Additive particles are elements whose semantics is usually defined with reference to an associate: a constituent of the host sentence of the particle that is distinguished in some way, often by prosodic means (which makes such elements belong to the more general class of focus-sensitive particles). Additive particles convey that their associate is not the only element to be subject to its predication, i.e. that there is at least one distinct alternative element for which the predication of the associate also stands (cf. Gazdar 1979; Krifka 1999 or Kripke 2009 among many others). The adverb too in English is a prototypical example of an additive particle. In (1-a), the semantic contribution of too is the presupposition (1-b), assuming that too associates with the noun John (indicated in small caps).

A (simplified) description of this particle is given by Krifka (1999) as in (2), where F stands for the (focalized) associate and P is the predicate which applies to F.2

Thus, in this analysis, additive particles assert the content of their host sentence and presuppose their additive component.4 They are also anaphoric, in that they require the identification of the alternative to their associate in order to be felicitously used.

English too does not impose any constraint on the scalarity of its associate. Other additive particles such as even have been analyzed as scalar additive particles, i.e. particles that do impose constraints on the scalar properties of their associate. Typically, the adverb even is given a representation as in (3) which crucially involves a constraint on the scale of the likelihood of propositions (again referring to, a.o., Krifka 1999).5

Applied to (4-a), this yields the presupposition in (4-b) (with small caps indicating the associate of even).

Another property of even that has been the subject of extensive discussion is its interaction with negation (cf. among many others Karttunen & Peters 1979; Wilkinson 1996 and the literature they generated). One of the key questions in the debate is whether it is necessary to postulate distinct entries for a negative polarity and a positive polarity even, or whether a unique semantic entry can account for all the readings of even, including those in negative sentences.

From a more general perspective, it has been argued that the behavior of additives under negation, the nature of the scale used by scalar additives and the position of their associate on this scale are parameters that differentiate scalar additives both within a given language and also across languages, thus defining the “landscape of additive particles” (Giannakidou 2007; Lee & Pan 2010). These parameters will be relevant in the description of Cantonese additive particles in the section that follows.

In Cantonese, additive particles form a varied set that minimally contains:

The precise description of each of these particles is difficult for at least two reasons. First, some of these elements have additional uses besides their additive ones. For example, dou1 can be used to convey universal quantification, so that (5) is potentially ambiguous between (5-a) and (5-b).

A satisfactory analysis of dou1 means that we have to decide whether it is necessary to assume two distinct meaning postulates for dou1, or an abstract one from which one could derive both (5-a) and (5-b). We will not address this question here, nor will we attempt at a precise description of the whole system of Cantonese additives.

Second, these elements are seldom used on their own, but are usually used in combination with other additive particles. Different additive particles belong to different parts of speech and occupy different syntactic positions, which makes it natural to use several of them in a single sentence as in (6) which involves three distinct additive elements: zung6, maai4 and tim1.

As a coarse way to measure the rate of co-occurrence between additive particles, we observed the frequencies of the additive particles mentioned above in the Hong Kong Cantonese Corpus (Luke & Wong 2015). We compared their total frequencies of occurrence with their frequencies when used in combination with at least one other particle in the set. The counts are raw: sentences were not analyzed one by one to filter out potentially irrelevant occurrences. The results are summarized in Table 1.

Apart from dou1, these results confirm that additive particles have a tendency to appear in combination with at least one other particle in the studied set.6 In the case of tim1, most of the co-occurrences are with zung6 (17 cases), but dou1 is also present (11 cases). The rest of the cases do not involve an additive considered in Table 1, or any additive at all.

This makes it difficult to tease apart the precise contribution of each particle: when particles are used in combination their meanings combine or unify in a way that is not trivial to decompose. Out of the particles in Table 1 only lin4 appears consistent in its interpretation. It always denotes a scalar additive akin to even and which scopes over noun phrases (Matthews & Yip 2010). The other ones all seem to allow both scalar and non-scalar interpretations. For example depending on its associate, dou1 will be interpreted in a non-scalar way (7-a) or scalar way (7-b) (reflected in the choice of translation by too or even).

This is not to say that dou1 is necessarily ambiguous. The scalar reading might just be a by-product of the topicalization of the object (see Section 6.2 for a discussion in a cross-linguistic perspective).

The high frequency with which these particles occur together led some authors to analyze some of these combinations as instances of particular constructions. For example, Law (1990) argues that zung6…tim1 is a discontinuous construction (a claim largely found in the subsequent literature dealing with the syntax of SFPs). Law argues that zung6 is optional in such constructions and when it is optionally deleted, the meaning of the construction is unaffected. Tang (2007) also points out that zung6 and tim1 are semantically redundant when used in the same clause, and that is one of the reasons why he considers zung6…tim1 as a discontinuous construction.

The reason we mention that frequent co-occurrence of particles is because it affects the way examples are constructed, especially for the particle that is the focus of this paper. In most cases, the most natural examples are produced by using a combination of particles rather than one in isolation, which makes it hard to pinpoint the contribution of the targeted element. The examples in this paper have been constructed with this in mind, eschewing particle combinations whenever possible for reasons of simplicity, while trying to keep examples as natural as possible.

Mirativity is defined as the grammatical marking of surprising or unexpected information either on the part of the speaker, or possibly of other discourse participants (Delancey 1997; Aikhenvald 2012). Mirativity is often discussed in tandem with the notion of evidentiality, i.e. the indication of the source of knowledge. This is because both notions have some overlap, and languages that do mark mirativity have a tendency to mark evidentiality as well (Aikhenvald 2012). Nevertheless, mirative markers are considered to form a category that is distinct from evidentials. One of the most convincing arguments for that conclusion is that there exists mirative markers which are compatible with different evidential values, and which can be used in combination with different evidential markers (Aikhenvald 2012; Delancey 2012).

Aikhenvald (2015) notes that Cantonese is one of the very few isolating languages that have an evidential system (albeit a simple one) represented by the hearsay sentence final particle wo5. Besides this evidential particle, it has been argued that Cantonese also has a mirative marker: the particle wo3 which marks information that is unexpected for the speaker (Luke 1990; Matthews 1998; Hara & McCready 2017).7 Leung (2009) mentions that wo3 was initially used as an hearsay evidential, and only later became a mirative particle, in line with the general semantic change paths from evidentials to miratives discussed by Aikhenvald (2012).

Here, we will argue that Cantonese has more than one mirative marker, i.e. that the sentence final particle tim1 can also mark mirativity, although of a different type. We turn to the specifics of that particle in the coming section.

The Cantonese morpheme tim1 is usually considered to have two distinct usages: as a verb meaning to add which is restricted to specific objects such as drink or rice, and as a sentence final particle which is the one that we will focus on.

As an SFP, tim1 appears at the end of utterances, possibly with other SFPs, forming clusters. The order of particles in SFP clusters is not free. It has been observed that when an utterance involves a cluster of SFPs, tim1 appears as the first element of that cluster (Matthews & Yip 2010), suggesting that it occupies a syntactic position that is different from other SFPs (Law 2001; 2002). This will be discussed further in Section 3.2 where it is shown that this description is not entirely accurate.

As mentioned above, the semantic contribution of tim1 ranges between a non-scalar additive meaning to a purely mirative meaning, which we illustrate below.

In (8), the use of tim1 appears to be purely non-scalar and indicates that the addressee had at least one other bowl of rice before, without any overtone that would convey that the bowl of rice is (for example) an unexpected thing to have.

In (9), tim1 can be interpreted as a non-scalar additive, i.e. as indicating that Bob knows other languages besides Portuguese, with no apparent ranking of the languages. A scalar interpretation is also possible, though not mandatory, e.g. if it is forced by context, by using a specific intonation on pou4man2 ‘Portuguese’, or by combining tim1 with an optional scalar additive like zung6.8

As expected, the additive usages of tim1 in (8) and (9) involve the identification of an associate, i.e. tim1 appears to be a focus particle in the sense of Krifka (1999). This associate can be of various types: a noun phrase as in (9), or a predicate as in (10).

The only constraint that seems to bear on the nature of the associate is that it cannot be the subject of the host sentence of tim1, cf. (11).9

Besides these additive readings, the contribution of tim1 can also mark a form of surprise of the speaker. Thus in (12) the speaker is surprised at the fact that it’s raining, i.e. they did not expect it to be the case.

The use of tim1 in (12) is intuitively different from the additive readings. For a start, it does not seem to require an antecedent of some sort for its interpretation, i.e. it is not anaphoric as other additive particles are. In both (8) and (9), the interpretation of tim1 requires the identification of a contextually accessible element that belongs to the set of alternatives to the associate. In (12), this does not seem to apply, and (12) can felicitously be uttered as a discourse initial segment.

In Section 3, we present more systematic evidence of the differences between the additive and mirative tim1, but the main puzzle about tim1 can already be seen from the data presented so far. Given the variety of meanings that can be conveyed by the use of tim1, it is fair to ask whether it is possible to give a single unified description to the particle. If it is an additive particle, its meaning need to be characterized more finely (is it scalar? what kind of scales does it associate with? etc.), and its anaphoric or not properties need to be accounted for. First, we turn to previous analyses of this particle.

Lee & Pan (2010) (L&P) propose a single unified analysis for tim1. They argue that the additive and mirative readings are derived from one and the same lexical entry, and that their semantic differences arise as the result of using different scales and associating with elements of different semantic types. Beyond their description of tim1, their work is also an attempt to further make precise the cross-linguistic typology of scalar additives by showing that tim1 occupies a unique position in this landscape.

Their main claim is that tim1 has to be considered as a scalar additive (cf. Section 1.1). They claim that, like other elements of its class, its meaning involves:

They further claim that tim1 is distinct from other scalar additives (both in Cantonese and cross-linguistically) for three reasons. First, it does not constrain the type of scale it associates with (e.g. unlike Cantonese lin4 or English even which are restricted to the scale of (un)likelihood). Second, it does not constrain the position of its associate on the scale (e.g. unlike even in English or oute and akomi ke in Greek, cf. Giannakidou 2007), meaning that the associate of tim1 is not necessarily the low extremum of the scale it belongs to. Third, it is not sensitive to polarity (again unlike even).

L&P consider three kinds of scales that tim1 can rely on: degree scales, for example as expressed by gradable predicates, the scale of (un)likelihood, and what they call the “quantity” scale. The quantity scale is invoked for the seemingly non-scalar uses of tim1. In an intuitive way, the resort to this scale indicates that the cardinality of some contextually salient set is increased. The type of scale that can be used by tim1 depends on the semantic type of its associate. All types can rely on the quantity scale, which corresponds to a default when no other scale is accessible. Besides this, propositions (i.e. elements of type t) can use the (un)likelihood scale and elements that denote entities can also rely on the degree scale. We discuss and illustrate each of these scales below, using (13) for the degree scale, (12) for the unlikelihood one, and (9) for the quantity scale.

In (13), the associate is a gradable predicate which is associated with the scale of smartness.

In (9), the associate is the NP pou4man2 and the scale is the quantity scale.

Note that if there is a contextually accessible scale on which Portuguese ranks higher than some other language (e.g. in terms of prestige), then that scale can be used for the interpretation of tim1. In that case the meaning of tim1 would better be paraphrased by even and its semantics would be similar to those of (13). In (13), L&P claim that the resort to the quantity scale is blocked by the fact that the nature of the associate primes the access to a degree scale.

In (12), the associate is the whole sentence (type t) and the scale is that of unexpectedness (triggered by the type t).

The L&P analysis is attractive in that it is economical because it proposes a unified analysis to tim1, and because it makes explicit the factors that trigger the different readings associated with tim1. We also agree with many of its features, notably their separation of the existential and scalar component. However, it suffers from a number of problems which we present in this section, most of which are not specific to their analysis, but problematic for many approaches to additives.

This section reports the results of an experiment which aimed at testing whether there are acoustic differences between the mirative and additive tim1. The experiment was prompted by the intuition that the additive tim1 is longer than the mirative tim1, i.e. that in examples like (9) the duration of the particle is longer than in examples like (12).

This was tested in a production experiment. 11 participants (7 female) were recruited. All were students at the Hong Kong Institute of Education, aged between 19 and 25 years old (mean = 22), and were native speakers of Cantonese with no declared hearing problems. Participants listened to an audio stimulus in the form of a dialogue that provided a context for a sentence they were then asked to read out loud in what seemed the most natural sounding way to them.

The experiment tested one single binary condition: the additive or mirative reading of tim1. Target items all involved tim1 in sentence final position in the sentence to be read out loud by the participant. Two lists were produced which used identical target tim1-sentences, but used different contexts for the target items, triggering either a mirative or additive reading of the particle. In addition, filler sentences were also included in the experiment. The naturality of the sequences of utterances was checked with various native speakers beforehand. They reported no significant problems. Participants were debriefed and did not report any naturality issues either.

An example of item and additive and mirative contexts is as follows:

A total of 12 different target items and 36 fillers were produced, and each participant produced 6 target items after being exposed to an additive context and 6 in the mirative condition. The items were shown on a computer screen, using a latin-square design and pseudo randomization (using the IbexFarm platform, available at http://spellout.net/ibexfarm/).

Only the target utterances were analyzed (12 per participant). For each produced utterance, the /t^him/ segment was analyzed on Praat (Boersma & Weenink 2016) and the duration of the rime of /t^him/ was measured (i.e. vowel + coda).13 The results are illustrated on Figure 1.

The effects of the tested condition were assessed using model reduction and maximum likelihood ratio test and treating participants and items as random factors. The results show there is a significant effect of the condition leading to longer production of the rime component of /t^him/ in the additive contexts (χ² = 15.753, p-value = 7.219e-05).

Therefore the experiment confirmed the initial intuition. There is a significant difference in the production of the additive and mirative tim1 in sentence final position: the mirative tim1 is significantly shorter than the additive one, suggesting they might be distinct elements. In Section 6, we argue that mirative tim1 evolved from additive tim1, and use the observed shortening as one supporting observation.

Besides their semantic and acoustic differences, the additive and mirative tim1 also have differences in their syntactic distribution.

In Section 1.4 we mentioned the common observation that tim1 always appears first in clusters of SFPs (Matthews & Yip 2010).14 However, a closer look shows that tim1 can appear in at least two different positions in SFP clusters. In (25) tim1 is cluster-initial, but in (26) it is in cluster-final position.

Such observations are consistent with data given by Kwok (1984: 45) when discussing the combination of tim1 with the particle ge3 (which she analyzes as a specific construction, a claim we will not discuss here) as in (27).

The additive tim1 can thus appear in two different positions (cf. (25) and (27)), but the mirative tim1 can only appear in the final-cluster position, i.e. it is not possible to interpret tim1 in (25) as a marker of surprise (this was confirmed by several native speakers).

An additional observation is that the additive tim1 can felicitously be used in interrogatives, whereas the mirative cannot, cf. (21) vs. (28).

The data in this section shows that, depending on its interpretation, tim1 may or may not appear at the initial of an SFP cluster, and is or is not compatible with certain constructions (minimally: interrogatives). This naturally suggests that it should be possible to use tim1 twice in the same utterance, once in each of these two positions and with possibly two distinct readings. That prediction is actually borne out, as shown in (29) which is a modification of (26) which involves another occurrence of tim1 as an additive marker:

The possibility of using the same item twice in a sentence with different meanings is a typical test to show that an expression is ambiguous (Gillon 1990), which further strengthens our claim that there are two distinct uses of the particle tim1.

There is little room for doubting that the additive reading tim1 is indeed additive. Not only is it etymologically related to a verb meaning to add, but virtually all its descriptions in the literature describe it as an additive.

However, given the data covered so far, especially in the discussion of the L&P analysis, it is fair to ask whether the additive tim1 is non-scalar, in the manner of English too, scalar, or possibly ambiguous between the two readings. Examples such as (19) do not carry scalar overtones unless supplementary markers or context force a scalar reading (e.g. zung6 as in (9)). The remaining cases fall in two categories. On one hand are those that involve a reading of unexpectedness, and we argue that those belong to the mirative tim1, not the additive one. On the other hand are cases that involve an associate which is inherently scalar. It is those cases that we will rely on to show that the additive tim1 is indeed scalar, in line with the spirit of the L&P proposal. To illustrate we will use (30).

Example (30) is acceptable. However, if we minimally modify it by changing the quantity of books the speaker read yesterday, its acceptability degrades: if the quantity is equal to the number of books read today the sentence is borderline acceptable (31-a), and if it is inferior it becomes plainly degraded (31-b).

Note that this behavior crucially differs from the case of non-scalar additives like English too which are not sensitive to the quantity denoted by its associate:

Therefore, we take those examples as indicative that the additive tim1 does encode a form of scalar constraint.15

The case of (31-a) deserves a bit more attention. In Section 1.4 when discussing example (19) we mentioned that tim1 does not necessarily carry scalar overtones, i.e. it does not entail that its associate is strictly higher than its alternative on some scale. The examples above showed that tim1 does not allow the associate to be lower on the scale. But what about equality? Although the acceptability of (31-a) is disputable, other examples show that tim1 is tolerant of equality. For example, one can modify (9) as in (33-b).

If tim1 required strict superiority of some sort, as English even does, then (33-b) should sound degraded in the same way (34) does. This is because knowing Putonghua is not really surprising for A-Mei (assuming her to be a Hong Konger), or at least it should be less surprising than knowing French or Portuguese.

What appears in (33-b) is that all elements have equal status regarding the speaker’s point, i.e. that A-Mei knows a lot of language. In Section 2.2 we showed that the scale of unlikelihood was not an appropriate scale to work with tim1, echoing other work about scalar particles like even (Anscombre 1973; Kay 1990; Greenberg 2016). We will argue that the only scale that is relevant for the interpretation of tim1 is an argumentative scale. The constraint encoded by the additive tim1 is a scalar constraint of equality or superiority of its associate. In Section 5 we give a formal treatment of argumentation in terms of the relevance of a proposition to the goal defended by the speaker. This will naturally handle cases like (34) as well as other examples we have seen so far.

When discussing mirativity, Aikhenvald (2012) makes it clear that:

She proposes to distinguish five flavors of mirativity, each of which can be defined with respect to the speaker, the addressee or the main character in a narrative.

In order to characterize the mirative component of tim1, we will compare it to the wo particles described by Matthews (1998) as mirative elements. Matthews mentions two mirative wo in Cantonese: wo3, described as a marker of “noteworthiness” (Luke 1990), and wo4, which indicates “unexpectedness” (Kwok 1984). Matthews analyze the latter as a derived form which combines wo3 and a falling intonation toneme that strengthens the illocutionary force of wo3 (Law 1990), which is why we will focus on wo3 alone here.

The range of uses of wo3 is rather wide, covering reporting and story-telling, informing, reminding, suggesting, challenging positions, indicating noteworthiness, etc. (Luke 1990). Its core meaning was recently formalized by Hara & McCready (2017) as an expressive particle which indicates that its prejacent is either unexpected or corresponds to an unexpected discourse move. They define unexpectedness as the negation of the consequent of a normality conditional, i.e. a conditional which specifies conclusions that can be drawn under normal circumstances (Reiter 1980).

The following monologue shows a typical use of wo3 as a marker of unexpectedness. In the context, the speaker is talking about her child.

Hara and McCready analyze the example by considering that since both parents are not liars, a reasonable (but defeasible) assumption is to assume that their child is not a liar either, which explains the use of wo3 to mark that the information in the prejacent goes against what is normally expected. This analysis puts wo3 in the “counterexpectation” flavor of mirativity recognized by Aikhenvald.

In contrast with the mirative aspect of wo3, tim1 rather marks an aspect of surprise for the speaker and sudden discovery or sudden revelation or realization. This can be seen in the following minimal pair, adapted from (Matthews 1998).

In (36-a) the information is not new to the speaker, but the speaker indicates that the reported fact is unexpected in the light of other information they know about. In (36-b), the speaker necessarily just found out about the information reported (whether directly or by some form of inference) and expresses their surprise about it. Unlike (36-a), (36-b) is not compatible with a context in which the speaker already knew about the contents of the utterance and wants to report its unlikely character as in (35-b).

Another possible value that the mirative tim1 can convey is related to what Aikhenvald (2004; 2012) call deferred realization i.e. a situation “where-by the speaker gives a post-factum interpretation to what they may have observed in some way.” Aikhenvald mentions that this value is found in different languages (such as Quechua or Western Apache) and that “deferred realization is an integral part of mirative meanings in all systems where mirativity is associated with inference.” Deferred realization is what is at hand in (26). In that example, the speaker is of course aware of what they eat at night. The surprise element comes from the realization of the potential consequences of their eating habits, i.e. of the existence of an inferential link between eating at night and putting on weight.

Besides surprise, the mirative tim1 is often accompanied by a negative overtone. For example, many speakers interpret (37) as marking that the wedding proposal is both an unexpected and rather unfortunate event, most likely because the speaker has no intention of getting married.

However that negative overtone does not seem to be part of the core meaning of the mirative tim1. For example, a sentence like (37) can easily be continued with a statement of the speaker saying how happy she is without sounding contradictory. We will therefore not integrate this reading into our analysis and consider that the negative overtone is a conversational implicature derived from the mirative reading. One way to derive the implicature is to consider people might generally hold negative attitudes to events they are not prepared for, although this default preference is easily overridden.

Finally, it should be noted that mirative tim1 is in complementary distribution with wo3 which is expected since they denote complementary values of mirativity. The additive tim1 is however compatible with wo3 (cf. (25) above), further strengthening the claim that the two usages of tim1 should be distinguished.

The probabilistic framework we use is based on standard intensional logic to which a probability measure P has been added. Thus, rather than dealing with binary truth-values, we are dealing with probabilities interpreted in a Bayesian way, that is as subjective degrees of belief (Jeffrey 2004). In that setting, belief update is modeled by conditioning: upon learning that a content φ is true, the probability measure P is replaced by a measure P′ such that λx.P′(x) = λx.P(x|φ), where P(x|φ) is the conditional probability of x given φ.16 In other terms: prior beliefs are replaced with posterior beliefs which correspond to the prior beliefs conditionalized by grounding the information that φ is true.

In that probabilistic setting, it is possible to characterize a notion of argumentation which will prove useful for the description of the tim1 particles. We will consider that an utterance of content p is an argument for a conclusion H iff P(H|p) > P(H), i.e. iff learning that p is true increases the belief in the truth of H. This is a probabilistic interpretation of the linguistic notion of argumentation first postulated by Anscombre & Ducrot (1983) and which was notably developed by Merin (1999) to characterize different aspects of communication as well as the semantics of some discourse markers. One of the tenets of argumentation theory is that every utterance in a discourse is oriented towards an argumentative goal, i.e. that in all instances of communication it is assumed that speakers speak to a point (Merin 1999). Sometimes this goal is explicit or contextually obvious, sometimes not. In those latter cases the argumentative properties of the utterances in discourse constrain the possibilities to figure out that goal. In other words, the goal is not a purely contextual element, but one that is semantically constrained by the set of argumentative elements used in an utterance.

With that notion of argumentation it is possible to measure the strength of an argument. This can be done by a variety of means (Merin 1999; van Rooij 2004). A classical way to do it is to consider a relevance function (rel) which is such that p is an argument for H if rel(p, H) >0. The higher rel(p, H) is, the better the argument will be. If rel(p, H) is negative, then p is a counter-argument for H. While there are many definitions of rel (e.g. Good’s weight of evidence, cf. Good 1950), all can be mapped to the quantity P( H|p ) P( H ) \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ {\textstyle{{P\left({H|p} \right)} \over {P\left(H \right)}}} \] \end{document} , called the impact factor, which measures the difference between the prior belief in H and the posterior belief in H, after learning that p is true. The impact factor can directly be rewritten as P( p|H ) P( p ) \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ {\textstyle{{P\left({p|H} \right)} \over {P\left(p \right)}}} \] \end{document} via Bayes’ rule. In that quantity P(p|H) is called the likelihood. This term is used here in a precise way, and is crucially a property depending on two elements: the content asserted by the speaker (p) and the goal H the speaker is aiming at (rather than the asserted content alone). Intuitively, the likelihood component corresponds to the probability of observing what the speaker asserts, if what they argue for is true. This can be illustrated in (38).

In (38), the speaker is arguing that the referent of he is a priest (=H) and uses the fact that he is wearing a cassock (=p) as evidence. The argument is effective because it is very likely for a priest to wear a cassock, i.e. P(p|H) is quite high, and it certainly is higher than the prior probability for an arbitrary person to wear a cassock (P(p)), meaning the argument has a high impact factor, or relevance. This is a property we will use in the following sections.

Linguistic argumentation theory claims that the semantics of some natural language items is intrinsically argumentative. For example, the connective but conveys argumentative opposition (Anscombre & Ducrot 1977; Winterstein 2012b), while operators like only reverse the argumentative orientation of their prejacent (Ducrot 1973; Winterstein 2012a). In line with these claims, we attribute an argumentative component to both versions of tim1.

The argumentative properties of a sentence are usually correlated with their degree properties if they express some, although this is not necessarily the case. For example the quantifiers a few, some, most, all form a quantitative scale (each expresses a quantity that is higher than the one below it) which is usually correlated with argumentative strength. Thus in (39), the higher the quantifier on that scale, the better the argument will be for the explicitly targeted conclusion John knows Deleuze quite well.

Beyond quantifiers, any expression which denotes a degree will usually map onto an argumentative scale (Fauconnier 1975; Ducrot 1980). Thus quantized NP form argumentative scales, as do gradable predicates modified by intensifiers. This all boils down to the fact that we can assume a unique scale for tim1, namely an argumentative scale which is relative to the speaker’s goal. This scale is partly contextual, but is also constrained by linguistic expressions: the isomorphy of degree scales and argumentative scales is one example, but there are plenty other cases which form the core of the works in the framework of argumentation within language (see e.g. Anscombre & Ducrot 1983; Merin 1999 for a wealth of examples and applications).

Finally, from a compositional point of view, argumentative constraints can be likened to expressive meanings conveyed via conventional implicatures (CI), in the sense of Potts (2005) and McCready (2010). This distinguishes these meanings from presuppositions. CI contribute new meaning, but not in an at-issue way, which explains their scopelessness properties and the difficulty to deny them. Presuppositions also share these properties but unlike CI they do not convey new material and can rather be seen as conditions for a felicitous assertion (McCready 2010). In this paper, we will use the framework of the ℒ CI + \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ \mathcal{L}_{\rm{CI}}^{+} \] \end{document} logic (Potts 2005; McCready 2010) to formalize the way the content of the tim1 particles is introduced. This is done for reasons of convenience since that logic is commonly used in theoretical works describing elements that convey CI.

The analysis we propose for the additive tim1 borrows from the argumentative one proposed for French même ‘even’ by Anscombre & Ducrot (1983: 57–67) (A&D) to capture the scalar constraint, and the one for non-scalar additives proposed by Winterstein (2011) to deal with matters of anaphoricity and association. The argumentative description of scalar additive items treats them as imposing that both their prejacent p and antecedent q must be arguments for some conclusion and that their prejacent presents a stronger argument that their antecedent. Formally, we write it as: ∃H : rel(p, H) > rel (q, H)>0.

For matters of anaphoricity, an additive like too can be described as encoding an existential presupposition along with an argumentative constraint which imposes the argumentative similarity of its prejacent and its antecedent (Winterstein 2011) (modulo some substitutions that we eschew here for clarity purposes).

Taking these two approaches together, we formalize the core meaning of additive tim1 as follows. In a nutshell, it indicates that its prejacent is at least as strong an argument as its antecedent for some conclusion. The antecedent is determined by considering the set of accessible alternatives to the associate. We assume the following notations:

We use the notation defined earlier, and analyze tim1 as an element that conveys the truth of its prejacent and introduces an additive presupposition along with a conventional implicature stating that its prejacent is argumentatively stronger than its antecedent. This can be formalized by analyzing tim1 as in (40). We use the ℒ CI + \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ \mathcal{L}_{\rm{CI}}^{+} \] \end{document} framework of McCready (2010) to model these dimensions of meaning, using their ♦ operator to conjoin at-issue and expressive meaning (i.e. we treat it as a mixed expressive).17

As such the representation does not make explicit the binding between the variable y existentially bound in the presupposition and the free variable y that appears in the CI content (so named to make it clear what it should be bound to). We take the latter to be akin to a pronoun which can be dynamically bound to the presuppositional variable (noting that such technical issues are not specific to the case of tim1).

Analyzed as in (40), tim1 leaves the at-issue of its host unchanged, its main function being to introduce the presupposition and the CI.

The alag (p, p′) relation is defined as: ∃H : 0 <rel(p′, H) ⩽ rel(p′, H) i.e. p must be at least as good an argument as its antecedent p′ for some H, corresponding to the argumentative goal targeted by the speaker. As argued above, this scalar relation can be based on the contextual argumentative properties of the elements in play, or based on their linguistic properties (e.g. in the case of quantized NP and gradable predicates).

One thing that stands out in (40) is the different nature of its contribitions: a presupposition and a CI. The status of the former is rather straightforward and in line with much of the existing literature: it has the expected projection properties, and can be bound to a previous statement (and sometimes needs to be, cf. our discussion in Section 2.2.1). The scalar component also projects in the way expected of conventional implicature, i.e. it is not altered by interrogation or an embedding under modals: for example (41-b) conveys that three beers for Bob is argumentatively stronger than the two beers John had (e.g. to argue that there was a lot of drinking going on that night).

Among the differences between presupposition and conventional implicatures are: (i) the fact that CI cannot be bound in the manner of presuppositions, (ii) the usual ineffability of CI compared to presuppositions, (iii) the speaker-oriented nature of CI (Potts 2005; McCready 2010). All these apply to the scalar component of tim1 (and scalar additives in general), and also apply to the component conveyed by mirative tim1 described in the following sections. Example (42-b) shows both the difficulty in providing a proper paraphrase of the scalar component (in the antecedent of the conditional), and to bind the use of tim1 to that expression.

The scalar component of tim1 is also speaker-oriented in that the speaker does not present it as backgrounded and purportedly shared information (though it can be), but takes in charge the scalar relation between prejacent and antecedent. Thus in (43) the speaker conveys, in a non at-issue way, that they believe checking e-mails during class is at least as annoying (probably more) than coming late to class. The two conjuncts could be reversed with the opposite reading.

The description in (40) covers all necessary aspects of the additive tim1: its need for an antecedent (via its presupposition) and its scalar component in the form of an argumentative superiority constraint conveyed as a CI. We illustrate this on two examples we introduced so far. First is (19) (adapted with a proper name).

According to the description of tim1 in (40), tim1 requires two arguments, which are: its associate x (Portuguese), and its scope Q (Bob knows). It then conveys three components. First, it conveys an at-issue meaning corresponding to Q(x) (=Bob speaks Portuguese). Second, it triggers a presupposition according to which Bob knows another language, different from Portuguese. To satisfy this presupposition, the language in question must be salient/identified in the context, i.e. it cannot be accommodated ex nihilo, which entails that by itself (19) cannot be used discourse initially. Third, it conveys a conventional implicature according to which Bob knowing Portuguese is an argument which is at least as strong as its antecedent for the goal the speaker is arguing for.

The question of whether one argument is stronger than another is partly a contextual one. In the case of (19), we can consider at least two different contexts. In the first one, the speaker is trying to convince the addressee that Bob knows a lot of different languages. In that context, all languages have similar argumentative power: asserting that Bob speaks either of them will have the same (positive) relevance to the argumentative goal Bob speaks many languages. However, if speaking Portuguese has some high contextual relevance (e.g. if the goal is Bob will easily settle in Brazil) then that will remain compatible with the semantics in (40).

Beyond such contextual properties, argumentative relations are also partially defined lexically, typically in the case of expressions denoting quantities or degrees which map onto argumentative scales, as shown in Section 5.1 above.

To illustrate, we repeat example (13) and indicate the associate x and scope Q of tim1.

With those elements, tim1 conveys an at-issue meaning according to which Bob is smart to a certain degree (higher than John). In addition, it presupposes that somebody else is smart to a certain degree (present in the context) and conveys a conventional implicature according to which the smartness of Bob is argumentatively at least as strong than John’s. This is immediately verified given that argumentative scales map onto degree scales, i.e. since Bob’s degree of smartness is above John’s, then whatever goal is supported by smartness will be affected more by Bob’s degree than John’s (e.g. their suitability for a task).

We can show that the scalar constraint of tim1 is indeed argumentative by seeing the behavior of tim1 in cases that involve discourse segments which stand in systematic argumentative opposition. A first example of that sort involves the Cantonese adverbs zaang1di1 ‘almost’ and gan2gan2 ‘barely’. For the purpose of the paper we will assume that these adverbs function like their English equivalents. From the argumentative point of view these elements are interesting because even though an expression of the form barely X entails that X is the case, barely will act as an argumentative reversal operator, i.e. the whole expression will argue against whatever X might argue for (Anscombre & Ducrot 1983). Thus, example (44-b) is degraded, even though the speaker conveys that John was on time. This is due to the reversal effect of barely.

Almost is the dual of barely: while it negates its prejacent, it retains its argumentative properties (Anscombre & Ducrot 1983; Jayez & Tovena 2008). Let’s now examine the case of (45).

The example is degraded, but it should be observed that in it, the English paper is already finished, unlike the Chinese paper. Therefore, on the scale of advancement/completion, the host of tim1 offers a property that is higher on some scale than its antecedent. But tim1 does not seem to be able to access that scale and is not felicitous in that example. This is easily explained if we assume that tim1 relies on an argumentative scale and is thus sensitive to the fact that expressions like almost finished and barely finished are in systematic argumentative opposition.

Another example that supports the hypothesis that tim1 relies on an argumentative scale involves the adversative connective daan6hai6 ‘but’ which imposes that its two conjuncts stand in argumentative opposition (Anscombre & Ducrot 1977; Winterstein 2012b). It is easy to check that the use of the additive tim1 is not compatible with the use of an adversative connective (46).

The same constraint explains why (15) is not felicitous.

The reason is that it is difficult to abduce a goal for which both the fact that A-Mei was born in HK and the fact that she studied in Argentina are co-oriented (meaning they mostly argue in favor of opposite sets of conclusion). However, if one gives an explicit goal such as A-Mei has seen a lot of different places in her life, then the acceptability of the example readily improves.

Finally, the argumentative view also help make sense of the counter-examples of Greenberg (2016) discussed in Section 2.2.2:

The explanation of (17) is that a disjunction of the form A or B is not argumentatively co-oriented with either A or B. To see it, one can use the probabilistic interpretation of argumentation and check that generally P(H|A ∨ B) > P(H) does not entail nor is entailed by P(H|A) > P(H), meaning that the expressions are not co-oriented.

L&P strongly argue for a differentiation between even and tim1. We on the other hand gave tim1 a core semantics that is very close to that of even: both have at their core a constraint of argumentative superiority (strict or not), and both rely on one unique scale. It is thus fair to review the arguments of L&P against a conflation of the two elements.

One of their arguments has to deal with the polarity sensitivity of even, and we will not challenge it. It is clear that tim1 is not sensitive to polarity the way even is, and nothing in the semantics we gave to tim1 predicts it should be. A second argument they give is that tim1 can associate with different scales whereas even is restricted to the scale of unlikelihood. We discussed and challenged that assumption. We argue instead that the scalar part of tim1 uniquely relies on an argumentative scale, like even. Unlike even, the scalar constraint is not a strict superiority one. Lastly, L&P argue that, unlike even, tim1 does not mark its associate as the extremum of a scale. They illustrate this property of tim1 by giving example (47) (their (27), including the choice of translation).

Example (47) is indeed acceptable, and can be interpreted as showing that knowing Portuguese is not the maximally unexpected property that can be predicated of the referent. However, if we take this sentence to be a test for the extremum quality of the associate, then even also passes this test, cf. the acceptable (48).

L&P seem to consider that (48) should be out, but all native speakers we consulted confirmed that this is an acceptable discourse. In any case and whatever the actual status of even in (48), the analysis we propose does not mark the associate of tim1 as the extreme point of a scale, so we predict the possibility of a sentence like (47).

For the description of the mirative tim1, we will also use the argumentative framework. In a way its description is less complex than its additive counterpart since it does not involve an associate nor a presupposition.

In a nutshell we propose that mirative tim1 encodes a constraint of high relevance towards the speaker’s goal. We will show how such a constraint accounts for both the surprise and deferred realizations readings of mirative tim1 discussed in Section 4.2. Formally, the proposed content for mirative tim1 is given in (49) (where the typing follows Potts 2005 and McCready 2010).

Where p is the at-issue content of the host of tim1 and highrel(p) = _def ImF(p, H) ≫ 1, where H is the argumentative goal targeted by the speaker. ImF is the impact factor discussed in Section 5.1 (of which relevance is a monotonous increasing function Merin 1999) which is such that: ImF( p, H )= P( H|p ) P( H ) = P( p|H ) P( p ) \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ {\rm ImF}\left({p, H} \right) = {\textstyle{{P\left({H|p} \right)} \over {P\left(H \right)}}} = {\textstyle{{P\left({p|H} \right)} \over {P\left(p \right)}}} \] \end{document} (via Bayes’ rule). In other words, the constraint states that the prejacent of the mirative tim1 must be especially relevant to the speaker’s goal.

How does a piece of evidence yield a high impact factor then? Basically, we can look at the meaning of a high impact factor in any of the two following ways (remembering that we are dealing with probabilities, i.e. quantities in the [0, 1] interval):

What the impact factor tells us is that the effect of p on H is conditioned by both the prior beliefs in p and H and the likelihood component P(p|H). This means that asserting a content p by itself will not guarantee a high impact factor: that component also needs to be a good diagnostic for H. Thus if the belief in p is very low, then the impact factor will be high as long as there is a causal link that relates p and H. This is one area where the argumentative approach differs from most approaches to the scale encoded by scalar additives: the scale is not related solely to a property of the prejacent, but also involves a contextual element: the speaker’s goal (which is otherwise partly constrained by the linguistic structure of the utterance, cf. earlier sections).

What sort of causal link are we talking about here? The simplest way to think about it is along the lines developed by Pearl (2009) about causal Bayesian networks (CBN). A CBN is a directed acyclic graph in which the nodes represent random variables and the links represent the causal dependencies between variables. Links are oriented, and this orientation represents the hypothesis that the root of the link causes a change on the variable at the other end. The strength of the causal links is measured by conditional probabilities. When observing (changing) the value of a variable, the values of the other nodes are modified by updating the joint probability distribution represented by the whole CBN. CBN offer a flexible and compact way to represent probabilistic knowledge bases. They also offer a simple way to illustrate how the use of tim1 can convey its high relevance. Our cases are rather simple in that they involve only two nodes: the goal targeted by the speaker H and the proposition p asserted by the speaker. Here we assume that the causal direction is such that p has an effect of H. This is represented on Figure 2.

We can use Figure 2 to discuss two of the mirative examples we discussed so far. First, we mentioned that (12) conveys a strong surprise of the speaker regarding rain.

We will assume that the goal targeted by the speaker of (12) is something akin to H_bad = Things are bad. The way to understand the effect of this assertion is that the speaker indicates that the value of P(p) is high (near certainty), thus modifying the value of the corresponding node in the CBN. This will then affect the probability of H accordingly. If P(p) was already high, then the revised probability of H will not be altered much, meaning p will have had no impact. This is what explains the reading of unexpectedness of p in (12).

Second, we described (26) as conveying a deferred realization of the speaker, i.e. the realization that (s)he will probably get fat if (s)he keeps on eating night snacks.

Regarding the speaker’s goal, one might argue that (s)he is either targeting something like H_bad above, or the closely related H_fat = I will get fat. We will adopt the latter here, remembering that for many people H_fat entails H_bad.

Here, it is not credible to assume that the speaker had a low prior belief in the content of the assertion since one is supposed to be aware of one’s eating habits. Instead, we argue that what the speaker targets in the CBN is the strength of the link between p and H, i.e. their beliefs about the causal link between the two. An extreme case could be to go from a CBN in which no link exists between the nodes (which amounts to an assumption of causal independence) to one in which such a link exists. Therefore the modification of the CBN in this case is not about the value of the nodes but about the links between nodes, i.e. about causality. In the case of (26) this revision comes as the result of the assertion of the other speaker.

In both cases we analyzed the argumentative component as a conventional implicature, i.e. as new content conveyed by the speaker in a non at-issue way. This means that the speaker uses the particle to indicate the high relevance of its context. Depending on which belief state we can attribute to the speaker, his indication of high relevance might be interpreted as either a signal of his low prior belief, or his signaling that he adjusted his beliefs regarding the causal link at hand.

The constraint of high relevance also highlights a difference between wo3 and tim1. The mirative tim1 can be used with a prejacent whose probability is high (as in (26)) unlike wo3, which requires a low prior of its host (or of the speech act conveyed by the prejacent, see Hara & McCready 2017). Essentially wo3 is not argumentative, even though it is probabilistic in a way. Its semantics are not directly related to the speaker’s goal, unlike that of tim1.

On a final note about the mirative tim1, and as yet another piece of evidence to distinguish it from its additive sibling, we can observe that nothing in the description we have given prevents mirative tim1 from being involved in an adversative construction, unlike the additive tim1. This prediction is borne out, as in (50-b) which shows that a tim1 segment can be introduced with the adversative daan6hai6:

Now that we have given a formal description of the contents of both the additive and the mirative tim1, we can turn to the question of their relationship. As can be seen from (40) and (49), both descriptions have in common that they express that the relevance of the host of tim1 is higher than something: higher than that of tim1’s antecedent for the additive tim1, higher than some threshold in the case of the mirative tim1. The major difference between them is the fact that the mirative tim1 stands alone, and seemingly lost its anaphoric character.

We argue that there has been a semantic change of the additive meaning towards the mirative one. The core semantics of the element has not changed, but it underwent a pragmaticalization of its meaning (Mosegaard-Hansen 2008). Traugott & Dasher (2002) give several criteria that characterize such a change in meaning. The relevant ones for tim1 are the following:

All of these criteria apply to describe a move from the additive to the mirative tim1. First, the mirative tim1 refers only to the speaker’s own (subjective) surprise. The existential presupposition of the additive involves more objective considerations about the presence of an antecedent etc. Second, the additive tim1 refers to the internal structure of its host and is directly linked to its truth-conditional interpretation, again via its existential presupposition. None of this is relevant in the interpretation of the mirative tim1 which merely adds an expressive content indicating aspects to related to the processing of its host (which can be modeled as detailed in the previous section). Finally, unlike the additive, the mirative tim1 scopes over the whole utterance, as indicated by its syntactic position (in line with the hypothesis of Law 2002 about the scope of SFPs in Cantonese). More specifically, the mirative tim1 is only felicitous with assertions, whereas the additive one is for example possible in questions. This can be taken as evidence that the additive tim1 operates below the level of speech acts, unlike the mirative one.

We will therefore hypothesize that the mirative tim1 is a latter innovation based on the additive tim1, which itself was derived from the verb tim1 meaning ‘to add’. The evolution path for the morpheme is given in (51).

Having described the evolution path, we now turn to the mechanism behind the change. We argue that the semantic shift involved is based on metonymy. This is because one typical characteristic of metonymic change is that it presents a continuum between meanings, usually allowing for cases that allow both interpretations simultaneously (Mosegaard-Hansen 2008). The case of tim1 fits that picture quite well: in many instances tim1 occupies a sentence final position and is potentially interpretable as both an additive and mirative element as in (52). Furthermore, its tendency to be phonologically reduced patterns well with other cases of grammaticalization.19

The speaker of (52) can be taken as merely using tim1 as an indication that A-Mei had a bowl of rice in addition to something else. But the utterance can at the same time also be taken to indicate the surprise of the speaker at this fact, and the relevance of it.

Ideally, these claims should be completed with diachronic corpus data showing the evolution of the marker. This is made difficult by the fact that, even though Cantonese has some written tradition (Snow 2004), finding reliable and authentic diachronic data is rather challenging (but see Leung 2006 about the evolution of wo5 and related particles). The most readily available data for contemporary Hong Kong Cantonese is found in two corpora: the Hong Kong Cantonese Corpus (HKCanCorp, Luke & Wong 2015), which contains both casual conversations and radio show transcripts, and the Hong Kong Cantonese Adult Corpus (HKCAC Leung & Law 2001), which contains transcripts of radio shows. For diachronic data, the Corpus of Mid-20th Century Hong Kong Cantonese (Chin 2015) offers transcripts from movies ranging from 1943 to 1970. While these corpora might not be as far away in time as we could wish, we can still observe differences in the distribution of the additive and mirative readings of tim1. We extracted all occurrences of the particle in these corpora and classified their use as either additive or mirative. The classification was done manually by taking into account the context of use of each utterance, using clues such as the presence of an antecedent, and native speaker’s intuitions to determine the most probable reading of the particle. The results are summarized in Table 3.

The data suggest that the additive meaning is significantly more frequent than the mirative one in the Mid-20th century corpus compared to contemporary corpora (Fisher exact test: p = 0.03), in line with our hypothesis that this meaning diachronically came first. This could also be due to a difference in genre: movie scripts are more planned than radio shows or conversations, but even then this would be consistent with the idea that the additive meaning is more deeply entrenched than the mirative one. In addition, the Mid-20th century data also involves 4 instances of tim1 as a verb, against none in the later corpora, which is also consistent with the idea that the verbal usage predates the others.

Having laid out a path for the semantic shift of the particle tim1, we will now look at whether similar changes occurred in other languages.

There is evidence in many languages of a relationship between non-scalar and scalar additives, with the scalar uses coming after the non-scalar ones. English even and French même are two cases in point. Both have at their origin an expression that indicates the equality or the identical character of its host and its antecedent, rather than an indication of superiority of some sort (cf. Old English efen > even, Ancien Français meïme > même). In French, that use is still accessible in some contexts via the expression de même which is synonymous to the non-scalar additive aussi ‘too’, cf. (53).

The scalar use of non-scalar items is also discussed for Basque ere by Etxeberria & Irurtzun (2015). These authors argue that in a context like (54) the particle ere can receive a scalar interpretation, but that this is not part of the semantics of the particle. Rather this comes as an effect of the focalization of some constituents in the sentence, in a way that appears similar to the example (7-b) we discussed above.

Japanese mo ‘too/even’ is another case of an element that can be seen as either non-scalar or scalar, cf. (55) for which a translation of mo by even appears best suited, but not necessary: a non-scalar additive would also be appropriate.

There thus seems to be a tendency for non-scalar items to develop a scalar reading. This is probably also what happened to tim1, based on its original verbal meaning. The way we formalized the various meanings of tim1 helps to represent how its meaning evolved. It first acquired its existential presupposition, to which the scalar meaning was later added, along the lines of the mechanisms proposed by Etxeberria & Irurtzun (2015), which entrenched the scalar component in the meaning of tim1 (though note that tim1 still has a non-scalar flavor in some of its uses, which we captured in our formalization by making it encode non-strict superiority). The mirative use of tim1 represents another stage of the evolution in which the presupposition has disappeared and only the conventional implicature remains (i.e. the most subjective of the components).

The question is thus whether that last move is observed in other languages. For English, there already have been analyses that treat even as a mirative element (e.g. Zeevat 2009). More in line with our proposal, Kim & Jahnke (2011) consider a relatively understudied usage of even in utterance final position, which they claim also bears a mirative value. Example (56-a) is a case in point which marks the (pleasant) surprise of the speaker, and contrasts with (56-b) which does not convey that overtone.

Aside from these, there are many cases for which the use of even does not have the anaphoric requirement of its non-scalar counterpart too, cf. (57).

In (57), no explicit alternative for John’s cocktails is mentioned, though the first discourse segment provides the grounds for satisfying the presupposition (a similar example would work in Cantonese with tim1). Example (57) is still additive in the sense that the interpretation of the particle even (or tim1) requires an alternative. However, these examples offer the starting point from which such particles eventually will lose their presuppositional requirement by loosening the condition that the antecedent has to be clearly identified: an example like (57) ensures the existence of an antecedent, without being specific about its nature.

As a last note, the case of Japanese mo offers an interesting challenge. In some of its literary uses, mo appears to have a non-additive reading that is usually described as a “softener”, e.g. as in (58) where the use of mo downplays the importance of its prejacent.20

It is not clear whether and how this meaning can be derived by using the set of hypotheses we have laid out so far, since it seems the reading of mo is more or less opposite to a mirative one. Future work might hopefully shed light on the matter.