2.1 Argument structure nominals

Grimshaw (1990) argues that deverbal nouns are ambiguous between three readings: Argument Structure Nominal, Result Nominal, and Simple Event Nominal. ArgStrNs inherit the event structure of the verb, which accordingly licenses full argument structure as in (5a). Such nominals denote events, which are compatible with predicates of events such as took a long time, but not with predicates of individuals such as was on the table in (5a). When the deverbal noun does not realize arguments, it receives a more lexicalized reading. On the one hand, the noun assignment in (5b) is a result nominal (ResN) and denotes an entity just like article, which makes it compatible with was on the table. On the other hand, assignment may also combine with took a long time on its simple event nominal (SEvN) reading as in (5c), which denotes a conceptual event similarly to underived nouns such as war and event.

According to Grimshaw, this event reading is not enough evidence to say that simple event nominals inherit the verb’s event structure, since lexical nouns like war also have such readings (cf. Lieber 2016a). Event structure is specific to verbs, depicting their aspectual make-up (whether they are mono- or bi-eventive), which requires and licenses argument structure (see Rappaport Hovav & Levin 1998 and later work). For instance, when a bi-eventive event structure is present (as with causative verbs), arguments are realized hierarchically (i.e. internal arguments come first) and identify the two sub-events of the verbal event. In the absence of an internal argument, there is no evidence for event structure in the use of assignment in (5c). As Roy & Soare (2013) put it, the noun in (5c) denotes a ‘conceptual’ but not a ‘grammatical’ event.²

Grimshaw offers several other tests to distinguish between ArgStrNs and ResNs/SEvNs, which we do not dwell on here, as we will restrict our attention to those relevant for our discussion of SynCs below (see Alexiadou & Borer 2020 for an overview).³

2.2 Synthetic compounds

SynCs have formed the grounds for long theoretical debates between lexicalist and syntactic models of word formation (e.g. Roeper & Siegel 1978; Selkirk 1982; Lieber 1983; 2004; Ackema & Neeleman 2004; Harley 2009 and others; see Borer 2013: ch. 12 for an overview). Although our final analysis will be couched within the syntactic framework of Distributed Morphology, it is not our aim to argue for a syntactic theory over a lexicalist one, as this debate is orthogonal to our present goal to highlight some contrasts that any analysis should account for.

We focus on the behavior and morphosyntactic structure of nominal SynCs in comparison with the corresponding ArgStrNs headed by suffix-based nominals as in (1). This means that we do not aim to offer a unitary analysis for all constructions that have been previously labeled as synthetic compounds. It remains to be investigated whether our proposal for true SynCs may also need to include (some) SynCs whose non-heads resemble oblique arguments (as in (6a)) or those headed by participial heads as in (6b):

Nominal SynCs have received contradictory treatments in relation to ArgStrN. Below we present some of their properties following previous literature: we first focus on Grimshaw (1990) and Borer (2013), as the latter builds upon and argues against the former. We continue with Lieber’s (2016a, b) analysis, which shares similarities with and differences from both other accounts. We close with a discussion of event implication in ArgStrNs vs. SynCs.

3.1 Arguments and modifiers in deverbal compounds

Both Borer (2013) and Lieber (2016b) argue that external argument non-heads are possible in SynCs but with a crucial difference in their modeling. Namely, for Borer, who considers argument realization to be hierarchical and bound to syntactic event structure as in Grimshaw (1990), this entails that SynCs do not realize argumental non-heads but just some modifiers that receive a contextual argument-like reading otherwise available also in root compounds. SynCs are root compounds for Borer (2013). Lieber has a purely conceptual understanding of argument structure (ignoring syntax-semantics linking) and takes the non-heads of SynCs to realize true verbal arguments. In her Lexical Semantic Framework, the head of a SynC has the same semantic structure as an ArgStrN. By means of a Principle of Coindexation, the non-heads of the two SynCs in (17) get coindexed either with the first or the second argument of the base verb celebrate, depending on their semantic specification. If the non-head is sentient, like family, it may get coindexed with the first argument of the verb as in (17a), and if it is not, like birthday, it can only be coindexed with the second argument.

In what follows we bring evidence that the empirical picture is somewhere in between Borer’s (2013) and Lieber’s (2016b) approaches. Namely, we argue that non-heads that look like external arguments in SynCs are modifiers and not arguments, and that only those that are interpreted as internal arguments can be true verbal arguments. For ease of presentation, we refer to SynCs whose non-heads receive an internal argument-like interpretation as Obj-SynCs and those whose non-heads resemble an external argument as Subj-SynCs. We keep the term argument for syntactic arguments licensed by event structure as in Grimshaw (1990) and Borer (2013) (unlike Lieber 2016a). This means that we argue for a systematic difference between Subj-SynCs as in (17a) as taking a modifier non-head like root compounds and Obj-SynCs as in (17b) as possibly realizing an internal argument non-head.

A first difference between Subj- and Obj-non-heads is that when they co-occur they require a strict order in which the Obj-non-head must be closest to the head noun as in (18) (cf. Selkirk’s 1982 tree pasta eater vs. *pasta tree eater and McIntyre 2009 on the internal cohesion of compounds with internal arguments). If they were both modifiers like in root compounds, as Borer argues, they should be free to appear in either order, which is clearly not the case.

The test in (18) shows that the Obj-SynCs are more internally cohesive than the Subj-SynCs. Compounds are known to be syntactically inseparable: inseparability is a standard test for compoundhood (cf. ugly [black bird] vs. black ugly bird, the latter forming a phrase), but they are also known to exhibit this property to various degrees between compoundness and phrasehood (see Ziering & van der Plas 2020). At a reviewer’s suggestion, we also tested the order of Subj-SynCs in comparison to what may count as root compounds and found that the former are less separable than the latter: e.g. informants prefer spring [jury hearing] over jury [spring hearing] and winter [farm production] over farm [winter production], showing that the Subj-SynCs jury hearing and farm production are more cohesive than spring hearing and winter production, the latter coming closer to phrases.

However, when an internal argument is also present, the Subj-non-head is preferred as the outmost non-head, following the modifier: informants prefer state [spring [tax collection]] over spring [state [tax collection]] for the interpretation ‘collection of taxes by the state in spring’ and college [spring [transfer admission]] over spring [college [transfer [admission]] for ‘admission of transfer by colleges/the college in spring’. This means that spring tax collection and spring transfer admission are more cohesive compounds than the corresponding Subj-SynCs in (18b) and (18e), which are also compounds. If Subj-non-heads were arguments like the Obj-non-heads, we wouldn’t expect them to be separable by modifiers, just like Obj-non-heads cannot be separated by Subj-non-heads in (18). This difference between Subj- and Obj-SynCs suggests that the latter is closer to a modifier-head configuration (and ultimately, a phrase) than the former, which lexicalizes a predicate-argument structure as a compound.

Despite (18), one may still argue that Subj-SynCs involve external arguments, which may be syntactically more independent of the head. Below we provide further contrasts between Subj- and Obj-non-heads that support the proposed argument/modifier difference.

The second difference between Subj- and Obj-SynCs is in terms of productivity. Taking them both to have a similar make-up (whether as root or argumental compounds), a prediction of both Borer’s and Lieber’s analyses is that Subj- and Obj-SynCs should be similarly available. Yet, there is a striking difference between the two. Obj-SynCs are just as productive as their corresponding ArgStrNs: see (19c, d). It is hard, if not impossible, to find a verbal or ArgStrN construction that does not already have an established Obj-SynC.

By contrast, Subj-SynCs cannot be productively constructed, as noted already by Selkirk (1982); Grimshaw (1990); Bobaljik (2003) and most of the literature. The examples in (20) could at best be interpreted as Obj-SynCs, to the extent that the base verb is transitive.

As highlighted by Borer and Lieber, Subj-SynCs do exist; however, they are much less frequent than the Obj-SynCs. In an independent study on noun-noun SynCs built on transitive verbs and automatically collected from natural text corpora, for a dataset of 1864 SynCs on which all three American native speaker annotators converged, Iordăchioaia (2019) reports that only 12% were labeled as Subj-SynC, by comparison to 68% for Obj-SynCs and 20% for other modifier readings (e.g. a surprise demolition). This means that Obj-SynCs are five to six times more frequent than Subj-SynCs. If Obj-SynCs and Subj-SynCs had a similar syntactic status (as for Borer or Lieber), this asymmetry would be entirely unexpected. Most importantly, if they were structural arguments introduced by event structure, they should be available for any possible verbal construction, as they would be compositionally licensed. As noted just above (19), only Obj-SynCs behave like this.

Further examples of Subj-SynCs are given in (21), which shows that many of these SynCs involve a limited set of collective non-head nouns, which refer to organizations or groups: e.g. government, state, police, parliament, court, union, community, opposition, church, council, jury, media, as observed in Abrosimova (2017).

This shows that many Subj-SynCs belong to a restricted class with conventionalized non-heads that are typically conceptualized as agents/subjects. Yet, this subject-like interpretation cannot originate from verbal event structure, i.e., the non-heads of these compounds are not external arguments. If they were, they would be formed by a compositional event structure and there would be no explanation as to why those in (20) are not possible, while objects always are. Therefore, compounds as in (21) match the pattern of root compounds posited by Borer, unlike Obj-SynCs.⁶

Bobaljik (2003) makes a similar point concerning compounds headed by -ee nominals as in (22), which Lieber (2004; 2016b) takes to involve external argument non-heads. First, Bobaljik shows that the non-heads of these compounds typically cannot appear as by-phrases, which standardly introduce external arguments in ArgStrNs; see (23).⁷

Second, he notes that the non-head in these compounds is restricted to institutions, it excludes persons. The examples in (24) could at most be interpreted as dvandva compounds, while such non-heads make typical external arguments for the base verbs. This in line with Abrosimova’s (2017) observation about the collective flavor of subject non-heads in (21), although not all Subj-SynCs headed by action-denoting suffixes behave like participant-denoting -ee nominals. In general, Iordăchioaia (2019) shows that there are even Subj-SynCs headed by SEv nominals, which do not show external argument behavior (see examples in §3.2).

Bobaljik concludes that compounds as in (22) receive a reading such as ‘employed/evacuated/bombarded under the auspices of or on behalf of the city/UN/US’, a modifier interpretation typical of root compounds. The Subj-SynCs in (21) are slightly different, which shows us the diversity of these modifier readings by contrast to structural external arguments.

In the next section, we provide structural tests that show that Obj-SynCs indeed present event structure properties that license argument structure, while Subj-SynCs do not.

3.2 Event structure in synthetic compounds

In §2 we saw already that Lieber (2016b) brings corpus evidence against Borer’s (2013) claim that SynCs lack event structure properties. Lieber’s data in (13) indicate that SynCs license by-phrases, purpose clauses and aspectual adjectives. Below we add further evidence from Iordăchioaia (2019) that it is in fact only Obj-SynCs that have these properties, and Subj-SynC do not, in support of our claim that only the former involve argumental non-heads, while the latter involve modifiers. Iordăchioaia (2019) carried out a questionnaire study with minimal pairs of Obj- and Subj-SynCs and found a clear contrast between them in terms of event structure. In (25) and (26) we see data that test the presence of aspect by means of in/for-adverbials and aspectual adjectives. As the two examples show, the grammaticality of the Subj-SynCs is substantially lower than that of the corresponding Obj-SynCs.⁸

Besides aspect, Voice properties that indicate the presence of event structure licensing external arguments were also tested with similar results. For the verbal domain, modifiers like intentionally, deliberately, carefully and control into purpose clauses are known to require the presence of Voice (see Kratzer 1996; Alexiadou et al. 2015 and references therein), as shown in (27a) and (28a).⁹ (27b, c) illustrate modification by the adjectives intentional and deliberate, and (28b, c) purpose clauses with SynCs. Again, we find a strong contrast between Obj-SynCs, which are fully grammatical, and their corresponding Subj-SynCs, which are unacceptable.

This contrast is particularly surprising, if we consider that Voice is responsible for the licensing of external arguments: if anything, we expect the SynCs whose non-heads are interpreted as subjects to be compatible with Voice properties and not those whose non-heads are interpreted as objects. And yet, (27b) vs. (27c) and (28b) vs. (28c) show precisely the opposite. Despite the fact that such Obj-SynCs do not overtly realize the external argument of the base verb, they prove to include the necessary event structure of verbal constructions with Voice. That is, in (28b), they must include the variable of an implicit external argument that controls the PRO subject of the purpose clause just like the overt agent in (28a).¹⁰

3.3 Interim conclusion

To conclude on our empirical picture of SynCs, we saw in §2 that both Subj- and Obj-SynCs are available, as observed by Borer (2013) and Lieber (2016b), and that SynCs also exhibit clear properties indicative of event structure, as argued in Lieber (2016b), contra Borer (2013). While Borer argues that all SynCs are root compounds and involve modifier non-heads, Lieber proposes the opposite, namely, that all SynCs are argumental, realizing either internal or external arguments. The presence of event structure properties is problematic for Borer’s analysis but compatible with Lieber’s semantic account. In this section we have shown, however, that the empirical picture is in between the two scenarios. In particular, we provided evidence that there is a strong structural contrast between Obj- and Subj-SynCs and that only the former exhibit event structure properties and are truly argumental, as in Lieber’s (semantic) approach, while the latter behave like root compounds, as in Borer’s approach.

4.1 Background on Distributed Morphology

Marantz (2001; 2013) and Arad (2005) propose two types of word formation in Distributed Morphology (DM): i) inner derivation (from the root) and ii) outer derivation (from a stem), as illustrated in (29):

The two display different properties. Inner derivation presents the following: 1) negotiated (idiosyncratic) meaning of the root in the context of the functional morpheme (e.g. in combination with little n the root √globe may mean ‘sphere’ or ‘the world/planet’, but a realizes only the latter meaning in global; Marantz 2013); 2) selectional restrictions (i.e. some roots are better than others with a particular morpheme: see adj. √malic-(i)ous/*y vs. √clums-y/*ous, Arad 2005); 3) the meaning of the construction depends on root semantics independent of argument structure operations from functional structure. Outer derivation has opposite properties: 1) compositional meaning predicted by the stem (e.g. [glob-al]-ize ‘make global’/*‘make into a sphere’ preserves the adjectival meaning it is derived from); 2) no selectional restrictions (see [malici-ous]-ness, [clums-i]-ness); 3) the meaning of the construction may involve arguments from functional structure. The third point will be illustrated in (30) below.

4.2 Argument structure nominals in Distributed Morphology

With these tools we can analyze the different readings of deverbal nominals from §2.1. For the ArgStrN reading in (5a) the suffix nominalizes the full event structure of the base verb up to AspectP (Alexiadou 2001) as a stem-based derivation: see (30). The root √demolish undergoes head-movement up to n and acquires the syntactic and semantic contribution of each functional head on its way. The presence of AspectP allows the adverbial in two hours, and VoiceP accommodates the by-phrase as the external argument. ResNs/SEvNs as in (5b) and (5c) do not involve any event or argument structure, so they are root-derived as in (31) (cf. (29a)). The root √assign is polysemous: it may denote an event or its result. Both readings are preserved in the negotiation of its meaning with the n head realized by the suffix -ment, and we obtain the SEvN and the ResN assignment as the output of allosemy (Marantz 2013; see Wood 2021: 5 for a recent summary of layering approaches to nominalizations).¹¹ In the syntax n is an abstract nominalizer, for which a suffix allomorph will be lexically inserted at Spell-Out (see Late Insertion) depending on the root. This is how n is spelled out as -ion for √demolish and -ment for √assign (see Embick 2010: ch. 2 for a detailed discussion).

4.3 Synthetic compounds in Distributed Morphology

A first analysis of SynCs in DM is offered in Harley (2009), which relies on root incorporation. Like us, Harley assumes that SynCs include internal arguments (but does not discuss the existence of Subj-SynCs), and, for a compound like truck driver, she takes the verbal root √drive to incorporate the internal argument noun truck and to then get nominalized as a compound by the suffix -er, without ever creating a verb, as shown in a simplified version in (32). This way, Harley excludes unavailable compound verbs like *to truck drive. However, as Borer (2013) points out, this analysis cannot account for SynCs that include an overtly verbalized verb (e.g. root verbal-iz-ing, water pur-if-ication), given the assumption that in (32) there is not even a simple verb in the structure of such compounds.

Iordăchioaia et al. (2017) investigate SynCs with -er nominal heads in Greek and English and formulate a distinction between SynCs that involve incorporation of their root non-heads, as in Greek, and English SynCs, whose non-heads are morphologically complex (e.g. dispos-ition lifter) and should not be incorporated. In (33), we adjust that analysis to our argumental Obj-SynCs such as house demolition by the army. Following Iordăchioaia et al., the non-head house moves to Spec nP, as its lack of a DP structure prevents it from receiving (prepositional genitive) case in an ArgStrN configuration, and it thus cannot stay in its original argumental position (see Longobardi 1994). This accounts for the internal cohesion of SynCs, which we observed in (18), without assuming root incorporation as in Harley (2009). Given that the non-head nP moves to Spec nP, there is no formation of a compound verb such as *to house demolish. Moreover, this structure contains a vP layer for the base verb, which can accommodate verbalizing suffixes to account for SynCs like root verbal-iz-ing and water pur-ifi-cation.

However, in contrast to Iordăchioaia et al. (2017), the data in (13b), (25) and (26) indicate that aspectual properties are also available in Obj-SynCs, which leads us to posit a verbal event structure identical to that of ArgStrNs in (30), including AspectP. Consequently, the only difference between ArgStrNs and SynCs concerns the non-head, which is a full DP in the former and stays in situ receiving genitive case, while it is a simple nP in the latter and must move to Spec nP for lack of case. The last position of the nP non-head is the same as in root compounds such as house roof, but in SynCs the non-head is an argument introduced by the event structure of the vP, which gives it the interpretation of an internal argument, unlike the non-heads of root compounds, which are interpreted by means of encyclopedia or context (see Delfitto et al. 2011). The so-called Subj-SynCs in (21) are built on a ResN head without verbal event structure (see (31)) and have a structure identical to that of root compounds built on lexical nouns.

The structure in (33) strictly accounts for the morphosyntax and the lexical semantics of argumental SynCs but for a DM approach it may raise questions about the locality domains involved and the availability of contextual allomorphy or allosemy in compounds. Although such cases have been pointed out for other languages in Harðarson (2021), we are not aware of any contextual allomorphy or allosemy that would appear in English SynCs and be absent in the corresponding ArgStrNs. To the extent that the meaning of the non-head globe were indeed sensitive to the root of the compound head in globe trotter ‘world traveler’ vs. globe smasher ‘smasher of spherical objects’, exhibiting a case of contextual allosemy in a SynC, as Harðarson (2021) claims (cf. §4.1), it seems to us that the same meaning distribution would appear in the corresponding ArgStrNs (the trotter of the globe vs. the smasher of the globe). We agree though that more intricate examples of such contrasts may become available upon further study, and then we assume that Harðarson’s (2021) account could cover those as well.

4.4 Presence and absence of Voice in -er synthetic compounds

While eventive SynCs headed by -ing or Latinate suffixes were found to exhibit both external arguments and aspectual properties, this does not carry over to participant SynCs headed by -er nominals, as Lieber (2016b) also observes. -er nominals are known to show more restricted verbal properties even as ArgStrNs (e.g. the driver of the truck; see Alexiadou & Schäfer 2010; Roy & Soare 2013). We focus here on -er nominals that denote external arguments of their base verbs (i.e. agents, causes, instruments), as this is the most productive meaning of the suffix and the one that originates in compositional event structure (see Booij 1986; Booij & Liber 2004; Alexiadou 2014; Lieber & Andreou 2018 on less productive patterns). Since they usually denote external arguments, such -er nominals have been argued to nominalize the VoiceP of their base verb (see Schäfer 2008a; Alexiadou & Schäfer 2010; but cf. Baker & Vinokurova 2009). The question is whether -er SynCs also involve VoiceP like the corresponding ArgStrNs and whether they always do so.

Importantly, -er SynCs could be argued to involve VoiceP, as long as they denote persons; instruments and inanimates, in general, cannot act agentively, as required by Voice. While ArgStrNs built on -er nominals usually denote agents, as shown in Alexiadou & Schäfer (2010), SynCs are more diverse. Alexiadou & Schäfer (2010: 19–20) and others before (see §2.2.4) have shown that many -er SynCs denote both persons and instruments/tools, as in (35).

We propose that the different readings of -er SynCs and their contrast to ArgStrNs originate in the presence, respectively, absence of Voice. In essence, we argue that -er SynCs are ambiguous: they may or may not include Voice with effects on their interpretation. On the one hand, in the presence of Voice, we obtain the person/agent reading of the er SynC, whose functional structure is identical to that of -er ArgStrNs, as shown in (35). Unlike in (30), the -er suffix does not form action but external argument nominals. That is, it appears in a configuration in which it becomes coreferential with an external argument variable in the event structure of the verb, which cannot be independently occupied by another DP (see Booij’s 1986 analysis with -er binding the verb’s theta-role). Following Schäfer (2008a), the suffix -er – and eventually the noun – is coindexed with the external argument introduced by Voice, which must be an agent/person.

This analysis suggests that person-denoting -er SynCs and the corresponding ArgStrNs share the event structure in (35), which is what we are aiming for. Like in the case of eventive ArgStrNs and their corresponding Obj-SynCs, the difference is that the internal argument in SynCs is a bare nP and must move to Spec nP, while in ArgStrNs it receives of-genitive case.

On the other hand, in the absence of Voice, an agentive reading of a SynC is unavailable, and the closest possible compositional interpretation that the -er SynC may obtain is that of an instrument/tool, which cannot act on its own in an event but may contribute to its coming about. They behave like the causer instruments described in Alexiadou & Schäfer (2006). Concretely, this means that, instead of an external argument variable provided by VoiceP (see (35), Schäfer 2008a), -er would bind the causer of a causative vP like causers of unaccusative verbs, which also lack Voice (e.g. The door opened from the wind). Such causers typically take over the subject position of the causative verb when an agent is missing: cf. The wind opened the door (see v-CAUS in Alexiadou et al. 2015: 30–36, which corresponds to our v in (36)). We obtain the instrument-denoting -er SynCs from the structure in (36), which uses Schäfer’s (2008a) implementation of the suffix -er, now nominalizing a simple vP structure (without VoiceP) and coindexing its variable with the x_i variable of a causer in this case.

4.5 Event implication

Let us now turn to the contrast between ArgStrNs and SynCs in terms of event implication reported in §2.2.4. If we take person-denoting -er SynCs to be essentially headed by ArgStrNs, we predict that they should pattern similarly in terms of event implication: see (14), (15), and (16). Recall that the debate on event implication is unsettled with respect to person-denoting compounds, since it seems unnatural to say of somebody that s/he is a truck driver without their ever having driven a truck. In the case of instrument readings, this situation is clear: a lawn-mower need not even be functional to qualify as such. Our analysis in (35)–(36) offers a straightforward explanation for this dilemma: the complex structure including Voice of person-denoting -er SynCs also triggers event implication, but the smaller structure of instrumental compounds does not.

That the crucial projection relevant for event implication is Voice is confirmed by the use of Voice-modifiers with these compounds. In (37), the addition of careful (or accidental, see Schäfer 2008b, cf. Reed 2011) eliminates the instrument interpretation of these compounds.¹² The default reading of lawn-mower and coffee-grinder seems to be that of a tool rather than a person, but once we introduce the Voice modifier, the tool reading disappears and only the person reading is possible:

In our study we also checked whether careful is interpreted in relation to the action or the person: i.e. whether a careful lawn-mower denotes a possibly careless person who carefully mows the lawn (action reading) or a generally careful person who may carelessly mow the lawn (person reading) (cf. beautiful dancer in Larson 1998). While the person-related reading is available, the event-related one is always preferred; accidental is only compatible with the latter. In addition, our consultants find that, e.g. a careful/accidental lawn-mower must have mown the lawn at least once, confirming that the presence of Voice also brings about event implication, just like in ArgStrNs.¹³ Without Voice-modifiers, these SynCs are underspecified, but child murderer is not, as murder is an agentive verb and always requires Voice with event implication, leaving no room for Voice-less instrumental readings of murderer.

A final note is in order. Given our focus on the fully productive and compositionally formed readings of -er nominals, as also acknowledged in Lieber & Andreou’s (2018) corpus work as the most frequent ones, we cannot address all possible readings of -er nominals here. Lieber & Andreou (2018) identify several contextual factors that contribute to the interpretation of -er nominals but they do not offer a comprehensive account of this interaction either (see p. 214). We believe that the influence of context is modulated by the kind of (modal or quantificational) operator that eventually binds the event variable embedded by the suffix, and this is a matter to be dealt with by formal semantics in the Logical Form, which is not within our scope (see how habitual readings are accounted for in Romanian supine nominals in Iordăchioaia & Soare 2015). Here we have argued for a structural difference between the compositional agentive vs. instrumental readings of -er SynCs in terms of presence/absence of Voice, which is supported by (37) and further facts with idioms in section 5.

To conclude, we have argued that true SynCs realize internal arguments, while those with apparent external arguments are root compounds. The former include event structure with (at least) a vP, which introduces the internal argument as in (33). The latter are headed by ResNs/SEvNs as in (31), and their non-heads, although syntactically occupying the same Spec nP position as in SynCs, are interpreted as modifiers (rather than arguments) of the heads. Importantly, we also showed that Voice (with agentive properties) is available in most SynCs, with important consequences on the interpretation of -er SynCs and event implication. We essentially argued that person-denoting -er SynCs involve Voice, while instruments do not.¹⁴ These insights on Voice in SynCs will be further tested and confirmed by looking at idioms.

5.1 Verbal idioms

There is a long discussion in the syntactic literature about how complex a verbal structure can become an idiom. An old idea defended by Marantz (1984; 1997) is that an idiomatic reading certainly includes internal arguments and may include Voice (i.e. DP_Agent + [V + DP_Theme]_idiom) but never the agent that Voice realizes: that is, idioms do not include fixed agents. Harley & Stone (2013) carefully investigate apparent challenges and provide further support for this generalization. Our idioms also conform to this generalization, but we will see that we find some important contrasts as to whether Voice (without the agent) is part of the idiomatic reading or not, which bear on the available idiomatic SynCs.

To the extent that agent-oriented modifiers such as carefully and intentionally diagnose a Voice projection, only the idiom to kick the bucket is incompatible with Voice in (43). To break the ice in (43c) allows both adverbs, while the other idioms are fine with intentionally and less so with carefully. This contrast must be aspectual: unlike intentionally, carefully imposes an accomplishment reading, which the idioms in (43b, d–f) do not allow out of context.

All the idioms in (43b–f) present evidence for a possible Voice projection,¹⁹ although some may be ambiguous. For instance, to catch the eye has two idiomatic readings: a non-agentive one ‘to become apparent’, in which both adverbs are excluded, and a possibly agentive one ‘to meet the glance of another’. (43d) illustrates the latter, which involves Voice.

In view of the structural variation that idioms present (Fraser 1970; Nunberg et al. 1994; Punske & Stone 2014; Stone 2016) and the availability of Voice in (43b–f), we must distinguish between different types of idioms depending on whether they include/exclude Voice. Voice has been proposed to be part of idioms in Folli & Harley (2007), who argue that some idioms do not allow passivization because they project their own Voice, which conflicts with passive Voice. Punske & Stone (2014) and Stone (2016) use this proposal to argue that, in English, idioms like kick the bucket also project Voice and block passivization. Yet, (43a) indicates that this idiom does not project Voice – at least, not agentive Voice. In addition, the passive should not be blocked by the presence of Voice since it applies to agentive constructions. It seems more plausible to state that kick the bucket is an idiom that excludes Voice, because its interpretation is unaccusative (‘to die’): see Burzio’s (1986) generalization.²⁰²¹

The fact that all the other idioms are compatible with Voice-modifiers suggests that their idiomatic reading either requires the projection of agentive Voice, whose external argument is not part of the idiom (see Marantz 1984), or it does not require Voice but does not conflict with Voice either, yielding two possible readings. In view of our discussion of -er SynCs in §4.4, idiomatic SynCs can help us to determine which verbal idioms belong to which category, as we expect those based on idioms that require Voice to yield only person readings and those based on idioms that do not require Voice to yield instrument readings but possibly also person readings when Voice is included.

5.2 Idiomatic -er synthetic compounds

We saw in §4 that SynCs include a vP with an internal argument and may also realize VoiceP. If a verbal idiom forms an -er SynC, this means that the idiom could be a vP or a VoiceP as the compound would be compatible with both. However, in view of §4.4, the instrument reading of -er SynCs comes about only in the absence of Voice. Thus, an idiom that obligatorily projects Voice should not derive instrument -er SynCs. In other words, if an idiomatic -er SynC has no instrument reading, this indicates that the verbal idiom includes Voice, which imposes a person/agent reading on the compound. Another implication is that idiomatic -er SynCs with an instrument-only reading will point to a verbal idiom without Voice. Turning now to the verbal idioms in (43b–f), they form the following idiomatic -er SynCs:

According to (44a), whistleblower allows only the person reading as an idiom, which indicates that the verbal idiom to blow the whistle must include (agentive) Voice; the idiomatic reading disappears in the absence of Voice on an instrument reading of the SynC. Ice-breaker offers the opposite pattern: as an idiom it has an instrument reading, i.e. ‘a game people play to get to know each other’ or ‘a line used to start a conversation’. This idiom must be formed below VoiceP: enforcing a person denotation with Voice on ice-breaker triggers a literal interpretation. This contrast between to blow the whistle and to break the ice is supported by the use of non-agentive causers. Causers exclude Voice and are incompatible with to blow the whistle (e.g. *This will blow the whistle) but felicitous with to break the ice (This will break the ice).

Another argument for the two SynC idioms whistleblower and ice-breaker as exhibiting opposite patterns that obligatorily include, respectively, exclude VoiceP comes from their use with adjectival Voice-modifiers, as in (45). Whistleblower is compatible with them on an action-modifier reading (cf. (37)), as predicted by the obligatory Voice and a derivation as in (35). Ice-breaker marginally allows them, whereby the ratings in (45b) require a person denotation, which is not readily available (cf. (44b)). This SynC would have to be derived by means of the structure in (36) as instrument-denoting.

By comparison to the verbal idiom in (43b), we observe that careful is much improved in (45a). This means that the aspectual restriction of the verbal idiom to blow the whistle is lost in the SynC, which is explained by the simple fact that the SynC does not inherit the aspectual value of the verb (see §4.4). In addition, the verb in (43c) fares much better with Voice-modifiers than the SynC in (45b). For us, this means that the SynC has become lexicalized as an idiom without Voice, while the verbal idiom to break the ice is also formed below Voice but is compatible with Voice, projected above the idiom domain.

The corresponding idiomatic event ArgStrNs also support this contrast between to blow the whistle and to break the ice. As we saw in (39e), an ArgStrN with an overt external argument is rated better than one with a covert external argument as in (39c) for the idiom to blow the whistle: that is, Voice is required by both the ArgStrN and the verbal idiom, and the overt external argument improves the idiomatic ArgStrN. By contrast, the ArgStrN of to break the ice worsens with an overt external argument ((41a), (41g) vs. (41d)). Here, Voice is not a necessary part of the idiom and enforcing its realization in the ArgStrN yields a literal interpretation, just like in the case of person-denoting -er SynC in (44b).

Having discussed these two opposite patterns of SynCs in terms of presence/absence of Voice, let us turn to the SynCs in (44c–e), whose behavior indicates a less clear status. Our consultants prefer eye-catcher as a person (see (44c)), but this SynC also allows an instrument reading attested in corpora (see (46a, b)). This means that eye-catcher can be used as an idiom both with and without Voice, yielding person and instrument readings, respectively (cf. the ambiguous verbal idiom in (43d)). (39b) and (41c) show that our consultants find bean-spiller and leg-puller less natural than the other idiom compounds and, if at all, they receive a person reading in (44d–e). This preference is confirmed by the few corpus attestations for bean-spiller in (40d) and for leg-puller in (46c). The corresponding verb idioms realize Voice (cf. (43e–f)), and the instrument reading is marginal in SynCs in (44d–e) and unattested in corpora. These facts indicate that, to the extent that bean-spiller and leg-puller become more established, they would join the pattern of whistleblower in (44a), including Voice and referring only to persons.

To summarize the picture on the interaction between verbal idioms and their corresponding SynCs, we can identify three categories of verbal idioms in terms of presence/absence of Voice:

• I.    Idioms that reject Voice, because their meaning conflicts with Voice: e.g. unaccusative to kick the bucket (see (43a)) and non-agentive to catch the eye (see the instrument SynC in (44c));

• II.   Idioms that obligatorily include Voice, since the lack of Voice brings about a literal meaning in SynCs: e.g. to blow the whistle (see (43b), (44a), (45a)) and quite likely also to pull one’s leg (see (43e)/(44d)) and to spill the beans (see (43f)/(44e));

• III.  Idioms that are formed below VoiceP but whose interpretation is mostly compatible with Voice: see agentive to catch the eye in (43d)/(44c) and to break the ice in (43c)/(44b).

The corresponding idiom -er SynCs behave similarly to their base verbs and are split accordingly in the three classes above. The exception is ice-breaker, which seems to have become lexicalized on an instrumental reading and, despite its verbal idiom’s belonging to class (iii), with optional Voice, it rejects Voice and belongs to the first class, like the instrumental eye-catcher. These SynCs are derived by our structure in (36) only. Class (ii) SynCs are derived by means of (35), and class (iii) SynCs (e.g. eye-catcher) are ambiguous between (35) and (36), like the non-idiomatic -er SynCs in §4.4.

The unavailability of the SynC bucket-kicker may receive at least two explanations. First, the apparent direct object of to kick the bucket is not an internal argument (the subject is), which, in our approach, conflicts with the structure of argumental SynCs (whether agentive or instrumental). However, this would leave the option of a root compound open, which is not available either. A second reason may be the implausibility of the idiom as a one-time event to be conceptualized as a generic activity (as compounds usually are). In conclusion, besides the structure of the idiom, semantic-pragmatic conditions may also play an important role in preventing the formation of the idiomatic compound bucket-kicker.

In this section we also saw that the contrast between idiomatic SynCs and ArgStrNs is not as strong as presented in Borer (2013). At least following the judgments of our consultants and corpus evidence, we showed that both SynCs and ArgStrNs are similarly (un)acceptable for verbal idioms, beyond the semantic-pragmatic differences triggered by the full DP vs. nP status of their internal argument. One difference we noticed in (39) vs. (41) is that the presence of an agent improves eventive ArgStrNs built on an idiom that requires Voice (see the blowing of the whistle in (39c, e)), but worsens those built on an idiom that does not require Voice (see the breaking of the ice in (41a, d)). This can be explained by the fact that ArgStrNs based on -ing must include Voice and, if Voice is not necessary for the idiom, the idiomatic reading will not be preserved in the corresponding ArgStrN. -er SynCs, however, may receive instrumental readings in the absence of Voice and are available for such idioms as for ice-breaker in (44b).