2.1 Like as a hedging particle

The expression like in English presents a constellation of uses and functions, which have been categorized through a number of taxonomies (see, in particular, D’Arcy (Reference D’Arcy2005) for extended discussion). In the present paper, we focus on the discourse particle use of like, whose pragmatic function has been informally described in terms of hedging (Schourup Reference Schourup1985, Jucker & Smith Reference Jucker, Smith, Jucker and Ziv1998, Siegel Reference Siegel2002, Sharifian & Malcom Reference Sharifian and Malcom2003, Dinkin Reference Dinkin2016, Dinkin & Maddeaux Reference Dinkin and Maddeaux2017).

Intuitively, in all of the examples above, like does not add to the propositional content of the utterance. Consistent with the typical behavior of discourse particles, it instead modulates an aspect of the relationship between the speaker and the proposition, in this case signaling that the speaker has some sort of weakened degree of commitment towards the assertion. It has been suggested that like ‘is used to express a possible unspecified minor nonequivalence of what is said and what is meant’ (Schourup Reference Schourup1985: 42), ‘indicates that the closeness of fit between the utterance and the thought it represents is looser than the hearer may otherwise have expected’ (Jucker & Smith Reference Jucker, Smith, Jucker and Ziv1998: 185), and signals ‘that the phrase it is attached to is detached slightly from commitment to a literal reading’ (Dinkin Reference Dinkin2016: 238). As suggested by such paraphrases, the proposed weakening of commitment varies depending on the particular nature of the content.Footnote ^[6]

2.2 Like in mirative contexts

In addition to the contexts above, like is also commonly found in situations in which the speaker seems to find the embedded proposition surprising or unexpected, similarly to what happens in mirative constructions (see Section 1). Examples of such scenarios, which to the best of our knowledge have not yet been described in the literature, can be seen in the naturally occurring sentences reported below, in (4). As can be seen in such cases, like commonly occurs with other indicators of surprise (e.g. exclamatives such as Whoa! in (4)); in other cases, it is, however, found on its own, as in (5).Footnote ^[7]

Intuitively, the use of like in the above contexts signals that the following facts are somewhat surprising or odd: that the rapper Lil Wayne is smart (4a), that the speaker’s former friend is now rich (4b), that the speaker won again (4c). In (5a), the surprise stems from the fact that even five-month-old chips taste good (5a), in (5b) that a guitar player’s hand moves quickly while he’s playing or in (5c) that someone who was formerly a janitor is now a millionaire.Footnote ^[14] Notably, the contribution of like closely resembles the one attributed to mirative evidentials in the literature as summarized by Rett & Murray (Reference Rett, Murray and Snider2013: 457), where these forms have been described as signalling a lack of ‘psychological preparation’ on the part of the speaker (DeLancey Reference DeLancey1997: 35) or as marking ‘a more or less spontaneous reaction to a new, salient, often surprising event’ (Aikhenvald Reference Aikhenvald2004: 197).

A crucial property of like, moreover, is that the effect of surprise is crucially tied to the hearer, and not just to the speaker. In other words, for like to be felicitous, it is not sufficient that the speaker finds the proposition surprising; it must be the case that the speaker believes that the hearer will also find p surprising. This is indicated by three observations. First, the use of like is not felicitous in a situation in which the speaker is indeed surprised, but already knows that the hearer does not find p surprising. For example, let us suppose that John has long been telling Sue that Bill has become rich, but that Sue for some reason has always refused to believe him. Let us now imagine that Bill pulls up in a fancy car in front of them, showing that John was right after all. While it is felicitous for Sue to convey her surprise via an exclamative, it would be odd for her to do so with like.

Furthermore, similarly to mirative constructions, this use of like is generally constrained by what Rett & Murray call ‘the recency restriction’ in their work on evidentials: the explicit marking of surprise needs to be made within a reasonably short time after the content of the proposition has been comprehended by the interlocutors. The example below, modified from Rett & Murray (Reference Rett, Murray and Snider2013), shows that exclamative intonation and like both share this property. If surprise is expressed at a later stage, as in (7b), the use of a mirative marker is infelicitous.

The example above might suggest that both like and exclamatives behave in the same fashion with respect to this restriction. Upon further examination, however, it can be noted that in the case of like, the recency restriction essentially applies to the hearer rather than the speaker. This is supported by the observation that the particle can still be felicitous when the restriction is violated on the part of the speaker, as long as it still holds from the hearer’s perspective – for example, in a context in which the speaker has long known that p, but has reason to believe that the information is nevertheless new and surprising for the interlocutor. In contrast, this is not the case for exclamatives, which are degraded in this context.Footnote ^[15]

Finally, the hearer-oriented nature of like is shown by the fact that the use of the particle is odd in contexts without any addressee, contrary to other markers of surprise.

As will be discussed in Section 4, the apparent hearer-orientedness of the surprise effects conveyed by like will motivate an analysis that treats the particle as operating over the Common Ground of the participants, i.e. a conversational space shared by the interlocutors.Footnote ^[16] Before proceeding any further, however, let us consider a possible objection to the claim that like operates as a mirative marker. The skeptical reader might point out that in the examples above, intonation seems to be doing a lot of the work to convey the speaker’s surprise. Like might therefore simply be filling a prosodic pause linked to the speaker’s unpreparedness to learn the content of the proposition, without providing any independent semantic or pragmatic contribution. We argue against this view on the grounds of two pieces of evidence. First, the prosodic contour of the examples with like is distinct from the contour of a typical exclamative declarative; we discuss the prosodic behavior of like in more detail in Section 3.2. Second, while simply eliminating like from the same examples does indeed convey a mild sense of surprise, it does not achieve the same effect of bewilderment that the inclusion of like does.

The fact that like occurs in these contexts is puzzling. Contrary to the cases discussed in the previous section, none of the contexts above seem to suggest that the speaker is less than fully committed to the assertion. More specifically, the attested co-presence of other mirative markers – e.g. exclamative intonation in (4b), or markers of full speaker commitment such as totally in (4c) – indicates at least impressionistically that the speaker does in fact thoroughly endorse the assertion. This raises the issue as to whether (and how) the use of like in (10) relates to the hedging uses presented in the previous section. Before addressing this issue, we first offer and discuss a series of diagnostics that illuminate the different pragmatic and distributional properties of the hedging and mirative uses.

3.1 Hedging and mirative uses: both are non-at-issue

While they appear to contribute different effects, both mirative and hedging like share two important properties. First, neither of them is part of the at-issue content of the utterance, i.e. neither contributes to the proposition that represents the ‘main point’ of what the interlocutors are addressing in the discourse (Tonhauser et al. Reference Tonhauser, Beaver, Roberts and Simons2013). This property of like is revealed by two diagnostics. First, both hedging and mirative like fail to interact with logical operators such as negation or modals, similarly to what has been observed for presuppositions and conventional implicatures (Potts Reference Potts2005, among others). This property is shown in the examples below: while like can occur to the right of negation (11) or a modal (12) in surface linear order, its contribution always ‘escapes’ the scope of these modifiers, suggesting that the particle is encoded on a different level from the rest of the proposition.Footnote ^[17]

Second, like cannot be directly agreed or disagreed with by the interlocutor with responses that deny (or affirm) the truth of the proposition. Instead, like can only be challenged with the use of constructions that call into question the more general felicity conditions of the utterance, such as the widely discussed Hey, wait a minute! response (henceforth HWAM; see Shanon Reference Shanon1976).

Having ascertained that both hedging and mirative like are not part of the at-issue content, we now move on to show that these two uses behave differently according to a variety of criteria.

3.2 Teasing apart hedging and mirative uses

The first set of diagnostics concerns the prosodic properties of like. On the one hand, hedging like does not present a specific intonational profile. Siegel (Reference Siegel2002) observes that it can be surrounded by pauses, which surface as parenthetical commas in the written transcription on a par with appositives. However, it is also possible to find cases in which hedging like is prosodically integrated with its surrounding material, as shown by the many examples attested in the literature and on the web that are written without any punctuation. By contrast, mirative like is necessarily followed by a longer pause, represented (henceforth) with ellipses (which are also commonly found in online uses, as well). As a result, while a hedging interpretation is normally available when like is prosodically integrated, a mirative reading becomes unavailable if there is no pause following like, as shown by the examples below.

The following prosodic contours extracted from Praat illustrate this difference. The utterances come from a native speaker of American English as part of a conversation found in the Lambada transcript of the Santa Barbara Corpus of Spoken American English (SBCSAE) (Du Bois et al. Reference Du Bois, Chafe, Meyer and Thompson2000). The two utterances are taken from a single speaker, and are found in the same section of the transcript, in which the speaker is telling a story.

The use of like in (16a) is a clear instance of like as a hedging marker, in which Miles gives a rough time estimate. In (16b), however, Miles expresses the surprise one might have at the reported turn of events. In the graph below in (17), we show the pitch contours for uses of hedging like. Example (17) shows the integrated prosody of like, where there is no major pause preceding or following the particle.

This type of contour differs from the pitch behavior of mirative like, which is characterized by both a long pause following like and a unique prosodic signature on the predicate, all over each other. This pause is also indicated with ellipses in the written transcript above.Footnote ^[19]

A second difference concerns the compatibility of like with other modifiers. Mirative like is perfectly felicitous, and in fact widely attested, with markers that indicate full commitment to the proposition on the part of the speaker, such as totally or definitely. Hedging like, on the other hand, cannot co-occur with such markers. This is shown in (4c) above, reproduced below in (19b), where totally is part of the original sentence. This stands in contrast to examples such as (19a), modified here to illustrate the infelicity of such modifiers in hedging uses.

A third difference between hedging and mirative uses surfaces in contexts in which like is embedded under the matrix subject of a reportative predicate, such as say in (20). In hedging uses, the effects of like can be ascribed either to the speaker or to the matrix subject, whereas in mirative uses, the surprise effect contributed by like must exclusively be ascribed to the speaker, and not to the matrix subject.Footnote ^[20]

A fourth difference is that while both mirative and hedging like are part of the non-at-issue content, they interact with other components of the utterance’s meaning in distinct ways. For example, Siegel (Reference Siegel2002) observes that hedging like can have an effect on the truth conditions of the sentence, at least indirectly. This is shown in the dialogues in (21). Here, speaker A is objecting to speaker B’s denial by countering that her use of like as a hedge makes her first utterance true, contrary to what the hearer suggests (21a). The same maneuver is, however, not possible for mirative uses, where the use of like cannot be used as evidence for contesting the truth-value judgment of the hearer (21b).

Mirative like, moreover, contrary to its hedging counterpart, interacts with the illocutionary mood of the utterance, i.e. with the properties that pertain to the speech act that the speaker is producing. First, mirative like is constrained by choice of speech act in that it is restricted to assertions, appearing to be infelicitous in non-assertive moves such as the posing of constituent questions or the use of imperatives. Below, (22a) cannot be interpreted as an order in which the speaker is surprised at the possibility that the addressee bring a knife; (22b) similarly cannot be interpreted as a question asked by a speaker who is surprised at whoever the person bringing the knife could be.Footnote ^[21]

In contrast, hedging like presents no such restrictions, and can seamlessly operate under the scope of questions or imperatives.

Second, mirative and hedging like engender different types of unacceptability when their content is overtly denied in the continuation of the utterance. This diagnostic was first utilized by Murray (Reference Murray2010) to highlight a difference between evidentials receiving an indirect versus mirative interpretation in Cheyenne. Murray notes that while explicitly denying the contribution of the former gives rise to a logical contradiction – similarly to what happens when one asserts that it is simultaneously raining and not raining – denying the contribution of the latter instead engenders an effect resemblant of Moore’s Paradox, where the oddness is rooted in a violation of the felicity conditions of the assertion, rather than the logical relation between parts of its content. While the judgments are admittedly subtle, the two uses of like seem to pattern in exactly the same way: while denying hedging like has a contradiction-like effect, denying mirative like gives rise to an instance of Moore’s Paradox.

3.3 Interim Summary

In the previous section, we have compared the grammatical properties of the hedging and mirative uses of like. While both uses are encoded as part of the non-at-issue content, they present distinct prosodic profiles, and behave differently with respect to a variety of compositional diagnostics. Table 1 summarizes them.

Table 1 Hedging versus mirative like.

What this discussion suggests is that mirative occurrences of like appear to be treated differently by the grammar, suggesting that they ought to be considered as a genuinely distinct usage from the hedging one. On these grounds, we can now move on to address the following question: what is the common core shared by the two uses? That is, what is it about the underlying meaning of the discourse particle like that leads to the emergence of these two (seemingly) unrelated pragmatic effects? We begin with the hedging use and then move on to the mirative.

4.1 Precision, halos and widening: an informal characterization

It has been observed that we are often imprecise in the way that we communicate (Lasersohn Reference Lasersohn1999). For example, if a driver is driving at a speed of 69 mph, we will normally accept as true a statement that the driver is proceeding at a rate of 70 mph. Strictly speaking, such a statement is not true, but it is nevertheless acceptable. By the same token, if ten people out of three million are awake in the city of Chicago, we are unlikely to take issue with the statement that ‘everyone in Chicago is asleep’, even though this would again not be a true description of the current state of affairs.

The amenability of (certain) expressions to be interpreted imprecisely has been captured by the suggestion that such expressions come equipped with a pragmatic halo – a set of objects of the same denotation type, which differ only in ‘pragmatically ignorable’ ways (Lasersohn Reference Lasersohn1999). Crucially, the size of the halo is determined contextually (Lasersohn Reference Lasersohn1999). In particular situations, e.g. talking sports at a bar, we might be more willing to apply a larger margin of tolerance to the interpretation of an expression than in others, e.g. discussing the set-up of a chemistry experiment. Informally, we propose that hedging like functions as a halo widener. More specifically, the particle signals that the expression it modifies comes with a larger pragmatic halo than the one that would normally be tolerated in that context. This, in turn, has the effect of broadening the set of the denotations of that expression that may be considered.

4.2 Expanding halos: a semantics for hedging like

Various authors have offered different perspectives on the proper treatment of pragmatic halos and their size (i.a. Krifka Reference Krifka, Krämer, Bouma and Zwarts2006, Lauer Reference Lauer, Guevara, Chernilovskaya and Nouwen2012, Sassoon & Zevakhina Reference Sassoon, Zevakhina and Chereches2012). In the present paper, we follow Morzycki’s (Reference Morzycki2011) proposal to recast halos as a set of denotations that bear a contextually determined degree of resemblance to the denotation of the linguistic form they apply to. The notation we make use of henceforth does not hinge on a particular conceptual stance regarding the notion of imprecision; rather, it is adopted to highlight the parallel between the contribution of like in hedging and mirative contexts, as we now discuss in the remainder of the paper.

The particular model we adopt captures imprecision through the proposal that the interpretation of a linguistic expression is constituted by a set of alternatives that includes the semantic value of the expression itself, in addition to objects of the same denotation type that bear (at least) a minimum degree of resemblance to the original denotation. The notion of resemblance is modeled via a cross-categorial ‘approximateness’ relation ${\approx}$ , which holds between two objects if they are similar to at least degree d in context C, where d is a real number consisting of a value between 1 and 0.

Crucially, different contexts impose different similarity orderings as well as different standards of required similarity. The higher the minimum degree of resemblance is, the fewer alternatives qualify as similar and the smaller the halo is, and vice versa: the lower the degree is, the more same-type denotations will be admissible as legitimate interpretations of the original one and the larger the halo is. To see how this works in practice, let us consider the expression ‘$20’ in (27).

Let us now consider three different contexts. In Context 1, (27) is uttered by a college student during a conversation at a bar, a scenario in which a relatively large amount of deviation from the literal meaning is tolerated (d set at 0.7). In Context 2, (27) is uttered by a frequent customer of the store to a person who has asked about the price of the shoes, a situation in which a higher degree of precision can be expected (d set at 0.9); finally, in Context 3, (27) is uttered by a shop attendant in response to an inquiry by a customer, a situation in which virtually no deviation from the literal meaning is to be expected (d set at 1, the maximum). The resulting interpretation can be captured as follows.

We are now in the position of characterizing the semantic contribution of hedging like. We suggest that the modifier widens the size of the halo of a linguistic expression, signaling that a less stringent standard of precision ought to be adopted in interpreting the expression in the context. More specifically, modifying (the halo of) an expression $\unicode[STIX]{x1D6FC}$ with like amounts to fixing the degree of similarity required for a same-type semantic value $\unicode[STIX]{x1D6FD}$ to be be part of the halo of $\unicode[STIX]{x1D6FC}$ to a lower value ( $d^{\prime }$ , below) than the one imposed by the context (i.e. d). The effect is that of expanding the set of admissible interpretations that the target expression can receive in the communicative situation in which like is used. To state this meaning more formally, we compare the set of admissible values for $\unicode[STIX]{x1D6FC}$ and its halo with and without modification via like in (29). Crucially, in (29b), the halo of $\unicode[STIX]{x1D6FC}$ is parameterized to  $d^{\prime }$ , a degree that is lower than  $d$ .Footnote ^[22]

We can see how this would work with the contexts above: for each of them, the effect of like would be to relax the minimum degree of similarity required for computing the halo of the expression.

Note that imprecision is not necessarily rooted in cardinalities or amounts – the observed effect in expressions such as ‘$20’ is only one instantiation of how halos, and modifiers such as like, operate. In the case of ‘European’ above, for instance, paraphrases for like such as approximately or about do not appear to be accurate. However, the underspecified nature of halos correctly predicts that, in the right context, imprecision could be observed for virtually any type of linguistic expression, as long as it is possible to compute a similarity ordering between its semantic value and some salient same-type alternative set. One such ordering could involve predicates denoting the property of being from a non-European country, where lower degrees would instead incrementally admit as legitimate alternatives the denotation of predicates picking out the property of coming from other countries, where the lower the degree of precision is, the greater the tolerated distance of the country from Europe is.

5.1 Preliminaries: context set, plausibility, assertions

We begin by introducing the basic ingredients of the analysis. Despite our relentless quest for knowledge, we never truly have a full picture of the state of the current world. For example, at the moment of writing, we are not in the position of knowing whether our best friend is still at their office, whether it’s raining in Chicago, etc. What we can do, however, is to entertain different hypotheses about how the world could be with respect to these issues, and progressively discard them as we learn more information – for example, if we find out that our friend is on vacation, then we can rule out the hypothesis that they are working. The upshot is that while our knowledge will never allow us to ‘identify a single world as the actual world’ (Pearson Reference Pearsonto appear), it can help us to establish at each moment in time those worlds that can be considered viable candidates for the actual world as well as those worlds that can be ruled out. In this view, conversation can be seen as a collective endeavor to pool our resources en route to narrowing down our set of candidates: each conversational move can be seen as a step towards collectively learning more about the state of the world and, at the same time, discarding alternatives that are no longer compatible with what we know.

In technical terms, we follow Stalnaker (Reference Stalnaker1978, Reference Stalnaker2002) in representing each conversational state in terms of the Common Ground (henceforth CG), a notion that helps us to characterize two aspects that are central to communication. On the one hand, the CG represents what the participants already know at a certain point in the exchange. From this perspective, the CG can be seen as the repository of those propositions that are mutually taken to be true by all conversational participants – in Stalnaker’s terminology, the presuppositions of the speakers. On the other hand, the CG allows the speakers to keep track of what worlds are still possible candidates. Informally, such worlds are those that are compatible with what the interlocutors know about the current world. More formally, such a set is obtained by taking the intersection of all the sets of worlds representing the propositions contained in the CG. We follow Stalnaker in calling this set the Context Set of the conversation (henceforth CS), that is, the set of worlds that are recognized by speakers as ‘live options’ for representing the current world. In sum, we have the following.

As we will discuss shortly, we suggest that mirative like precisely intervenes on the structure of the CS. We assume that each CS is bound by a pragmatic restriction that prevents outlandish-though-compatible worlds from being taken into consideration as candidates, and we argue that like serves as an invitation to the hearer to relax such a restriction, effectively re-admitting such worlds into the CS. Before explaining how like interacts with this process, we discuss the nature of the CS more closely.

As a first step, we treat the CS as a set of doxastic alternatives, each of which represents a possible world that is still in contention for being the current world on the basis of what the conversational participants believe to be true. We call this set $\text{CS}_{w,G}$ , where G represents the group of participants and w represents the actual world.

Note that this representation mirrors Hintikka’s (Reference Hintikka, Davis and Hockney1969) and Pearson’s (Reference Pearsonto appear) representation of the individual doxastic states of each discourse participant, which contains the set of candidates that a particular individual considers to be in contention. Along these lines, we essentially treat the CS as a collective doxastic state – that is, as a set of doxastic alternatives that need to be compatible with the shared beliefs of all of the participants, rather than with the beliefs of a single speaker. The second step towards understanding how mirative like operates – and, perhaps, the most important conceptual move of our analysis – is the following: doxastic states, including CSs, are pragmatically restricted, just as are the interpretations of linguistic expressions. Specifically, we tend to exclude from our doxastic states outlandish worlds, that is, worlds that are too distant from the current one, even if they are in principle compatible with what we know. To see a concrete example, consider again the following proposition.

Let us imagine that the CG contains only worlds in which the propositions above are true: the person in question comes from a low-income family; they were very unsuccessful at school; they had a merely average work ethic; and they were rather unambitious. Crucially, no such worlds are incompatible with worlds in which such a person is now rich. They could have won the lottery, or suddenly have had a brilliant idea. However, while these possibilities cannot be excluded, they are, at best, highly remote. Given what we know about this person, it is exponentially more likely that the actual world will turn out to be one in which such a person is not rich. This makes the worlds in which the person is rich so outlandish that, for pragmatic purposes, we can rule them out from our CS. In other words, even if we do not have any information on this person’s current income, we can proceed under the assumption that this person is not rich, purposefully ignoring possibilities in which they actually are. Evidence supporting the idea that outlandish worlds are routinely ignored when we engage in conversation comes from the domain of modals, that is, expressions that operate by quantifying over possible worlds (Kratzer Reference Kratzer, von Stechow and Wunderlich1991). Consider the following example, from Klecha (Reference Klecha2014).

In the exchange above, Alice is – strictly speaking – correct. There is one possible world in which one could avoid getting wet by wrapping themselves up in duct tape, making Bryan’s use of have to too strong. Yet, Bryan is still using language in a felicitous way, to the point that Alice’s reply is likely to come across as unnecessarily pedantic. While not being categorically ruled out by what we know about the current world, worlds in which people cover their skin with duct tape to fight the rain are so unlikely that they routinely escape the modal base of have to, qualifying Bryan as a savvy speaker. Note that, crucially, different contexts might impose different standards on how outlandish a world must be to be ignored. Let us consider the following two contexts, also from Klecha (Reference Klecha2014): a science olympiad, where teams compete to solve engineering problems, and a Rube Goldberg device olympiad, which has the same rules but encourages participants to solve their problems in creative, roundabout ways.

In the first context, the option of building a Rube Goldberg device, while not impossible, requires such a high amount of time and procedural complexity that it can be safely ruled out as an unreasonable possibility, much like the option of covering one’s body with duct tape in (34) above, leading us to judge the sentence in (35a) as true – construction of the bridge is the right thing to do. In the second context, however, the use of complicated devices is the defining trait of the competition. As such, worlds in which we build one of them are no longer outlandish, but fall squarely within the domain on which the modal operates, engendering the intuition that the sentence in (35b) – according to which the obvious solution is to build a bridge – is false. Through this exercise, it is possible to see that the pragmatic practice of excluding outlandish worlds is not just a minor detail about the conversational setting. Rather, it is deeply ingrained in our way of interpreting and processing meaning, to the point that it affects our judgments about a sentence containing an operator that quantifies over worlds.

Going back to the main issue under discussion, it therefore seems reasonable to posit that plausibility-based restrictions on possible worlds should not just be taken into account to analyze the semantics of modal operators, but should also be incorporated into our understanding of how we reason about possible worlds more generally. This, crucially, also includes the process whereby we compute candidates for the actual world based on the information that we have in our CG. We thus propose to enrich the notion of a CS by suggesting that, for a group of participants  $G$ , CS includes worlds that are not only compatible with what the speakers know/believe, but that are also reasonable.

To model this second property, we make use of the following ingredients. First, following Klecha (Reference Klecha2014), we propose to measure the outlandishness of a world by means of ST, an operator that applies to two worlds v and w and returns the degree of stereotypicality of v given what we know in w relative to a context C.

Second, we enrich the meaning of ST(v)(w) with a parameter $\unicode[STIX]{x1D703}$ , representing the minimum threshold of stereotypicality that a world must have (with respect to the evaluation world) to count as plausible.Footnote ^[23] To have access to such parameters, we assume that the CS itself is parameterized not just to a group of participants and a world of evaluation, but also to a threshold, thus providing the required elements to assess the reasonability of a world. With these tools, we are ready to formalize the notion of reasonable CS informally sketched out above. $\text{CS}_{G,w}^{\unicode[STIX]{x1D703}}$ will contain those worlds that are compatible with what the speakers mutually believe (per the definition of CS), as well as those that are greater than or equal in plausibility to the threshold $\unicode[STIX]{x1D703}$ in C.Footnote ^[24] In more formal terms, for a world v and an evaluation world w:

We argue that, by means of using like, the speaker signals to the hearer that the CS should be expanded to include worlds that, due to their outlandishness, were previously excluded from contention. As we discuss below, this move serves as a strategy from the speaker to facilitate acceptance of their assertion in contexts in which the assertion is especially likely to be rejected – that is, in contexts in which all $p$ -worlds are highly implausible, and thus excluded from the set of worlds under consideration to begin with.

5.2 The problem of updating with outlandish worlds

To see how this contribution can be modeled, we begin by reviewing the process through which we update our pool of candidate worlds. First, we follow Stalnaker and much of the subsequent literature in assuming that conversation is aimed at narrowing down the CS, so as to inch closer towards a representation of the current world. This goal is pursued by means of uttering linguistic assertions. Specifically, every time that we accept a proposition asserted by our interlocutor, we eliminate from the CS those worlds that are not compatible with the proposition – that is, the worlds in which the proposition is false – via set intersection. The CG resulting from an accepted assertion will be one in which the assertion has become a presupposition; the ensuing CS will be one in which only the worlds in which the asserted proposition is true are preserved, while the others are ruled out. To see how this process works, consider first a simple assertion, such as (39).

Furthermore, let us imagine that there are four possible worlds: two in which p is true, w11 and w22; and two in which p is false, w33 and w44.

Let us now imagine a two-party conversation between Sue and John, taking place in an actual world w1 in which p is reasonably plausible. For instance, such a world is one in which the speakers know that the person in question comes from an already well-off family, showed strong motivation and displayed a rare talent since the early stages of their education. In such a world, all possible worlds have relatively high stereotypicality values. On the one hand, it is reasonable that the person is now rich, given the advantageous circumstances; on the other hand, it is likewise reasonable that the person is not rich, given the fact that motivation and talent, while conducive to financial success, are by no means sufficient to attain it. As such, for a plausibility threshold set at 0.1, they all comfortably make the cut to be in the CS of the conversation involving the group of participants  $G$ (i.e. Sue and John).

Let us now imagine that Sue learns that the person in question is now indeed rich, and wants to share this information with John, her interlocutor. Following the standard Stalnakerian view, the update procedure proceeds as follows. First, Sue asserts that p; second, if John accepts the assertion – or, to put it more precisely, unless John has any explicit objection to it – the CG is updated by intersecting the worlds in the CS with the worlds in which p is true. This process narrows down the CS, maintaining in it only the worlds in which the asserted proposition is true and eliminating those in which it is false. Example (42) provides a step-by-step breakdown of the process.

Let us contrast this situation to the actual world w2 described in the previous Section 5.1, where both Sue and John believe that the possibility of the person being rich, given the circumstances, is very remote. Here, all worlds in which the person is rich have very low stereotypicality value. Hence, for $\unicode[STIX]{x1D703}$ set at 0.1, the CS of a conversation between Sue and John contains only $\neg p$ worlds.

What happens if, in w2, Sue learns that the friend wins the lottery and intends to assert it, so as to share this news with John? Since there are no p-worlds in the CS, intersection with p worlds would lead to the empty set, that is, to an inconsistent CG.

Needless to say, such an effect would be highly disruptive for the collective endeavor in which the conversation participants are engaging, jeopardizing their epistemic quest for identifying the actual world; in terms of discourse, indeed, a conversational state with an inconsistent CG is in crisis (Farkas & Bruce Reference Farkas and Bruce2010). Such a state of affairs is identical to a state resulting from asserting a proposition that contradicts what is in the CG, an act that Stalnaker labels as self-defeating (Stalnaker Reference Stalnaker1978: p. 44). Should we then conclude that the stereotypicality restrictions on the CS make it impossible for us to assert, and therefore to turn into common knowledge, a proposition that is only true in implausible worlds? This seems to be too strong a constraint. First, the worlds in which the friend is rich, though outlandish, are after all compatible with the actual world, and are therefore not impossible. In addition, it is well known that pragmatic restrictions excluding outlandish worlds are defeasible (Klecha Reference Klecha2014): they can be lifted, slackened or tightened by the interlocutors throughout the conversation. As such, what needs to be done to resolve the issue is to first re-admit outlandish worlds into the CS. Once this is done, it will then be possible to eliminate the $\neg p$ worlds and, eventually, add the proposition to the CG via a regular update operation. We suggest that mirative like precisely serves the purpose of facilitating this operation.

5.3 Mirative Like: addressing scrutiny, expanding context sets

While the listener always has the possibility of autonomously considering remote worlds, this is not guaranteed to happen. In fact, assertions proposing counter-expectational updates are very likely to undergo special scrutiny before p becomes common knowledge. In particular, it has been observed that a natural reaction to such proposals is a ‘double-checking’ move: a response whereby the addressee explicitly asks the speaker to confirm the appropriateness of adding p to the CG, deferring any decision on the acceptance of the proposal until they receive such a confirmation. Typical examples of such double-checking reactions are really, did that really happen and similar expressions (see Romero & Han Reference Romero and Han2004 for extensive discussion of double-checking moves).

We suggest that like, by lowering the stereotypicality restrictions on the CS, serves as an explicit marker to win over the hearer’s potential skepticism and to facilitate acceptance of the proposal without further scrutiny. Specifically, we argue that the use of like facilitates the update in two ways. On a compositional level, it expands the pool of candidate worlds. As such, it creates the conditions for the hearer to accept the assertion and successfully eliminate the $\neg$ p worlds, while reducing the risk of leading the conversation into a state of inconsistency (see (44) above). On a pragmatic level, it indirectly signals that the speaker themselves acknowledges the outlandish nature of the asserted content – if this were not the case, there would have been no need to use like in the first place. This acknowledgement serves as a further attestation to the speaker’s cooperativeness, showing that they took the listener’s perspective into consideration, and that they are willing to go our of their way to make sure that the assertion enriches the CG – the main goal of any conversational exchange.

We implement this idea in the following way. To begin with, we follow Krifka (Reference Krifka2001) in the view that speech acts are functions that take a proposition and an input conversational state as argument, and return an output conversational state. Against this background, we assume that assertions can be represented through a multi-layered structure, which minimally encodes two components: the proposition p and the illocutionary content, i.e. the effect that the author of the speech act aims to obtain on the discourse state. In the standard Stalnakerian view, the illocutionary content of an assertion consists of a proposal to add p to the CG.

Following the procedure described above, this result amounts to generating an output CS of the conversation by intersecting the candidate worlds in the input CS with the worlds in which p is true. Following the notation introduced above, $G$ represents the group of participants, w the world in which the conversation is taking place and $\unicode[STIX]{x1D703}$ the degree of stereotypicality of the CS. To facilitate readability, we omit from the representation of the CS the condition that the candidate worlds must be compatible with what we already know, i.e. (i) in (38). We assume that the conditions remain in place, of course.

We suggest that, by using like, the speaker specifies that the p-update operation ought to be carried out not with respect to the Input CS, but to CS+, a widened CS whose threshold of stereotypicality is lower than the one of the Input CS – that is, the one that had been in place up until that moment in the conversation. To see how this is implemented in the dynamics of assertion, let us first compare the internal structures of CS and CS+. The two sets are identical, with the exception that the threshold for filtering out outlandish worlds in CS+ is  $\unicode[STIX]{x1D703}^{\prime }$ , that is, a lower one than the one in the original set. This ensures that CS+ is a superset of CS; it contains all the worlds that were already candidates before, plus some outlandish-but-CG-compatible candidates that were previously excluded.

We propose that assertions modified by like differ from regular assertions by making use of CS+ in the illocutionary proposal, as opposed to CS.

In this view, mirative like indirectly modifies the illocutionary force of an assertion. While the essential effect of the assertion remains the same – namely, adding p to the CG – the proposal put forward by an assertion modified by the particle operates over a different input, one that includes (at least some) outlandish worlds as candidates. This captures the desired effect: by expanding the CS, the use of like ensures that at least some of the worlds in which p is true are also part of the pool of candidates for the current world. Crucially, this reduces the risk that the update of the assertion leads to an inconsistent CG, putting the listener in a better condition to accept the proposal. Note that use of like, however, does not guarantee that the update will actually go through. Following the standard view of assertion, whether the proposal is accepted remains out of the speaker’s hands and is ultimately up to the addressee’s discretion. Use of the particle when uttering an assertion true in outlandish worlds, however, will increase the likelihood for this to happen.Footnote ^[25]

Before returning to the comparison between hedging and mirative like, two observations are in order. First, the association between like and surprise emerges as a side effect. Like does not mark surprise per se; if the speaker is, however, accompanying an assertion with a signal that remote worlds should now be considered, it follows that those worlds are indeed those in which the asserted proposition is true, triggering the inference that p is unexpected. In this view, the contribution of mirative like crucially differs from the one of markers that directly convey the speaker’s feeling of bewilderment towards the content of what they are asserting, such as exclamative intonation and markers such as wow! (Rett Reference Rett2011); at the same time, the mirative effect of like arises in a similar way to what has been claimed by Zanuttini & Portner (Reference Zanuttini and Portner2003)’s theory of Wh-exclamatives in English, where surprise and unexpectedness have also been modeled in terms of a domain-widening operation from canonical to less canonical scenarios (see Zanuttini & Portner Reference Zanuttini and Portner2003: Section 4.2 for details). Second, by targeting a shared space in the conversation – as opposed to a private one – the effect of particle crucially involves both interlocutors, and not just the speaker. This correctly captures the observation that the felicitous use of the particle requires that the asserted content must be hard to believe for the addressee as well, and not just the speaker. Again, this seems to make like different from mirative markers that are more inherently anchored to the speaker’s perspective, such as exclamatives (see (8)–(9) above).

We now return to the central theme of the article, and discuss how the proposed analysis helps to make sense of the differences between hedging and mirative uses of like.

5.4 Capturing the differences

As can be recalled from the previous discussion, mirative like can be distinguished from the hedging version by virtue of (i) being compatible with markers of maximal certainty, (ii) failing to impact the truth-conditions of the assertion, (iii) being unavailable in unbiased questions and command imperatives, and (iv) resisting shifting when embedded under the subjects of attitude or reportative verbs.

First, compatibility with markers of epistemic confidence stems from the fact that mirative like is used by the speaker to enhance the possibility that their proposed update is accepted by their interlocutor. As such, any assertion in which the particle is used shares with regular assertions the felicity condition that the speaker individually believes that p – that is, that the proposition is true in all the available doxastic alternatives in the speaker’s epistemic state (Hintikka Reference Hintikka, Davis and Hockney1969, Searle Reference Searle1969). As such, use of mirative like is perfectly compatible with markers that explicitly signal epistemic certainty towards the proposition. By contrast, hedging like, by widening the allowable pragmatic halo of an element within the asserted content, gives rise to a weaker assertion than what would have been produced without like. As such, even though hedging like does not directly lower the speaker’s commitment, adding a marker of confidence is inconsistent with the weakening effect associated with widening the halo, thus giving rise to a pragmatically incongruous behavior. In light of this, it is also possible to understand why hedging but not mirative like has an effect on the truth conditions of the utterance.

The halo-widening contribution of hedging like crucially affects which individuals fall into the extension of a predicate. For example, the fact that a wider range of prices can be accepted for interpreting ‘$20’ will ultimately impact the process whereby we assign truth to ‘The shoes cost $20.’ While this contribution does not emerge by directly modifying the descriptive content, it effectively makes the truth-conditions less stringent.Footnote ^[26] This does not happen with mirative like. Because the particle targets the illocutionary content of an assertion, it is inert with respect to the logical content of what is being asserted, thus failing to impact the truth-conditions of the statement.

The limitation of mirative like to assertions follows from the fact that the particle is used to facilitate the addition of a proposition to the CG; such a specification cannot be made in speech acts in which no update is proposed to begin with, constraining the use of like accordingly. Concerning questions, the illocutionary content does not encode a proposal; rather, it presents the listeners with two alternative routes through which the CG could be updated, and towards which the speaker has no particular commitment (see Farkas & Bruce (Reference Farkas and Bruce2010) for further discussion). Concerning imperatives, these speech acts aim at bringing about changes in the actual world, rather than attaining its correct representation. As such, the felicity conditions of these utterances are related to notions such as authority, performativity and preferences (see Condoravdi & Lauer Reference Condoravdi, Lauer and Ingo2011), and have little to do with the status of the proposition in the picture of the world shared by the interlocutors, making the contribution of like non-congruent. In contrast, because hedging like targets the linguistic interpretation of individual predicates, its use is insensitive to the specific type of speech act used in the context, as well as the pre-conditions of such a speech act. As a result, the use of hedging like is felicitous whenever the semantic interpretation of the content of the utterance is important, i.e. in virtually any type of utterance that makes use of natural language.

Finally, the strong tendency of mirative like to resist embedding is linked to its status as a speech act modifier. By modifying the input of the proposal made by the assertion, its contribution is inherently anchored to the participants in the here-and-now of the conversation. This also applies to situations in which the content of the assertion features other doxastic agents that in principle could serve as the anchor but that are not involved in producing the very utterance in which like is used. This is not the case for hedging like, however. Because this use of the particle conveys a metalinguistic commentary that is independent of the properties of the speech act in which it occurs, it can shift under any agent that could in principle produce such a commentary: the speaker, which is always an option, or other subjects involved in reportative or belief events.

6.1 Like: the common core behind hedging and mirative effects

While the hedging and mirative functions of like appear at first sight to be unrelated, the analysis outlined above suggests that they are in fact linked to the same core operator. In both uses, like relaxes a context-sensitive pragmatic restriction that determines the cutoff point for what elements are part of a pragmatic halo and a doxastic state, respectively. The contextual restrictions that the presence of like manipulates in both uses respond, moreover, to very similar pragmatic demands. Assuming a certain amount of deviation from the truth conditions of an expression and its actual interpretation allows us to describe the world in a perspicuous way, sparing us the burden of providing unnecessarily fine-grained details; similarly, ruling out outlandish worlds allows us to work with fewer candidates in our quest for achieving a representation of the actual world, sparing us ‘the cognitive difficulty of processing unexpected/non-stereotypical propositions’ (Klecha Reference Klecha2014: 144).

If the two uses fundamentally bring about the same type of manipulation over very similar pragmatic restrictions, how can we explain the fact that, as discussed in Section 2, only hedging like contributes a weakening effect, while mirative like does not? We suggest that the difference is not grounded so much in the contribution of like as it is in the distinct properties of the different semantic objects that like operates over. More specifically, we suggest that effects of weakening/strengthening are ultimately determined by the differential interaction between the size of the set as well as the pragmatic strength associated with pragmatic halos and context sets. In the case of halos, the larger the set of admissible alternatives to an expression is, the larger the amount of deviation from the expression’s literal interpretation is. This in turn dilutes the strength of the assertion: because more possibilities that are compatible with the speaker’s communicative intention remain open, the assertion will allow us to learn less about the state of the world than its like-free counterpart. The situation is different with context sets. Here, consideration of non-stereotypical worlds is a pre-condition for accepting an assertion that instead leads us to learn a lot about the world. For example, learning that a person that we did not expect to have financial success is actually rich allows us to eliminate many candidates for the actual world, leading us to gain more knowledge than we would have gained had we learned that this person, as expected, was not rich. Thus follows the intuition that assertions with mirative like, contrary to those modified by the hedging variant, are not weak; they are, in fact, quite informative.

The proposed analysis leaves one question open: how does the hedging/mirative polysemy of like relate to the other pragmatic and syntactic functions that this form can have in English? As discussed in Section 2.1, the empirical picture appears to be especially complex. D’Arcy (Reference D’Arcy2005) has individuated as many as nine separate functions, which, aside from the discourse particle use, include the following uses (among others).

In light of this rich constellation of uses and contributions, the question arises as to whether the analysis outlined above, or at least the common core that links the hedging and mirative contributions, can illuminate whether a similar semantic affinity connects the other uses as well. While an exhaustive answer would go well beyond the scope of this paper, we provisionally note that the notion of relaxing a standard of similarity between two linguistic expressions and two worlds – the essential commonality shared by hedging and mirativity – seems to be potentially relevant to at least the conjunction (49a), the suffix (49b) and the quotative uses (49c). The crucial notion tying all of these uses together is the fact that like is placing two distinct entities in a relationship of similarity to one another. All of these intuitively involve a comparable similarity relationship between two close-though-not-identical objects of the same linguistic type, such as two individuals for the conjunction use, two adjectives for the suffixal use or two speech events for the quotative use (see Davidson (Reference Davidson2015) for an extended semantic analysis of this use). Further illustrating the fact that like can be used to relate two objects as being ‘similar enough’ is the use of like in so-called similatives.

In Rett’s (Reference Rett2013) treatment of similatives, she argues that the two types of dancing here can be related along a variety of similarity criteria. Perhaps the most relevant similarity in an example such as (50) can be thought of as a similarity of manner, e.g. both Mary and John use their arms a lot while dancing. What all of these uses of like share is the fact that they place two objects in a relationship of rough similarity with one another. In the case of conjunction, the embedded proposition is in a relation of similarity to what Mary feels, and represents, moreover, that she is not quite sure about her success. In the case of its suffixal use, like indicates that the doll in (49b) is similar to a child, but is not quite one. In the case of the quotative complementizer, the direct quotation in (49c) places the content of Mary’s utterance in a relationship with Mary’s state. Finally, in the case of the similative in (50), the presence of like equates some manner of Mary’s dancing with one of John’s. Whether the proposed formalization of such a similarity standard in terms of a context-sensitive numerical parameter is adequate for all of these uses remains to be seen. However, it is at the very least encouraging to observe that, despite their important different syntactic and pragmatic properties, these uses of like could also lend themselves to a partially unified semantic analysis.

6.2 Mirativity: the cross-linguistic picture

Stepping back to the broader picture, an outstanding issue concerns the relationship between like and other dependent manifestations of mirativity cross-linguistically, in particular with the widely attested cases of evidentials. Needless to say, addressing such a puzzle in a comprehensive fashion would require a detailed comparative analysis of like and the known cases of mirative evidentials, which would extend well beyond the scope of the present paper; we nevertheless find it worthwhile to make several preliminary observations, focusing on the following question: what semantic/pragmatic property(ies) construe(s) like and narrative/indirect evidentials as a suitable natural class for the expression of speaker’s surprise?

Among the vast literature on evidentiality, we would like to mention two accounts aiming at connecting the encoding of indirect evidence and mirativity. In Rett & Murray’s (Reference Rett, Murray and Snider2013) work on Cheyenne, the link is modeled in semantic terms. Both indirect and mirative evidentials relate the at-issue proposition p to some contextually salient set E of epistemically accessible propositions. What determines the difference between the two meanings is the temporal relation between the utterance and the event of the speaker learning that p. If the assertion is made within a short period of time after the learning moment, E is valued as the speaker’s own expectations, triggering the mirative reading; if the assertion is made a long time after the learning moment, E is valued as the community’s expectations, triggering the indirect interpretation. It is difficult to directly compare this proposal with the account of the polysemy of like outlined in this paper. In particular, the effect of lowering the required threshold of precision, as done by hedging like, seems to have little to do with the encoding of information, making these analyses difficult to pit against one another. However, there are two ways in which the polysemy featured by like seems to be related to that of the evidentials examined by Rett & Murray. First, the availability of a mirative interpretation is semantically motivated by the logical form of the expression it is parasitic on. Whether it is about valuing a set of propositions, as evidentials do, or a pragmatic restriction over a set, as like does, mirativity arises through a structurally similar mechanism to the one that yielded the other reading. Second, in both accounts, mirativity is modeled as a speech act phenomenon: its contribution is not encoded as part of the propositional content, but pertains to the illocutionary contribution of an assertion. In both cases, this claim is substantiated by similar empirical properties, such as speech-act-level restrictions, Moore’s Paradox effects and resistance to perspective shifts (see Section 3.2). Accordingly, the question follows as to whether the compositional commonalities between like and mirative evidentials point to a more general cross-linguistic property of the expression of surprise, highlighting mirativity as a phenomenon that is inherently encoded as a speech act property, as opposed to other types of non-at-issue meaning. We see this as an important question for cross-linguistic semantics, and for linguistic theory more broadly.

Peterson (Reference Peterson2010), by contrast, suggests that mirative interpretations of evidentials are the result of an implicature, framing mirativity as a pragmatic phenomenon. The implicature arises whenever a speaker utters an assertion with an indirect evidential in a context in which they have direct knowledge of a situation. This move would violate Grice’s Quantity Maxim: since a stronger assertion could have been made – that is, one without an evidential – the speaker is clearly being under-informative. As a way of reconciling this linguistic behavior with cooperativeness, the evidential is re-interpreted as a marker of the speaker’s mental unpreparedness with respect to the proposition, thus imbuing the evidential with a flavor of surprise. Concerning like, it is not possible without detailed diachronic data to determine whether the mirative variant effectively emerged via a similar conversational implicature. While we think that this is a very plausible hypothesis, the fact that it is not possible to cancel the mirative contribution without generating infelicity indicates that, in any event, this effect has now become conventionalized as part of like’s lexical meaning, similarly to the cases discussed in Rett & Murray.

It can furthermore be observed that, from a synchronic perspective, hedging like and indirect evidentials are indeed both associated with speech acts that are crucially weaker than their unmodified counterparts: while asserting p, the speaker leaves open the possibility that things might be otherwise, either due to lack of direct evidence or by signaling that a strict interpretation of the sentence’s content might not apply. In this respect, what both hedging and indirect evidentiality share is that they leave room for a $\neg p$ option that would have been instead unavailable – or at least much more backgrounded – in the case of non-hedged statements or assertions backed up by direct evidence. Notably, an underlying $\neg p$ option is also present in the expression of surprise. In particular, it has been suggested that mental states of surprise arise from a contrast between the expectation that the proposition is false (hence $\neg p$ ) and the observation that it is actually true (hence p), which, likewise, contributes to making $\neg p$ salient. Giannakidou (Reference Giannakidou, Blaszack, Giannakidou, Klimek-Jankowska and Mygdalski2015) formalizes this intuition by proposing that, if a speaker s is surprised that p, then she must have believed that $\neg p$ , at a time $t^{\prime }$ prior to the time of utterance (see Giorgi & Pianesi (Reference Giorgi and Pianesi1997) and Giannakidou & Mari (Reference Giannakidou, Mari, Nadine, Berezovskaya and Schölle2016) for different proposals).Footnote ^[27] The emerging picture is one in which the availability of a $\neg p$ alternative provides a conceptual bridge between hedging like and indirect evidentials, on the one hand, and the effects of speaker’s surprise that these expressions can convey, on the other. More specifically, a hypothesis following from this observation is that constructions that independently leave room for $\neg p$ worlds are suitable linguistic forms to express the category of mirativity; the specific semantic/pragmatic mechanism through which surprise is expressed, however, will ultimately depend on the particular type of semantic contribution whereby each form makes $\neg p$ available. In this view, it is remarkable that a similar connection between $\neg$ p and mirativity seems to be at work for other expressions beyond evidentials and like. For example, the subjunctive mood in Italian is licensed either in situations characterized by a lack of commitment to the truth of p or under the scope of predicates that presuppose commitment to p, but express emotion and surprise (Giannakidou & Mari Reference Giannakidou, Mari, Nadine, Berezovskaya and Schölle2016). Such examples provide encouraging, if preliminary, evidence that dependent manifestations of mirativity might be found across many more linguistic domains than evidentiality, pointing to the expression of surprise as a phenomenon that encodes a general semantic/pragmatic core, but which nonetheless surfaces in different ways depending on the specific nature of the linguistic form that conveys it.