example is when a journalist uses discourse parenthetical reports to report the opinions of two parties who disagree with each other. For instance, in example (8), a toy variant of (13), which is discussed below, the speaker uses one report to express the point of view of the employees and another report to express the point of view of the boss, who directly disagrees with the employees. 10 Reports in Discourse (8) a. There was an explosion at the factory. b. The employees said that it was the fault of the boss c. but the boss said that it was the fault of the employees. We cannot infer speaker commitment to the embedded clause of either report by looking at (8) alone, but both reports are nevertheless used discourse parenthetically to offer possible explanations of the explosion. This is a problem for Hunter et al.’s account because the discourse parenthetical use of the reports calls for annotating the reports with Source, but the fact that the embedded content of the reports is not entailed precludes an annotation with Source. A further problem with Hunter et al.’s account is that even in cases such as (6b) and (7b), in which author commitment to the embedded content of a discourse parenthetical report can be inferred, commitment is almost always inferred by using world knowledge to reason about the reliability of the source(s) cited in the attribution predicate of the report. This kind of world-knowledge based reasoning, central to de Marneffe et al. (2012)’s study of veridicality, is generally independent of the reasoning used to determine rhetorical structure. In other words, Hunter et al.’s Source/Attribution distinction does not reflect a rhetorical distinction, and thus it fails to use the notion of discourse function from SDRT (or any rhetorical theory) to model the intuitive discourse function of discourse parenthetical reports. 3.2 No commitment Veridicality entails that the embedded clause of (1b) cannot be related to (1a) via Explanation, at least if there is any doubt about the speaker’s commitment to the truth of this clause’s content. This is intuitively correct: if the speaker is not entirely confident about the claim that John was out of town, she cannot be confident about the implicature14 that John’s being out of town explains why he wasn’t at the party. Still, in uttering (1b) she makes salient the possibility that John didn’t come to the party because he was out of town and performs something like an Explanation—a hedged Explanation, if you will. This intuition plays an important role in motivating attachment-based treatments of discourse parenthetical reports. However, some reports that seem in other ways to be discourse parenthetical carry no requirement of speaker commitment to the embedded clause and are not intuitively used to offer the content of the embedded clause as even a potential explanation or answer, etc. This happens when a speaker explicitly denies the content of the embedded clause or otherwise reveals that she thinks it is false. Simons (2007) discusses some such examples using question/answer pairs. (9) a. Which course did Louise fail? b. Henry, falsely, thinks that she failed calculus. (Simons, example (18)) These examples are complicated. At first glance, the report (9b) seems discourse parenthetical: the content of the embedded clause appears to be what makes the report relevant to the incoming discourse because it is this content that potentially provides an answer to (9a). On the other hand, in (9b), the speaker cannot be taken as actually offering the proposition Louise failed calculus as even a potential answer to the question posed in (9a); in fact, she makes it clear that she is committed to 14. A discourse relation that is not explicitly marked but inferred on the contents of its arguments is cancellable, although many instances will be very difficult to cancel. 11 Hunter that proposition’s not being an answer. As a result, relating the clause embedded under thinks to (9a) via the relation Question-Answer Pair would seem inappropriate.15 One could also use a negated speech or attitude verb (didn’t say, doesn’t think) or a negative verb (doubts, is skeptical) to block the associated speech act. Even more interestingly, from the perspective of rhetorical structure, is that we can undermine an intuitively discourse parenthetical reading by stringing multiple discourse units together to form multi-part responses. For instance, a speaker can disengage herself from the embedded content of a report with a Comment, e.g. (10), or a Contrast, e.g. (11) and (12): (10) Henry said she failed calculus. He always gets things wrong! (11) Henry said she failed calculus, but he’s wrong. (12) Henry said she failed calculus, but Anna disagrees. In fact, it can take many discourse turns to determine a speaker’s commitment to the embedded content of a report. The excerpt below is from an article in the New York Times.16The first two paragraphs of the article discuss an earthquake that hit Prague, Oklahoma in 2011 and the destruction that it caused. The excerpt provides the third and fourth paragraphs: At a packed town hall meeting days later, Ms. Cooper said, state officials called the shocks, including a 5.7 tremor that was Oklahoma’s largest ever, “an act of nature, and it was nobody’s fault.” Many scientists disagree. They say those quakes, and thousands of others before and since, are mainly the work of humans, caused by wells used to bury vast amounts of wastewater from oil and gas exploration deep in the earth near fault zones. And they warn that continuing to entomb such huge quantities risks more dangerous tremors — if not here, then elsewhere in the state’s sprawling well fields. These paragraphs work together to present possible explanations for the earthquake introduced in paragraphs 1 and 2. Let’s simplify the example for the sake of our discussion. (13) a. In November 2011, a 5.0 magnitude earthquake shook Prague, Oklahoma. b. State officials said that the earthquake was an act of nature that was nobody’s fault. c. However, many scientists argued that the quake was caused by wells used to bury vast amounts of wastewater from oil and gas exploration deep in the earth near fault zones. 15. In (9), we can infer a negative answer to (9a), namely that Louise did not fail calculus (Groenendijk and Stokhof, 1984). However, I would like to distinguish (9b) from a report such as (9b’): “Henry thinks/said that she didn’t fail calculus". In (9b’), the speaker is really offering the embedded content as a negative answer to (9a), albeit a tentative negative answer. In (9b), the speaker has an answer that is completely independent of what Henry thinks or says—that’s why she can judge that Henry is wrong. The point of bringing Henry into it, then, is not so much to use him as a source or as evidence for a potential answer, but to indicate that the speaker is aware that Henry might answer the question differently and thinks that Henry is confused. Thus, Henry’s commitments are rhetorically central in (9) in a way that they are not in (9b’). 16. ‘As Quakes Rattle Oklahoma, Fingers Point to Oil and Gas Industry’, by R. A. Oppel Jr. and M. Wines. The New York Times, April 3, 2015. 12 Reports in Discourse (13b) and (13c) oppose two viewpoints, much like (12). At this point in the article, we can only take the author to be presenting two possible explanations.17 The rest of the (lengthy) article makes it clear that the authors side with the scientists, however. The following paragraph, taken from the same article, reveals this endorsement: But in a state where oil and gas are economic pillars, elected leaders have been slow to address the problem. And while regulators have taken some protective measures, they lack the money, work force and legal authority to fully address the threats. At this point, the discussion shifts from trying to find an explanation for the earthquake to a discussion of why the local government has been so slow to address the problem brought up by the scientists. The assertions are no longer in the scope of reports, so we are dealing here with the authors’ commitments. The use of definites such as the problem and the threats presuppose the existence of the problem/threat that that the scientists introduce in (13c). The larger discussion of why the local government has been so slow in addressing the problem also presupposes that the question of what the problem is has been answered. As this paragraph reflects the authors’ commitments, we can infer that they have sided with the scientists, and thus have no commitment to the embedded content of (13b). We can make a similar point with (1). Imagine an utterance of (1) followed by either (14) or (15) (but not both): (14) She must be covering for him, though, because I spotted them together at the market this morning. (15) But he was only an hour away. He could have come if he’d wanted to. There must be some other reason. In judging (1) alone, an interpreter would be entitled to infer that the speaker is committed to the possibility of John’s having been out of town and is using this possibility to provide a possible explanation of why John didn’t come to her party. However, once the speaker continues with either (14), which denies that the potential explanandum holds, or (15), which denies that the potential explanandum actually explains John’s absence, an interpreter is no longer justified in inferring even a hedged Explanation relation between (1a) and (1b). The attitudes that the speaker takes towards the truth of the embedded content of a report and its discourse function are not revealed by considering only the discourse move immediately preceding the report. Starting definitions of discourse parenthetical reports, like that offered by Simons (2007) or the attachment-based proposals offered by various rhetorical frameworks, are driven by the intuition that the embedded contents of these reports play a certain discourse function. Intuitions are primed by considering pairs consisting of a report and a single, preceding discourse unit. The problem brought out by examples like (10)-(13) is that if we place these pairs in a larger discourse context, our intuitions about the discourse function of the report, and what inferences are licensed by it, 17. The fact that scientists are often taken to be more reliable sources than state officials with regard to causes of natural disasters might lead a reader of (13) to suspect that the writers side with the scientists and wish to adopt their explanation. However, we do not get that from looking at the discourse structure of (13) alone. This kind of world knowledge, discussed by de Marneffe et al. (2012), also plays an important role in the interpretation of reports in discourse, but the focus of this article is on how the rhetorical structure of discourse, and the anaphoric connections between discourse units, affects interpretation. 13 Hunter can change significantly. But then what should we say about the reports in such cases? Are they discourse parenthetical or not? What features shall we use to decide? Attachment-based accounts provide no clear answer. Simons’ discussion waffles on these questions as well. Of examples like (9) (her (18)), she says that: “main point content cannot be identified with the content of either the subordinate clause alone or the main clause. Rather, main point content emerges from the interaction between the subordinate clause content and the attitudes to that content expressed by the other predicates used.” This undermines her definition of discourse parenthetical reports. In the beginning, she defines a discourse parenthetical report as one whose embedded clause conveys the main point of the utterance, but later, she seems to suggest that reports like (9b) should count as discourse parenthetical. Although she notes this tension, she does not go on to work it out, as her focus is on other issues, so the ensuing discussion does not point to a path down which rhetorical accounts of discourse parentheticals can go. 3.3 Distinct functions In the previous sub-section, we discussed examples in which a report is offered as a response to a single discourse unit (e.g. Which course did Louise fail?), but in which the full discourse contribution of the report—the information that determines what conclusions can be drawn from the report—can only be understood by looking at chunks of discourse involving multiple discourse units. In this section, we note that the other direction is possible as well: sometimes a single report is a coherent response to more than one discourse unit. In other words, the embedded clause and the attribution predicate might each have a discourse function relative to the incoming discourse, but these functions are distinct and we might not catch both of them if we only consider a single incoming discourse unit (such as a question). Let’s return to (13), but alter it slightly to bring it more in line with the original excerpt. (16) a. In November 2011, a 5.0 magnitude earthquake shook Prague, Oklahoma. b. State officials said that the earthquake was an act of nature that was nobody’s fault. c. Many scientists disagree. d. They argue that the quake was caused by wells used to bury vast amounts of wastew- ater from oil and gas exploration deep in the earth near fault zones. (16c) describes the scientists’ attitudes, and (16d) intuitively elaborates on this description. Yet (16d) can only be an Elaboration on (16c) if we treat the attribution predicate as rhetorically relevant; the embedded clause alone does not talk about the scientists’ attitudes, but only about the object of their attitudes. Thus locally—that is, looking at the report in (16d) as a response or follow up to (16c) alone—the main point of the report seems to be to provide more information about what the scientists hold and the report is not obviously construed as discourse parenthetical. However, as we saw in the last subsection, if we look at the way (16b-d) (or the associated part of the original excerpt) is functioning in the larger discourse, we see that the embedded clause of (16d) serves a higher discourse function, namely that of conveying a possible explanation of why the earthquake occurred. On this level, it is the embedded content that seems more rhetorically important or that carries the main point. 14 Reports in Discourse Once again, we see that the effect of reports on discourse structure can be subtle and diffuse and the intuitions about main point and discourse relevance triggered by looking at pairs of utterances, e.g. (1), do not carry us through to a full account of reports in discourse. 4. Modeling Reports in Discourse The discussion from §3 highlights three issues that a theory of discourse parenthetical reports needs to address. First, because of hedged commitments, the use of a discourse parenthetical report can be taken to be at most a speech act that conveys the possibility of a rhetorical connection between two eventualities. Second, because of examples of no commitment, even an act of hedging seems too strong in many cases—a speaker might use a report that at first seems discourse parenthetical only to later reveal a lack of commitment to the embedded content. Third, because of distinct functions, both parts of a discourse parenthetical report can enter into discourse relations with the larger context; in other words, both can serve a discourse function in the sense defined by a theory of rhetorical structure. As these issues remain unresolved, our working definition of a discourse parenthetical report remains vague: a discourse parenthetical report is one in which the embedded clause plays a major role in making that report relevant to the preceding discourse. The aim of this section is to clarify what it is for a report to be discourse parenthetical and to develop a model of the discourse contributions of such reports. The model that I favor is presented in §4.2. But first, §4.1 lays out my reasons for rejecting a very different, but seemingly attractive, alternative model. 4.1 Separate layers of information? The problems of hedged commitments and no commitment stem largely from a conflict between the intuitive discourse function of a discourse parenthetical report and the kind of speaker commitments that such a function normally entails. An appealing hypothesis, then, is that discourse structures should track information about discourse relations and information about a speaker’s attitude to those relations in two separate dimensions (cf. Potts (2005)’s multi-dimensional account of (not-) at-issue content). The PDTB annotation approach suggests, although it does not entail, one such model. Recall that the content of the attribution predicate of a discourse parenthetical report never contributes to the argument of a discourse connective in the PDTB. Nevertheless, this information is stored alongside the annotations and could in theory be exploited during the interpretation of the annotations. Such a two-dimensional approach might also work to model Simon’s claim that the two clauses of a discourse parenthetical report “convey two different types of content” (Simons 2007, p. 1053), although because Simons does not provide a model or a formal notion of discourse function, she is not committed to such an approach. The problem with a two-dimensional model that treats the attribution predicate and the embedded clause as conveying two different types of content is that it does not square with the observation of distinct functions. If both parts of a discourse parenthetical report can enter into discourse relations, i.e., both parts make the same type of contribution to discourse structure, then whatever differences may exist between the types of content conveyed by the two clauses, it is not a difference that motivates a model in which the attribution predicate contributes the whole of its content to one 15 Hunter dimension or layer of discourse structure while the embedded clause contributes the whole of its content to another.18 An alternative approach would be to posit two layers of information, one that tracks rhetorical relations and one that tracks commitment to those relations, but allow the attribution predicate to contribute to both layers. Danlos and Rambow (2011) proposes an account along these lines. The general idea is to annotate the relations that intuitively hold between discourse units without regard for the speaker’s commitment to the content of those units. Then, as each labelled attachment is constructed, formulas are added to the annotation that contain information about the speaker’s attitude towards each argument of that relation. These formulas take the form: f(e, s) =Mod, Pol, where s is an agent of a propositional attitude, e is the eventuality that serves as the object of the attitude, and f is a propositional attitude function that maps a pair (e, s) to a judgment about e. Following Saurí and Pustejovsky (2009), these judgments have a modal component (Mod)—-which can take the value certain (CT), probable (PR), possible (PS), or unknown (U)—and a polarity component (Pol)—positive (+), negative (-), or unknown (u). Finally, each discourse relation in the annotation structure is subscripted with a source. In general, the source will be the writer or speaker, but in the case of reports, the source can be the source of the reported content. To illustrate, in (17), the Narration relation that intuitively relates Bill and Jane’s attending dinner and their going dancing would be attributed to Bill. (17) [Jane had dinner at my place on Thursday,]α [and Bill said that]βatt [she went dancing afterwards.]β This would (skipping a few details that are irrelevant here) yield the following, two-part annotation: (i) AttributionWr(βatt, β), NarrationBill(α, β) (ii) f(eα,Wr) = CT + ∧ f(eβ,Wr) = Uu ∧ f(eβ, Bill) = CT+ ‘AttributionWr(βatt, β)’ means that the writer is committed to an Attribution relation between βatt and β and ‘NarrationBill(α, β)’ means that Bill is committed to a Narration relation between α and β. ‘f(eα,Wr) = CT+’ means that the writer is certain that eα holds (similarly for ‘f(eβ, Bill) = CT+’), and ‘f(eβ,Wr) = Uu’ means that both the truth of eβ and the author’s attitude towards its truth are unknown.19 While this proposed annotation captures many intuitive features of (17), the proposed account fails to make the right predictions. First, (17) supports neither the inference that Bill is committed to Narration(α, β) nor the inference that he is committed to eβ having occurred; it justifies at most the conclusion that the speaker is committed to Bill’s being committed to these things—the speaker could be making all of this up just to get Jane in trouble. Second, there is a question of why we 18. Maier and Bary (2015), following work by Potts (2005), develops a different kind of two-dimensional model in which the embedded clause of a discourse parenthetical report contributes to one dimension and the entire content of the report (including the embedded clause again) contributes to another. Maier and Bary do not explore a discourse-based account of parenthetical readings, opting rather for an ambiguous semantics for reports that distinguishes between parenthetical (two-dimensional) and non-parenthetical (one-dimensional) reports. For this reason, I will not explore their account in the current paper; nevertheless, I suspect that the troubles brought out by distinct functions will apply to their account as well, as such examples count against a simple binary distinction in which only the matrix clause or only the embedded clause is active in a given interpretation. 19. It is unclear whether the speaker’s attitude towards the attribution predicate is recorded. It does not appear to be, though it is not obvious in Danlos and Rambow (2011) why it would be excluded, since it figures in discourse relations. 16 Reports in Discourse would want to track Bill’s commitments in the first place. The units βatt and β are only relevant to α insofar as the speaker is committed to there being a relation between α and some part of the report. Discourse structures track speech acts of the speaker because it is the speaker who chooses the discourse. It is also unclear why annotations of discourse structure must track the distinction between a speaker’s thinking an eventuality is possible and her thinking that it is probable. Distinctions like this will in general be entailed by the semantics of the embedding verb chosen (know is factive, etc.) or by the way that the report is used in the larger discourse context. This information doesn’t require an extra layer of annotation on discourse structure. Plus, even if we wanted to make such modal distinctions, they could be captured by appending modal operators to the relations (e.g., ♦Explanation(α, β)). We need modal relations in any case to handle responses like: Well, he was out of town, so maybe that’s why he didn’t come, in which the speaker expresses her commitment to the potential explanans but hedges her commitment to the relation itself. Tracking commitments with modalized relations also entails the requisite commitments to the arguments themselves: if a speaker is committed to ♦R(α, β) for a veridical relation R, then veridicality entails that her level of commitment to β be at least ♦β, which is consistent with either full commitment to β or commitment only to ♦β. Thus, by putting the modal on the relation, we get the right results for both the relation and the arguments. By contrast, merely putting the modal on the argument itself (which is equivalent to what Danlos and Rambow propose) does not achieve what is intuitively required—a modal relation. The import of the report in (1) is not correctly captured as Explanation(1a, ♦he was out of town), but as ♦Explanation(1a, he was out of town). The more general point is that if a speaker uses a discourse parenthetical report to hedge her commitment to a discourse relationR(α, β), then, as explained in §3.1 and §3.2, she is not performing a speech act whose content is R(α, β). Annotating (1) with the formula Explanation(1a, he was out of town) and then separately adding the information that the speaker is not fully committed to this content simply isn’t the right thing to do in a rhetorical theory, because there is no speech act of Explanation that has taken place; the speech act just is a proposal of a possible explanation. One has to take the semantics of speech acts into account in calculating what speech act took place; consequently, the veridical nature of veridical relations and their discourse function cannot actually be separated as a two-dimensional or two-layer theory would propose. The natural thing to do within a rhetorical theory is to see the way that speakers use discourse parenthetical reports as speech acts in their own right by assigning them a different kind of discourse relation and then associating these relations with their own semantics. I develop such an account in the next subsection. 4.2 An integrated model I develop my account of reports in discourse by extending SDRT and modifying attachment-based analyses of discourse parenthetical reports to handle the three issues outlined in §3. Let’s begin with the problem of distinct functions, that is, the fact that both parts of a discourse parenthetical report r can be rhetorically relevant by simultaneously being linked to distinct discourse units in the discourse graph for the discourse preceding r. To capture these data, we must allow both parts of a report to contribute to arguments of discourse relations, which in turn requires that we distinguish the attribution predicates of all reports from their embedded clauses by segmenting the two parts as separate discourse units. Thus a given report r will contribute two segments to discourse: the 17 Hunter attribution predicate, whose content entails ∃p.x said (thinks...) that p for some agent x supplied by the report, and the embedded clause, which specifies the content of p. Segmenting reports this way is somewhat controversial, because the attribution predicate does not denote a precise eventuality on its own; it merely tells us that something was said (or thought, etc.). Only when combined with the embedded clause does it adequately specify the requisite speech event (or attitude). This gives some merit to the PDTB and CDT choice to not treat reports as contributing two distinct segments despite the fact that reports do convey two eventualities: that of the saying or thinking, etc. and that described by what was said or thought, etc. The benefits of segmentation outweigh the complications, however. We have too much evidence that both clauses of a discourse parenthetical report can be discursively relevant to ignore one or the other. Plus, the idea of creating separate discourse units for intimately intertwined contents is already required for treating other phenomena such as presupposition in SDRT. In the case of presupposition, there is no syntactic unit that circumscribes a discourse unit for the presupposition; nevertheless, a separate unit for the presupposition is semantically motivated and required in the discourse structure.20 The next step in modelling the discourse function of discourse parenthetical reports is to de- termine how the attribution predicate and embedded clause are related to one another. Recall that Hunter et al. propose a special relation, Source, to handle discourse parenthetical readings. In con- trast, I propose that all reports be annotated with Attribution, regardless of whether they receive a parenthetical reading or not. That is, given two segments, α and β, where α labels the attribution predicate of a report r and β labels the embedded clause, I propose that r contributes the following structure to a discourse graph: α β Attribution Structurally, Attribution is a subordinating relation, which reflects the syntactic structure of discourse parenthetical reports. For its semantics, suppose we have an instance of Attribution(α, β) for some discourse units α and β, and let Kα be the DRS that represents the content of α, and Kβ , the DRS that represents the content of β (see Kamp and Reyle (1993)). Kα should entail that some agent x stands in some attitude A to a proposition p. For example, the attribution predicate of Jill said John was out of town entails the content said(Jill,p) for some proposition p (i.e. Jill said something). The formula Attribution(α, β) is then true just in case the content of Kα is true and the content of Kβ specifies the content of the proposition p. Or in dynamic semantic terms, update with Attribution(α, β) at a world w and assignment f extends f to an assignment g that classically satisfies Attribution(α, β) at w just in case w and f likewise support update with Kα and the resulting assignment g assigns p a subset of the w-accessible worlds w′ that support update with Kβ given the assignment g. More formally: • (w, f)JAttribution(α, β)K(w, g) iff g ⊃ f and ∃x∃p.Kα |= AKα(x, p) and (w, f)JKαK(w, g), andg(p) ⊆ {(w′, k) : (w′, g)JKβK(w′, k)} 20. Thanks to Laure Vieu for discussions on this point. 18 Reports in Discourse where AKα is an attitude predicate determined by the embedding verb, e.g. said, thought, etc., contributed by α. To capture the discourse function of discourse parenthetical reports, we need to model the relation between the report and the incoming discourse context. I propose that this relation has three features, which come together to make a report discourse parenthetical. First, the embedded clause must be related directly to a discourse unit introduced in the discourse preceding the report. This feature is common to all attachment-based solutions. Second, the relation must be distinct from the relation via which the attribution predicate is related to the incoming context (if the attribution predicate is so related). Third, the relation connecting the embedded clause to the preceding context will be a modal relation indicating a hedged commitment from the speaker.21 The function of (1b), for example, is to provide a possible explanation of John’s absence from the party. Modal relations are triggered by reports as follows. Suppose we have an instance of Attribution with the form Attribution(α, β) and suppose that there is a discourse link whose second argument is β and whose first argument, γ, is found in the discourse preceding α. The link between γ and β will be labelled ♦R for some discourse relation R. More precisely, we can add the rule PA, for parenthetical Attributions, to the logic of SDRT: PA: (∃e.e(α, β) ∧ l(e) = Attribution)→ (∃γ∃e′(γ