1.1 The puzzle of embedded V2

This paper examines the hypothesis that certain constructions, so called main clause phenomena, are associated with assertion. The empirical focus of this paper is embedded Verb Second [EV2], a well-studied main clause phenomenon since Andersson (1975).

In languages like Swedish and German, main clauses are obligatorily V2:¹

In embedded clauses, however, the empirical picture is more complicated. In complements of verbs like say, V2 is possible, though not obligatory:²

Verbs like regret and doubt, however, do not allow V2-complements:

The aim of this paper is to understand the conditions under which V2 is licensed in embedded contexts, and how to characterize the interpretation of EV2-clauses.

1.2 Embedded V2 and assertion

The general form of explanation for the above contrasts, following Hooper & Thompson (1973) [H&T] (building on earlier observations by Emonds 1970), is that the availability of embedded main clause phenomena is linked to the presence vs. absence of illocutionary force, and that verbs like think and say licence assertive complements, whereas verbs like doubt and resent do not. Looking at five classes of predicates, Classes A–E, H&T propose that Classes A (non-factives like say), B (non-factives like believe), and E (cognitive and perception factives like discover and see) license both asserted complement clauses and main clause syntax, whereas Classes C (non-factive ‘response stance’ verbs like doubt and accept³), and D (emotive factives like resent) license neither asserted complements, nor main clause syntax.

While H&T focused on English main clause phenomena (e.g. VP-preposing, topicalization, etc.), this pattern of distribution has generally been found to hold up also for EV2 (e.g. Wiklund et al. 2009, Julien 2009; 2015, Bentzen 2010, Wiklund 2010, Jensen & Christensen 2013, Djärv et al. 2017, Caplan & Djärv 2019, Djärv 2019):⁴

H&T’s idea that main clause phenomena are associated with assertion has also been present in the literature on EV2 since Andersson (1975). However, though this hypothesis has been widely adopted, it has proven difficult to pinpoint exactly in what sense V2-complements are asserted, and how to formally characterize this claim. Addressing the question about the precise interpretation of EV2-clauses is the primary goal of this paper.

With respect to this issue, current theories fall into two general camps: those appealing to the discourse status of the embedded proposition [p], and those appealing to the presence of a particular type of commitment to p (both in line with research on matrix assertions since Stalnaker 1978 and Hamblin 1971, a.o.). See for instance Truckenbrodt (2006), Wiklund (2010), Gärtner & Michaelis (2010), Krifka (2014), Julien (2009; 2015), Woods (2016), Caplan & Djärv (2019), and Djärv (2019).

Besides the embedding verb, another important observation that has been used to motivate these different types of accounts is the effect of negation on the availability of EV2. As shown in (6), verbs like believe lose their ability to embed EV2-clauses when negated; mirroring the pattern found with inherently negative verbs like doubt (3-b)–(4).

According to Truckenbrodt (2006), EV2 is licensed when the embedding context entails a commitment to a belief that p (see Section 3.2). Thus, EV2 is available under believe, but not under not believe or doubt. As Caplan & Djärv (2019) point out, however, this also predicts that EV2 should improve under verbs like doubt, when negated. While judgements to this effect have in fact been reported (e.g. Wiklund 2010), it’s not clear how robust they are. An experimental study from Djärv (2019), (looking at German EV2), replicated the judgements in (2)–(6), but did not find a significant effect of negation on the acceptability of EV2 for verbs like doubt.⁵⁶ Caplan & Djärv (2019) observe instead that the availability of EV2 across verb types and polarities tracks the potential for introducing the embedded proposition as discourse new information, as illustrated in (7), and argue based on this that the availability of EV2 across embedding environments is better explained in terms of the discourse status of p as old vs. new information.

Thus, while both type of accounts make the same predictions for verbs like think and say, with respect to negation, we find that the overall effect of negation across H&T’s predicate types is better accounted for in terms of the discourse status of the embedded proposition. This is because the commitment-based account predicts an interaction with negation for the response stance verbs, whereas the discourse novelty-based account makes no such prediction, given that these verbs, regardless of their polarity, require p to be discourse old.

As we shall see in Section 2, however, examination of naturally occurring data shows us that commitments also play a role, though in a different way from how it’s previously been characterized. On previous accounts (Section 3.2), EV2-clauses represent assertions that are either fully anchored to the current discourse context, including the speaker, or fully anchored to the reported context, including the matrix subject. I show that at their core, EV2-clauses represent discourse moves whereby the speaker adds the embedded proposition to the conversational table as an issue for discussion (in line with the claim that EV2 involves an attempt to update the context). In many cases, this is accompanied by the speaker expressing a public commitment to this proposition, as in matrix assertions. However, I also identify cases that involve a commitment shift. Here, the assertion is not anchored to the speaker, but to a contextually specified commitment anchor, which is identified by the matrix clause, but does not need to correspond to the matrix subject (contrary to previous accounts). This allows the speaker to raise the issue represented by the embedded clause with the force of an assertion, without committing to it. The speaker can thereby immediately reject or take issue with the proposal it represents. Crucially, even though the assertion is anchored to a contextually specified commitment anchor, the discourse move still pertains to the current conversational context, rather than to the reported context, as on previous accounts.

In addition to these assertive discourse moves, this paper also identifies a third type of discourse move associated with EV2. Here, the EV2-clause functions as a type of biased question, similar in form and function to (unembedded) rising declaratives (e.g. Lisa won?). This shows us that EV2 doesn’t need to be associated with assertive discourse moves at all, but —depending on properties of the matrix clause and prosodic factors— can also be associated with inquisitive discourse moves. This presents a problem for accounts that analyse V2 in terms of an assertive operator or feature (Section 3.2). Interestingly, while these EV2-clauses represent inquisitive discourse moves at the discourse-level, they are interpreted as regular embedded declaratives within the complex sentences in which they occur. This raises important questions about the composition of these clauses, which I discuss in Section 4.4.

To account for this data, I propose that the interpretation of EV2-clauses is best captured by the statement that they share the conventional discourse effects associated with main clause declaratives, in the sense of Farkas & Roelofsen (2017). On their account of matrix declaratives and interrogatives, differences in the conventional discourse effects of these sentence types follow from their semantic differences (modelled in terms of inquisitive semantics), combined with a basic convention of use, applicable to both sentence types. This proposal —supplemented with the novel assumption that in embedded contexts, the matrix clause may be used to identify a contextually specified commitment anchor for the discourse move represented by the embedded clause— thus provides a unified account of the various assertive and inquisitive uses of EV2-clauses. It allows us to capture the insight from previous work that EV2 is associated with assertion, while offering the flexibility to account also for the inquisitive uses and the commitment-shifted assertions identified here.

1.3 The role of the matrix verb

In addition to their conventional discourse effects, we also find that EV2-clauses can be associated with a similar range of pragmatic discourse effects to matrix declaratives: they may be used for clarification, to signal agreement or disagreement with a previous claim, to persuade, remind, or reassure the addressee, make a promise, etc. In these embedded contexts, we find that the particular pragmatic effects associated with the EV2-clause are closely linked to both the specific matrix verb and other properties of the matrix clause (e.g. person, tense, sentence mood). This observation allows us to address a second central question in the study of embedded main clause phenomena, regarding the precise role of the matrix predicate. Addressing this question is the second goal of this paper:

With respect to this question, current theories fall into three general types: accounts arguing that certain predicates select for a particular type of clause, which is either compatible with main clause syntax (as in Wiklund et al. 2009) or effectively blocks it (as in Kastner 2015);⁷ accounts that take the compatibility between the embedding verb and main clause complements to be essentially pragmatic (e.g. Julien 2009; 2015; Jensen & Christensen 2013; Woods 2016; Caplan & Djärv 2019; Djärv 2019); and accounts where the compatibility is fundamentally semantic (Krifka 2014) (see Section 3.2).

As we shall see, the data examined in Section 2 shows us quite clearly that the role of the matrix clause is fundamentally pragmatic. Specifically, I argue that the role of the matrix clause is to frame the specific type of discourse move represented by the EV2-clause. On this view, the compatibility of a given verb and a V2-complement depends on the specific discourse move represented by the embedded clause together with the overall function of the matrix sentence.

1.4 Outline

Section 2 presents the data. Section 3 summarises the empirical findings and briefly evaluates previous formal accounts of EV2. Section 4 presents the proposal, and discusses the issue raised by the inquisitive EV2-clauses identified in Section 2.3, regarding the mismatch between the discourse effects and the semantic interpretation of these clauses. Section 5 discusses the role of the matrix clause. Section 6 summarises.

2.1 Assertive discourse moves fully anchored to the current discourse

The EV2-clauses discussed in this section represent assertions by the speaker, in the current discourse context (such uses have been discussed in previous literature, since Andersson 1975, e.g. Wechsler 1991; Julien 2009; 2015; Woods 2016; Caplan & Djärv 2019; Djärv 2019). As shown in (9)–(15), such cases can further be associated with a range of more specific uses, also available to matrix declaratives.¹⁰

In (9), the speaker clarifies, re-states, or agrees with a previous claim that p. In (9-a) and (9-b), the previous claim that p was by the speaker, as signalled by the matrix clause (‘I just meant’, ‘I’m just saying’), marked in boldface. In (9-c), the speaker expresses their agreement with someone else’s point of view, as indicated by the matrix clause (‘I agree with those who think’).¹¹

We also find examples where the speaker directly contrasts their opinion that p with an alternative point of view in the context (10)–(11). These cases differ in terms of the point of view highlighted. In (10), the point of view highlighted is that of the speaker, as indicated by the matrix clause (‘I’m myself of the opinion’, ‘I would be more inclined to think’). In (10-c), the quotation in the embedded clause signals that the speaker is taking issue with and rejecting a previous claim that a ban is “the solution to all problems”.

In (11), the point of view highlighted is that of the addressee. Unlike in (9) and (10), where we saw first person matrix clauses with verbs of saying, thinking, or opinion, here, we find second person matrix clauses with cognitive factives like get it, learn, realize, and know (‘don’t you get that’, ‘you have to learn’, ‘when will you realize’, ‘you know that p, right?’).

In (10)–(11), the speaker contrasts their opinion that p with the point of view of someone in the current discourse context; highlighting either their own point of view (10) or the addressee’s (11). However, we also find “contrastive” uses of EV2 where the point of view highlighted is that of someone in the reported discourse context:

We also find a large number of cases where the matrix clause is used, broadly speaking, to motivate the assertion represented by the embedded clause. (13-a) and (13-b) are persuasive, or argumentative, as signalled by the matrix clause (‘if you’ve done so and so, you will know’ or ‘based on so and so, I’m saying’). In (13-c), the matrix clause justifies the embedded assertion (‘now I’ve learned that’).

We also find cases where the speaker provides a reminder that p (14). Here, we get direct reminders with matrix clauses like ‘never forget that’ (14-a) or ‘know that’ (14-b), as well as generic reminders, with matrix clauses like ‘it’s often forgotten that’ (14-c).

Finally, we find promises and reassurances, involving first person subjects and the verb promise.

In all of these cases (9)–(15), it’s clear that the embedded proposition p is asserted by the speaker. As shown in (16), the complex sentences discussed in this section can all be paraphrased as two simple matrix clauses with the same conversational effects as their complex counterparts.

Additionally, in all of these cases, it would be odd for the speaker to follow up with a rejection of the embedded proposition.

To formally characterize these assertive uses of EV2-clauses, I adopt Farkas & Roelofsen’s (2017) analysis of the conventional discourse effects of matrix declaratives (building on Farkas & Bruce 2010): the issue of p is pushed onto the conversational table as an issue for discussion, and p is added to the speaker’s discourse commitments.¹²¹³ Following Farkas & Roelofsen (2017), I also distinguish the conventional discourse effects of a given sentence type from the pragmatic discourse effects of the utterance (e.g. whether it is intended to be informative, or provide a reminder, clarification, justification, or a promise, etc.), which depend on the particular utterance situation. Adding to the picture from Farkas & Roelofsen, the data discussed here shows us that in complex sentences, the embedded clause may be part of the speaker’s assertion. In this case, the utterance situation which frames the discourse move represented by the embedded clause includes the matrix clause. Thus we get clarifications with matrix clauses like ‘I just meant that’ or ‘I’m just saying that’, cases where the speaker takes issue with the interlocutor’s position with matrix clauses like ‘don’t you get that’ or ‘when will you realize that’, promises with ‘I promise you that’, and reminders with ‘don’t forget that’.

Note additionally that the proposition represented by the matrix clause is also asserted by the speaker. For instance, in (15-a) (‘I promise that this cage isn’t small, it’s big’), one possible response would be Den är visst liten. (‘It is small.’), taking issue with the assertion represented by the embedded clause (that the cage isn’t small). However, another possible response would be Lovar du det? (‘Do you promise?’), following up on the assertion represented by the matrix clause (that the speaker promises that p). Thus, a complex assertion of this type introduces two possible future directions for the conversation.¹⁴

In addition the these cases, we also find two kinds of uses of EV2-clauses where the speaker does not express a commitment to p: assertive discourse moves anchored to a contextually specified commitment anchor (Section 2.2) and inquisitive discourse moves, which function as a type of biased question regarding the issue of p (Section 2.3).

2.2 Assertive discourse moves with commitment shift

In matrix declaratives and in the cases we saw in the previous section, the discourse move of adding the issue of p to the table and the expression of a public commitment to p are both anchored to the speaker. In the examples discussed here, however, they come apart. As in the previous section, the EV2-clauses in (17) represent assertive discourse moves. The difference between the two cases is that here, the assertion of the embedded proposition is not anchored to the speaker, but to a contextually specified commitment anchor, which is specified by the matrix clause, but —crucially— doesn’t need to correspond to the matrix subject, as shown in (17-b). As shown in (17), this commitment shift allows the speaker to raise the issue of p at the level of the current conversational context, with the force of an assertion, without themselves committing to this proposal. The speaker can thus immediately reject or take issue with it, without making contradictory discourse moves.¹⁵

Similar cases have been discussed by Julien (2009; 2015), who suggests that Mainland Scandinavian EV2-clauses “represent a proposition that is asserted, either directly, by the speaker [as in Section 2.1], or indirectly, with the speaker reproducing a claim that someone else has made” (Julien 2009: 1). To account for these ‘indirect assertions’, Julien (2015) adopts Krifka’s (2014) account (which I discuss in Section 3.2), whereby EV2-clauses represent assertions by the matrix subject, semantically, and are evaluated with respect to the reported context.

From looking at the naturally occurring cases in (17), however, it is clear that these utterances do more, pragmatically, than simply reproduce a claim that someone else has made. Crucially, even though the assertion of the embedded clause in these cases is shifted to a contextually specified commitment anchor, the discourse move of adding the issue of p to the table still pertains to the current conversational context, rather than to the context provided by matrix clause. This is clear from the speaker’s follow-up, whereby the speaker denies p by asserting ¬p¹⁷ at the level of the current conversational context, where it then becomes available for further conversational uptake, as shown in (18), based on (17-b).

In the case of a denial by an actual discourse participant, a typical outcome is that one of the participants retracts their assertion or that the participants ‘agree to disagree’ (see Farkas & Bruce 2010). In (17)/(18), where the assertion of p is not anchored to an actual discourse participant, the effect of the speaker’s denial (18-b) is that the issue of p simply gets removed from the table; with ¬p remaining as an issue for further discussion. Evidence for this comes from the availability of the Swedish polarity sensitive response particles ja and jo in responding to the utterances in (17), with vs. without the speaker’s denial.

For positive questions and statements, such as (19), the positive response particle ja signals affirmation and the negative response particle nej signals rejection.

For negative questions and statements, however, as in (20), ja is infelicitous (20-a). Here, nej is used for affirmation (20-b), and the polarity-reversing particle jo is used for rejection (20-c). (See Holmberg 2013, and also fn. 17, for further discussion.)

Considering again (17-b), we find that without the speaker’s follow-up (21-a), the issue represented by the EV2-clause is available for an interlocutor to take issue with and reject. Since this is a negative statement (‘I haven’t raised him right…’), the interlocutor has to use the response particle jo in this case (21-b); a ja-response is infelicitous and a nej-response can only be used for affirmation (21-c).¹⁸

However, with the speaker’s follow-up (22-a)–(22-b), the interlocutor can now only agree with or reject the shifted assertion of p indirectly, by targeting the speaker’s follow-up assertion of ¬p, which is a positive statement (‘but I have’); either with a ja-response ((22-c); agreeing with ¬p), or with a nej-response ((22-d); rejecting ¬p). The jo-response, however —which we saw in (21) would have to be used in order to reject the shifted assertion of p represented by the (negative) EV2-cause— is now infelicitous (22-e).

What this suggests is that the issue of p is effectively removed from the table by the speaker’s follow-up. Interestingly, the fact that p is no longer available for the addressee to directly respond to in (22) doesn’t seem to be due simply to the fact that there is an assertion intervening between the (shifted) assertion of p in (22-a) and B’s response in (22-e). In (23) where A asserts p (directed at C) (23-a) and B intervenes with a denial of p, asserting ¬p (23-b), C can respond to A either by siding with and agreeing with B, effectively echoing their assertion of ¬p (in which case, a ja-response is used; (23-c)), or by responding directly to A with a rejection of their assertion of p (in which case, a jo-response is used; (23-d)).¹⁹

Here, I do not offer a formal account for how this happens, but intuitively, this seems to be linked to the fact that the assertion of p is not linked to an actual discourse participant. Rather, this commitment shift is used by the speaker as a type of conversational device, enabling them to raise the issue of p in the guise of an assertion, in order to immediately reject or take issue with the proposal it represents.

As in Section 2.1, the proposition represented by the matrix clause is also asserted by the speaker. Considering again (17-b) (‘I’ve heard both this and that…’), we find that an interlocutor can also respond to this by taking issue with the claim ‘I’ve heard that…’, for instance with a response such as När har du hört det? (‘When did you hear that?’). This assertion is fully anchored to the speaker: it would be infelicitous for the speaker to follow up with a denial, e.g. ‘but I’ve never heard that’.

2.3 Inquisitive discourse moves: embedded declarative biased questions

Finally, we find an interesting case where the discourse move represented by the EV2-clause is not that of an assertion, but of a question. This is illustrated in (24).

A sentence of this syntactic form could in principle be used in two ways. With falling intonation, it would naturally be interpreted as a type of evidential statement, where the matrix clause provides support for the speaker’s claim that p by pointing to the existence of a reliable source of evidence for p (cf. (13)). However, the question mark in (24) suggests that it involves rising intonation. In this case, the sentence is interpreted as a type of biased question, similar to rising matrix declaratives, such as Lisa won?. Here, the matrix clause signals that the speaker has some evidence for p, though not necessarily evidence that they take to be fully reliable.

It is worth noting that rising declaratives have been shown to also allow for non-inquisitive uses (e.g. Ward & Hirschberg 1985, Malamud & Stephenson 2015, Jeong 2018). (24), however, is clearly inquisitive. Evidence for this comes from the observation that a natural continuation by the speaker in (24) would be a follow-up question like Stämmer det? (‘Is that right?’), targeting the question represented by the EV2-clause. Another diagnostic, originally from Gunlogson (2008), is the availability of responses like ‘Oh, I didn’t know that’ vs. ‘Yes, that’s right’. Whereas ‘Oh, I didn’t know that’ signals acknowledgement of new information, ‘Yes, that’s right’ signals affirmation based on the speaker’s own evidence or judgement. As discussed by Jeong (2018), the Yes/Oh diagnostic is helpful in distinguishing between inquisitive and non-inquisitive uses of rising declaratives:²⁰

In (27-b), which involves an EV2-clause with falling intonation, a natural response would be Jaha (similar to ‘Oh I see’), signalling acknowledgement of the new information offered in the embedded clause; the same pattern we find in response to the unembedded falling declarative in (27-c).

In (28-b), which involves an EV2-clause with rising intonation, on the other hand, a natural answer would be Ja, det stämmer (‘Yes, that’s right’) or Ja, så är det (‘Yes, that’s how it is’), signalling an affirmative answer to B’s question. Here, the Jaha (‘Oh I see’) response is infelicitous. This is the same pattern that we find with the unembedded rising declarative in (28-c).²¹

The same applies to the naturally occurring example in (24), repeated in (29). Here, a natural response would be Ja, det stämmer (‘Yes, that’s right’) (29-a), offering an affirmative answer to the biased question represented by the embedded clause. A response such as Jaha, då vet jag (‘Oh I see, now I know’) —which would be felicitous if the sentence was pronounced with falling intonation— is not felicitous here (29-b). As in the previous two sections, the speaker also asserts the proposition represented by the matrix clause. An interlocutor can therefore also choose to ignore the question of whether p and instead respond to the speaker’s claim that they have heard p, as shown in (29-c).

3.1 Summary of results

In Section 2, I examined the discourse functions associated with EV2 in a sample of naturally occurring data. In this sample, we saw three clearly distinct uses of EV2-sentences. In the first case (Section 2.1), the EV2-clause is interpreted like a (falling) matrix declarative: the issue of p gets pushed onto the conversational table as an issue for discussion, and the speaker expresses a discourse commitment to p. Among these cases, we observed a range of more specific discourse effects, also available to matrix declaratives, which we saw are closely tied to the specific properties of the matrix clause. Additionally, I identified two cases where the speaker raises the issue of p, at the level of the current discourse context, without expressing a commitment to p. The first case (Section 2.2) involves a commitment shift. Here, the discourse move represented by the embedded proposition is also that of an assertion. However, in this case, the assertion is anchored to a contextually specified commitment anchor, thus allowing the speaker to immediately reject or take issue with the proposal represented by the EV2-clause. In the second case (Section 2.3), the discourse move represented by the embedded clause is not that of an assertion, but a type of biased question, similar in both form and function to (matrix) rising declaratives.

The data discussed in this section allow us to address both the question of the precise interpretation of EV2-clauses (5), and the question of the role of the matrix predicate in licensing EV2 (8). Regarding the first question, I propose that the interpretation of EV2-clauses is best characterized in terms of the statement that they share the conventional discourse effects typically associated with main clause declaratives, in the sense of Farkas & Roelofsen (2017): they add the issue raised by the embedded clause to the table as an issue for discussion, and they convey a commitment to the informative content of the clause. In the case of EV2-clauses with falling intonation, the commitment to p can either be anchored to the speaker or shifted to a contextually specified commitment anchor. EV2-clauses with rising intonation, on the other hand, function as a type of biased question, in which case, as I discuss in Section 4, the commitment is trivial.

Regarding the role of the matrix predicate, what the data discussed here shows us is that a given predicate is able to embed a V2-complement, to the extent that the matrix clause as a whole is able to frame the discourse move represented by the EV2-clause. As we have seen, this depends not only on the embedding verb, but also on other properties of the matrix clause such as person, tense, and sentence mood. Here, we have mainly seen verbs of saying and thinking, as well as cognitive factives, in line with previous observations (Section 1). Importantly, however, the different predicates are closely linked to the particular discourse move represented by the EV2-clause. For instance, we have seen clarifications with matrix clauses like ‘I just meant that’ or ‘I’m just saying that’, cases where the speaker takes issue with the interlocutor’s position with matrix clauses like ‘don’t you get that’ or ‘when will you realize that’, promises with ‘I promise you that’, and reminders with ‘don’t forget that’. Additionally, we saw examples involving accept, which —as I mentioned in Section 1— has been found in previous work to not take V2-complements. According to Caplan & Djärv (2019) and Djärv (2019), this is because accept does not readily allow the embedded proposition to be presented as discourse new information.

In the case we saw here, however, the matrix clause was ‘one has to accept that’ (12-a). Similar cases were also noted by Caplan & Djärv (2019), involving matrix clauses like ‘can’t you just accept that’ and ‘accept that’.

Thus, it seems like the ability of accept to embed V2-complements is linked to its ability to frame an assertion of p, in cases where the speaker wants to highlight that the addressee should —but doesn’t already— accept p. As in (16) above, the example in (31) can be paraphrased as two matrix assertions, and achieve the same discourse effects.

A similar point holds also for forget, for which we observed matrix clauses like ‘never forget’ or ‘it’s often forgotten’ (14). These contexts provide natural environments for the embedded clause to provide an actual reminder (by the speaker to the addressee) that p. This is unlike sentences like Anna forgot that Lisa won (cf. (30)), which can only be interpreted as statements about the attitude holder forgetting something.

To summarise, we have seen that EV2-clauses are associated with what appears to be the same range of discourse functions that are available to matrix declaratives, with the addition of the possibility of commitment shift offered by the embedding. In Section 4, I spell out my proposal for EV2 formally, extending Farkas & Roelofsen’s (2017) account of the discourse effects associated with matrix declaratives and interrogatives. First, however, let us briefly consider some previous assertion-based accounts of EV2, in order to see that the data discussed here motivates a new account of EV2.

3.2 Alternative accounts

On previous formal accounts of EV2, EV2-clauses represent assertions that are either fully anchored to the current discourse context, including the speaker, or fully anchored to the reported context, including the matrix subject. As we have seen, however, while EV2-clauses may represent speaker-assertions in the current discourse context (Section 2.1), they may also represent assertions which are evaluated with respect to the current discourse context, but which are anchored to a contextually specified commitment anchor, rather than the speaker (Section 2.2). Additionally, previous accounts of EV2 predict that EV2 should be associated exclusively with assertive discourse moves. As we saw in Section 2.3, however, EV2-clauses may also represent inquisitive discourse moves.

An influential account of EV2 comes from Krifka (2014) (focusing on EV2 in German), who models assertions as representing a change from a conversational state (modelled as a world-time index) where the speaker’s assertive commitment does not hold, to one where it does hold. Assertive states are modelled using the predicate assert (33), which “takes an index i, an addressee variable y, a proposition p and a speaker variable x, and yields the value true iff at i, x has assertive commitments with respect to the proposition p to the addressee y.” (Krifka 2014: 7).

Krifka (2014) assumes, after Rizzi (1997) and others (including Julien 2009 for EV2), that assertions are ForcePs, syntactically, and proposes the following semantics for such clauses:

Uttered at a context c, this will change the input index i minimally to the index i′ such that “at i′, the speaker x is committed to defend the truth of the proposition [I admire Sue] towards the addressee y” (Krifka 2014: 10).

Whereas unembedded assertions involve adding p to the speaker’s list of public commitments, embedded assertions, marked by V2-order, are proposed to describe an event of the matrix subject adding p to their list of public commitments. On Krifka’s analysis, the difference between a sentence with and without EV2 is that with a non-V2-complement, say selects (semantically) for a proposition (p), a CP, whereas with a V2-complement, say selects for an assertion (A), a ForceP (34):

The semantic difference between a sentence with and without EV2 is shown in (36): the non-V2 sentence in (36-a) involves the predicate say in (35-a), and the EV2-sentence (36-b) involves the predicate say in (35-b) and a ForceP-complement (34).

As shown in (36-b), an EV2-sentence like Mary sagte, sie hasst John is analysed as describing an event of Mary asserting that she hates John. Thus, on Krifka’s account, EV2-clauses always represent assertions by the matrix subject.

As we have seen here, however, EV2-clauses, at least in Swedish, represent discourse moves that are anchored to the current speaker and conversational context, either fully or partially. Even the cases identified here as involving a commitment shift (Section 2.2) still represent discourse moves whereby the speaker adds p to the current conversational table. Additionally, as we saw in Section 2.3, EV2-clauses do not in fact need to represent assertions, but can also represent inquisitive discourse moves. Another problem for this account is that it’s not clear how it would capture the availability of EV2 under verbs that do not describe speech acts. Krifka (2014) does acknowledge this problem in the context of verbs like denken (‘think’) and glauben (‘believe’), and suggests that “it is easy to see that a propositional attitude can be characterized by a speech act that an agent would utter if the agent has that propositional attitude. If Mary believes that Bill is at school, then she is willing to assert that Bill is at school, and hence believe can subcategorize for such an assertion” (p. 25). However, this type of explanation would not extend to a predicate like forget (14). As we have seen here, the ability of a given verb to embed a V2-complement is linked to its ability to appropriately frame the discourse move represented by the embedded clause, rather than to the type of event or attitude that it describes.

An alternative account of EV2 (also focusing on German) comes from Truckenbrodt (2006), extending a proposal by Gärtner (2002). On this account, V2 is linked to the presence of context indices in C. For main clause declaratives and interrogatives, Truckenbrodt analyse these as the speaker S wanting to change the world by changing the epistemic states of the speaker or the addressee A (i.e. conveying or requesting knowledge). Falling declaratives are characterized as adding to the common ground that S believes p and additionally calls for A to accept p. This is encoded by context indices in C: A, ⟨epist⟩, ⟨deont⟩, where A and ⟨epist⟩ are responsible for triggering V-to-C movement.

For main clause V2, A and ⟨epist⟩ are proposed to combine with ⟨deont⟩ and give rise to an attempted context update (37). For embedded V2-clauses, however, Truckenbrodt argues that they have assertive ‘proto-force’, which gets absorbed in appropriate embedding contexts, namely those that entail a commitment to a belief that p: “V-to-C movement is connected to assertional proto-force, which becomes assertional force (potential) in unembedded use and can be absorbed under appropriate embedding. The process of absorption is to account for the restrictions on embedded V-to-C” (p. 281). On this account, then, EV2-clauses (unlike main clause V2) do not have proper illocutionary force, but rather, illocutionary force potential, which gets ‘absorbed’ when the embedding contexts entails a commitment to a belief that p. This is then used to explain why EV2 is available under verbs like say and believe, but not under predicates like don’t believe or doubt.

Though the details of Truckenbrodt’s account differ in important ways from Krifka’s, the two accounts end up facing similar problems in light of the Swedish data discussed here. As we have seen, these EV2-clauses —just like matrix declaratives— represent discourse moves whereby the speaker adds the issue of p to the current conversational table as an issue for discussion. Thus, these clauses have proper illocutionary force, rather than ‘proto-force’. Moreover, EV2-clauses do not need to represent assertive discourse moves at all, but can function as questions.²²

A final proposal for EV2, which shares core features of both the current account and the above proposals, comes from Woods (2016). The primary empirical focus of Woods (2016) is embedded inverted questions in English, though she proposes an extension to EV2 in German and Mainland Scandinavian. Woods’ proposal involves both a grammatical component, which is responsible for triggering V-to-C movement, and a separate discourse component. At the level of the grammar, Woods links V-to-C to an Illocutionary Act Phrase (Speas & Tenny 2003). This phrase is argued to host two semantic elements: the (doxastic) assertion operator in (38), which can either be anchored to the matrix context or to the current conversational context, and a perspectival monster (Sudo 2012), which is responsible for fixing the context which determines the perspective holder for the embedded illocutionary act to either the current discourse context (including the current speaker) or the reported context (including the matrix subject):²³

In addition to this grammatical component, Woods’ analysis includes also a discourse component (p. 185–187, 193): the assertion operator of Farkas & Bruce (2010), which Woods invokes only for a subset of EV2-clauses, namely those cases where the assertion is anchored to the speaker and the current discourse context. On Farkas & Bruce’s (2010) model, if a speaker (or author; a) asserts a declarative sentence S[D] against an input context K_i, then that assertion will change the context in three ways. First, the propositional content of S[D] will get added to a’s discourse commitments (DC_a). Secondly, the syntactic form of S[D] and its denotation are pushed to the top of the conversational Table (T). Finally, the propositional content of S[D] gets added to a so-called projected set (ps), which is a set of common grounds that reflects canonical ways of settling the issues on the Table. These discourse effects are achieved via the assertion operator in (39).²⁴

With respect to the discourse component, Woods’ proposal is similar to the current account, the details of which are discussed next. The problem for Woods’ account is how to account for the commitment-shifted cases (Section 2.2) and the inquisitive uses of EV2-clauses (Section 2.3). The operator in (38) predicts that the commitment anchor and the conversational context in which the assertion is evaluated should always shift together. Thus, we expect that EV2-clauses should represent assertions that are anchored either to the current discourse context and the speaker, or to the reported context and the matrix subject (as for Krifka 2014). However, we do not expect to see the more complex discourse move identified in Section 2.2, where only the commitment anchor, but not the conversational context, is shifted. In terms of the inquisitive uses of EV2-clauses, note that even if we supplemented Woods’ proposal with an appropriate discourse-level question operator to capture these cases (instead of (39)), the operator in (38) would still result in the interpretation that the speaker believes p to be true, in conflict with the interpretation of these EV2-clauses as a type of question. As we shall see next, the current proposal —where the only theoretical machinery is the table model— offers a simple and unified account of all of these discourse functions available to EV2-clauses.

4.1 Semantic model (Farkas & Roelofsen 2017)

On Farkas & Roelofsen’s (2017) account of matrix declaratives and interrogatives, differences in the conventional discourse effects of these two sentence types follow from their semantic differences, modelled in terms of inquisitive semantics (e.g. Ciardelli et al. 2013; 2015), combined with a single ‘convention of use’ (building on Farkas & Bruce 2010). As a result, they argue, illocutionary force (assertion and question) operators are not necessary.²⁵

Traditionally, a distinction is drawn between declaratives and interrogatives, in terms of their semantic type.

While both sentences in (40) pertain to the truth of the proposition Lisa won, they differ in terms of their conventional discourse effects. In uttering the declarative in (40-a) in a world w, the speaker is typically assumed to (at least) express a public commitment to p (i.e. that p is true at w) and further propose to add p to the common ground (i.e. the set of propositions that are mutually accepted by the discourse participants; Stalnaker 1978). In uttering the interrogative in (40-b), on the other hand, the speaker is understood to ask the question of whether the proposition Lisa won is true at w or not. With this distinction in mind, declaratives are traditionally analysed as propositions ⟨st⟩, and interrogatives as sets of propositions ⟨st,t⟩. Using the terminology of inquisitive semantics, the semantic representation of a declarative is analysed in terms of its informative (truth-conditional) content, whereas the semantic representation of an interrogative is analysed in terms of its inquisitive content (the issue it raises).

In inquisitive semantics, on the other hand, the semantic value of a sentence is taken to capture both its informative and its inquisitive content; with propositions analysed as singleton proposition sets and interrogatives as multi-member sets; where propositions are viewed as information states. In uttering a sentence (whether declarative or interrogative), the speaker is understood to simultaneously do two things with respect to the current conversation: (i) provide the information that the actual world is contained in the informative content of the sentence, and (ii) raise an issue for which its resolution requires establishing one of the propositions expressed by the sentence.²⁶ For a declarative like (40-a) (represented by the singleton proposition set {p}^↓), this means that the speaker provides the information that the actual world is a member of p, and steers the conversation towards a state where p is common ground. In this case, the inquisitive content of the sentence is trivial, given that the issue raised by the utterance of the sentence is already resolved by its informative content. For an interrogative like (40-b), on the other hand, the informative move is trivial (as the informative content of the sentence covers the whole logical space W). The move of steering the conversation towards one of the information states expressed by the sentence, however, is non-trivially inquisitive, given that in order to reach such a state, the addressee will need to provide additional information, either that Lisa won (p) or that she didn’t win (¬p). (For more discussion of this perspective, see Farkas & Roelofsen 2017, as well as Ciardelli et al. 2013; 2015; Theiler et al. 2019, a.o.)

Farkas & Roelofsen (2017) argue that on this perspective, the conventional discourse effects associated with declaratives and interrogatives follow as a consequence of their semantic content, together with a single convention of use which applies to both sentence types. This convention builds on Farkas & Bruce (2010) and previous work. As in Farkas & Bruce (2010), a discourse context is modelled as in (41) (where a possibility is modelled as the set of possible worlds which are compatible with it):

The commitment set of a participant x, cs(x) is defined as the set of worlds that are compatible with all the possibilities that x is publicly committed to: cs(x) = ⋂commitments(x). Unlike on the Stalnakerian picture, the common ground is not taken to be a conversational primitive, but is derived from the commitments of each of the discourse participants: cg = ⋃{cs(x)|x ∈ participants}. The basic convention of use proposed by Farkas & Roelofsen (2017) is given in (42).

Thus, an utterance of a (falling) declarative sentence has the following effects on the context: (i) p gets added to the speaker’s discourse commitments, and (ii) the issue of p, {p}^↓, gets added to the conversational table; thus steering the conversation towards a state where the participants mutually agree that the actual world is contained in p.

To account for non-canonical uses of interrogatives and declaratives, such as rising declaratives and tag-interrogatives, Farkas & Roelofsen (2017) additionally invoke the clause type markers dec/int and closed/open. In English root clauses, dec/int are marked by declarative vs. interrogative word order, and in embedded clauses, dec/int are marked by the complementisers that vs. whether. In English root clauses, closed/open are marked by falling vs. rising intonation. They assume, based on data discussed in Section 4.4, that rising/falling intonation in English has semantic import in root clauses only (see Farkas & Roelofsen 2017: fn. 12). However, as we saw in Section 2.3, rising vs. falling intonation can have the relevant types of semantic effects also in embedded clauses (see Section 4.4 for further discussion).

On this account, rising declaratives like Lisa won? start out as sentence radicals denoting a singleton set {p}^↓, but take on the meaning of a question, {p,¬p}^↓, when they combine with the clause type marker open, which is marked by rising intonation (see Farkas & Roelofsen’s 2017 Section 4.2 for more detailed discussion):

Additionally, the rising declarative Lisa won? also signals that the speaker has access to some evidence for the possibility that Lisa won (see also Malamud & Stephenson 2015). To account for this, Farkas & Roelofsen (2017) add a special effect to the basic discourse effects discussed above:

Thus, an utterance of a rising declarative like Lisa won? has the following effect on the context: as for unmarked polar interrogatives, the speaker (a) raises the issue of whether Lisa won by placing {p,¬p}^↓ on the table; thus steering the conversation towards a resolution, a state where either Lisa won or Lisa didn’t win is common ground, and (b) expresses the (trivial) commitment that the actual world is an element of p⋃¬p (the set of all possible worlds, W). Additionally, by virtue of the special discourse effects, the speaker also signals that they have some evidence for p, to which they assign at most low credence.

4.2 Clausal embedding in inquisitive semantics

A consequence of this model is that declarative clausal complements denote sets of propositions, rather than propositions (Hamblin 1971). In terms of the composition with the matrix predicate, I follow Theiler et al. (2018; 2019), who argue in the framework of inquisitive semantics that clausal complements are uniformly of type ⟨st,t⟩ and that clause-embedding predicates take arguments of this type. The composition of a complex sentence like Mary believes that Anna won is illustrated in (45)–(46) (for discussion, see Theiler et al. 2019: 97–100):²⁷

4.3 Proposal for embedded V2

In Section 2, we discussed three uses of EV2-clauses: (i) assertive discourse moves which are fully anchored to the current discourse context, including the speaker, (ii) assertive discourse moves where the context update of adding the issue of p to the table is anchored to the current discourse context, but where the commitment to p is shifted to a contextually specified commitment anchor, and (iii) inquisitive discourse moves which function as a type of biased question, anchored to the speaker and the current discourse context.

4.4 Inquisitive embedded V2: discourse effects vs. semantic composition

On Farkas & Roelofsen’s (2017) account, rising declaratives start out as sentence radicals denoting a singleton set {p}^↓, but take on the meaning of a question, {p, ¬p}^↓, when they combine with the clause type marker open (43). However, if this shift is represented at LF, that means that the embedded clause in (53) will be an inquisitive proposition {p, ¬p}^↓ when it combines with the matrix verb (hear). This is not a problem for the compositional semantics as such; as shown in (55), hear is compatible with both declarative and interrogative complements.

Intuitively, however, the interpretation of the matrix sentence in (53) is that the speaker has heard that p, not whether p, thus suggesting that the EV2-clause denotes a non-inquisitive proposition {p}^↓ when it combines with the matrix verb, while still triggering the global discourse effect of a question with respect to the issue it represents.

According to Farkas & Roelofsen (2017), this problem is not expected to arise. They suggest, based on (56), that rising declaratives and tag-interrogatives cannot be embedded, stating that “even if we consider a sentence involving a plain declarative complement clause, like [John thinks that Amalia left], the commitment that the embedded declarative would induce, if uttered in isolation, does not need to be taken into account in determining the discourse effects of the entire sentence. All we need is the propositional content of the embedded clause” (p. 244).

However, as the embedded rising declarative in (24)/(53) shows us, rising vs. falling intonation can have these kinds of discourse effects also in embedded contexts. Moreover, we have seen that for EV2 in general, the discourse effects associated with the embedded V2-clause do matter for the discourse effects of the sentence as a whole.

The example in (53) thus highlights what is actually a more general issue for this analysis, which is that at some point, the embedded content needs to become accessible independently to the matrix predicate (for the semantic composition, potentially as an informative proposition) and to the discourse level (to have discourse effects at the level of the conversational context, potentially as an inquisitive proposition). This could potentially be achieved via some type of discourse-level operator, which would combine with the sentence radical denoted by the embedded clause and trigger the appropriate discourse-level effects (e.g. those of a biased question), while at the same time allowing the sentence radical to combine with the matrix clause (as an informative proposition).

Regarding the judgement in (56), a note is also in order. While the judgement in (56) is clear, the overall picture painted in this paper is that main clause phenomena are complex: their acceptability depend on a multitude of factors, including the particular properties of the matrix clause (rather than just the type of verb), together with the precise discourse move represented by the embedded clause. H&T, who discuss tag-questions (56-a) as one of their original main clause phenomena, note that they can in fact be embedded in appropriate circumstances (just like rising declaratives, as we have seen in this paper), including under verbs which don’t generally embed questions, like suppose:

The contrast between (57-b) and (57-c) is consistent with the empirical picture observed here, whereby the availability of embedded (assertive and inquisitive) discourse moves across embedding contexts depends crucially on the particular discourse effects associated with the embedded clause, together with highly specific properties of the matrix clause. I return to the question of the role of the matrix clause in the following section.