1 Introduction

At least since Marantz (1984), it has become widely accepted that agents qua external arguments have a different status from internal arguments, in that they do not condition the meaning of the verb as internal arguments do. As shown in the by now classic examples in (1)–(2), the meaning of the verb kill varies depending on the referential properties of its internal argument, but no such special interpretations of the verb arise on the basis of the referential properties of the external argument.

Building on Marantz’s observation, Kratzer (1996) proposed an analysis on which the above observation falls out from the fact that verbs denote functions from their internal arguments only. External arguments, including agents, are not semantic arguments of the verb, and thus do not compose with the VP by Function Application. Rather, a special functional head, dubbed Voice by Kratzer, introduces the external argument, and composes with the VP by the compositional rule Event Identification. Since the external argument is introduced VP-externally, it is never local enough to the verb to condition special interpretations in the same way internal arguments can as direct arguments of the verb.

While highly influential, Kratzer’s approach is not without its critics (e.g., Wechsler 2005; Bale 2007; Müller & Wechsler 2014; Wechsler 2020). Bale (2007), in particular, shows that Kratzer’s approach makes incorrect predictions regarding the availability of certain types of repetitive presuppositions with the sub-lexical modifier again. Again is an ideal modifier for testing the hypothesis that external arguments are always severed from verbs, because it triggers a repetitive presupposition whose shape is uniquely determined by the constituent to which it attaches. If the agent is always severed from its verb, then again should be able to scope between the VP and the agent, producing a repetitive presupposition that excludes the external argument, what Bale calls a subjectless presupposition. If Kratzer is right, again’s presupposition should always be satisfied by a previous event of the sort described by the VP regardless of the identity of the agent. As Bale shows, the empirical evidence suggest that Kratzer (1996) is not completely right: while transitive verbs seem to allow subjectless presuppositions readily, as expected on Kratzer’s analysis, intransitive verbs and stative transitive verbs uniformly resist subjectless presuppositions. This in turn leads Bale to conclude that not all external arguments can be severed from their verbs, and that both transitivity and the eventivity/stativity distinction are crucial factors in determining how external arguments relate to their verbs.

In this paper, we challenge Bale’s conclusions from two directions. First, we point out two classes of eventive transitive verbs that resist subjectless presuppositions: verbs of ingestion like eat and drink (Jackendoff 1992; Jerro 2019) and intentional verbs of killing like murder massacre (Beavers & Koontz-Garboden 2012; Ausensi et al. 2021). Second, we challenge the claim that resistance of subjectless presuppositions by a particular verb class requires an analysis on which the agent is not severed from verbs of that class. This is implemented by building on an intuition due to Hale & Keyser (2002), in the form of an anaphoric index encoded semantically in the verb. Stative transitive verbs and verbs of ingestion and intentional killing bear indices that are bound by the verbal functional head syntactically that introduces the external argument, here Kratzer’s Voice, via a mechanism proposed for reflexive pronouns in Kratzer (2009). Compositionally, this straightforwardly accounts for those cases where transitive verbs resist subjectless presuppositions with again in a way that is fully compatible with the idea that agents, and external arguments more generally, are uniformly severed from the verb.

We proceed as follows. Section 2 provides a background discussion of the proposal in Kratzer (1996) and the predictions it makes regarding subjectless presuppositions as discussed by Bale (2007). Section 3 demonstrates that just like stative transitive verbs, verbs of ingestion and intentional verbs of killing similarly resist subjectless presuppositions. Section 4 presents our analysis in terms of indices and binding, together with the empirical motivations about the specific lexical entailments of these two verb classes that necessitate reference to an agent argument. We also present further evidence from reflexive theme arguments conditioning the availability of subjectless presuppositions in support of the proposed analysis. Section 5 concludes the paper by reflecting on the larger question of whether agent arguments are syntactically and semantically represented in the verb, drawing on previous criticisms of Kratzer (1996) in Wechsler (2005; 2020) and Müller & Wechsler (2014), in line with the broader objective of this special volume.

2 Background

To account for the observations in (1)–(2), Kratzer (1996) proposes that external arguments are not semantic arguments of the verb. Formally, taking the type of individuals as e, the type of eventualities as v, and the type of truth values as t, transitive verbs denote functions from individuals to event predicates, type <e,<v,t>>, where the individual argument corresponds to the internal argument of the verb. An example using the transitive verb feed is illustrated below, with the thematic specification of the verb’s arguments given in Neo-Davidsonian fashion as functions from events to individuals.

Kratzer proposes that composition of the VP with the external argument is achieved through a separate functional head Voice, whose semantic contribution is to relate an individual to the event contributed by the verb. For example, agentive Voice will have the denotation below, a function of type <e,<v,t>>.

Notice now that upon composition with its internal argument, the verb phrase (VP) feed the dog denotes an event predicate of type <v,t> in (5). Neither the VP nor Voice can compose via Function Application, for neither is the right type to serve as an argument to the other.

To compose these two constituents, Kratzer proposes a rule of Event Identification. Taking as input two functions of type <e,<v,t>> (Voice) and <v,t> (VP), Event Identification produces as output a function from individuals to event predicates derived by saturating the arguments of the functions, conjoining the result, and then abstracting over the event argument common to both functions and the individual argument of the first function (6).

With the above ingredients in place, we can now see a full derivation of the introduction of a transitive verb’s arguments (7) (Kratzer 1996: 122)

Kratzer’s implementation ensures that the agent is never a true argument of the verb. It is this property that leads to the observations in (1)–(2). The proposal also makes a further testable prediction: given that the external argument is never contained in the VP, then any modifier that can target VP-type constituents should be able to make semantic contributions that exclude the external argument. One modifier with the right properties for testing this prediction is the presupposition trigger again. Again is standardly treated as a modifier of event predicates (type <<v,t>,<v,t>>) that introduces a presupposition that an event of the sort denoted by its argument occurred previously (von Stechow 1996; Beck & Johnson 2004; Bale 2007). Importantly, again’s presupposition is uniquely determined by the event predicate it takes as argument.

As Bale (2007) notes, if external arguments are always severed from the verb, then the VP comprising the verb and its internal argument is of the right type for again to modify. Additionally, since the VP contains no specification of the external argument, again will simply presuppose that an event with the same internal argument occurred previously. The external argument of this previous event may therefore differ from that of the asserted event, producing what Bale dubs a subjectless presupposition. As Bale demonstrates, such subjectless presuppositions are indeed attested with eventive transitive verbs, such as hit and kick. The contexts below specify that two distinct agents carried out the previous and the at-issue event, indicating that again’s presupposition indeed excludes the agent argument in the (a) examples. Note that the availability of a subjectless presupposition is sensitive to the linear placement of again (von Stechow 1996; Beck & Johnson 2004); in the (b) examples, again appears to the left of the agent argument, indicating it must take the agent in its scope and therefore include it in its presupposition, explaining the infelicity of the (b) sentences in the given contexts. The (c) examples involve passivized versions of the verb, standardly assumed to existentially quantify the agent argument introduced by a passive Voice head, and are therefore also compatible with the given contexts.

The examples in (9)–(10) suggest that Kratzer’s approach is on the right track, and there is indeed a VP constituent excluding the agent argument. Nonetheless, Bale makes a crucial observation; unlike (9)–(10), eventive intransitive verbs and stative transitive verbs uniformly reject subjectless presuppositions with again. Importantly, examples with unergative verbs like run in (11) show that subjectless presuppositions are not uniformly available with agent arguments.

On the basis of this evidence, Bale concludes that the availability of a subjectless presupposition with again turns on the transitivity and eventivity of the VP in question, which in turn suggests that not all external arguments can be severed from their verbs. Rather, stative transitive verbs and intransitive verbs uniformly denote functions from all of their arguments, ruling out subjectless presuppositions. More formally, stative transitive verbs are of type <e,<e,<v,t>>> (15a) and eventive intransitive verbs of type <e,<v,t>>. This is in contrast to eventive transitive verbs, which return an event predicate when supplied with an internal argument (15d), explaining their compatibility with subjectless presuppositions. The entries for examples of each verb class discussed above are provided in (15).

3 Intentional killing and ingestion verbs

Bale’s observations lead to the conclusion that certain classes of verbs must denote functions from their external arguments while others do not in (15a)–(15d), thereby explaining the distribution of subjectless presuppositions with again. Type-theoretically, this explanation seems adequate and correctly predicts when a subjectless presupposition is (un)expected with again. Nonetheless, this approach raises two important questions. First, on an empirical level, does Bale (2007)’s conclusion hold up across sub-classes of transitive verbs, whether eventive or stative? Second, does the fact that a particular class of verbs resists subjectless presuppositions necessitate the analysis that Bale suggests, on which the agent is not severed from verbs of certain classes?

In this section, we suggest that the answer to the first question is negative, illustrating with two transitive verb classes that resist subjectless presuppositions, verbs of ingestion and verbs of intentional killing. We also provide a negative answer to the second question, and propose a unified analysis of verbs that resist subjectless presuppositions without positing differences in the manner in which verbs ultimately associate with their external arguments.

Beginning with the first question, we identify two eventive transitive verb classes that resist subjectless presuppositions, contrary to what is expected given Bale’s observations. The first class consists of intentional verbs of killing, such as murder, assassinate, slay, etc. As shown at length by Ausensi et al. (2021), these verbs systematically require that any previous event satisfying again’s presupposition be carried out by the same agent, ruling out subjectless presuppositions.¹

Interestingly, as has been noted by many, semantically similar verbs like kill, which at least shares some subset of the lexical entailments of intentional verbs of killing, freely license subjectless presuppositions with again.

The second class consists of verbs of ingestion, such as eat, drink, swallow, etc. In contexts establishing distinct agents between the presupposed and asserted events, we see plainly that again does not produce subjectless presuppositions when modifying verbs of this class; a different agent ingesting the theme previously does not license the use of again.²

Taken together, we see that, contra Bale (2007), transitivity and eventivity are not the only criteria regulating the availability of subjectless presuppositions with again. There are clearly eventive transitive verbs that resist subjectless presuppositions, alongside intransitive and stative transitive verbs.

Moving on now to the second question, what might be the intuitive explanation behind why intentional killing and ingestion verbs resist subjectless presuppositions, while semantically similar verbs like kill do not? Our answer begins with the observation that both classes of verb make reference to the agent in more than one component of their meaning. For verbs of intentional killing, this lies in entailments about the agent’s intentions to kill, while for verbs of ingestion, this lies in reference to the ingester in the result state component of the verb’s lexical semantics. We address each of these points in turn.

As Ausensi et al. (2021) note, one key difference between kill and intentional verbs of killing, as suggested by the label used here, concerns the entailment of intentionality of the agent in carrying out a killing event. While intentional killing verbs require that the agent intentionally carry out the event, kill does not. Evidence for this comes from the possibility of non-animate causer subjects and modification by adverbial phrases that cancel or reinforce intentionality entailments, such as by accident or on purpose. Kill permits such subjects and adverbials while intentional killing verbs do not. This suggests that intentional killing verbs lexically entail intentionality.

Moving on to verbs of ingestion, there is a significant precedent for analyzing these as bi-eventive change-of-state verbs. On this treatment, these verbs encode an entailment of motion, in which a theme argument moves into a body part specified as part of the agent (Jackendoff 1996; Krejci 2012; Jerro 2019). For example, Jackendoff (1992) analyzes the verb drink as having the conceptual semantic structure as that of a verb of caused motion; the theme moves along a path to a location inside an individual’s mouth. Notably, as shown in (27), the individual whose mouth the theme ends up in is represented twice in the semantic representation: once as the causer of the event, and one as the inalienable possessor of the location the theme ends up in, as indicated through the co-indexing of the relevant arguments.

One can see some motivation for a semantics involving change of location in the following, where attempting to cancel an entailment of change of location into someone’s body is infelicitous with a verb of ingestion.

Other work has argued for similar representations for the lexical semantics of ingestion verbs on the basis of other entailments associated with this verb class. For example, Krejci (2012) argues that ingestive verbs like eat are lexically reflexive, such that the causer argument is coindexed with an argument of a transitive digested-by relation with the theme argument such that the causer digests the theme.³ On Krejci’s approach, the event structure template of an ingestive verb like eat in the style of Rappaport-Hovav & Levin (1998) is provided in (29) below; here, an actor x acts in a food-manipulating way that causes some object y to potentially become digested.

While this differs from Jackendoff (1992), for whom the relevant result state involves a change of location in (27), what is crucial is that in both proposals ingestive verbs are lexically reflexive, encoded via a form of coindexation between the agent of the event and holder of the result state of the event. If verbs of ingestion involve coindexation of their agent with an argument of the result state component of its meaning, then we have the beginnings of an explanation for why these verbs resist subjectless presuppositions with again: even if their agents are introduced outside of the lexical VP, an argument within the lexical VP must be coindexed with the agent, and this conindexation persists within again’s presupposition.

The idea that particular verb classes track their external arguments via an indexation mechanism has independent precedent in the analysis of stative transitive verbs of emotion, such as love and respect due to Hale & Keyser (2002). On their analysis, stative transitive verbs are built out of a corresponding nominal form bearing an anaphoric index in a relational, prepositional sub-structure. Importantly, this index, which intuitively references the ‘possessor’ of the emotional state in question, must not corefer with the closer theme argument in (30) and must instead be bound by the external subject experiencer argument, what Hale & Keyser (2002) refer to as an obviative binding relationship. That is, Mary loves her children intuitively means Mary has her own love for her children and cannot be interpreted as Mary has her children’s love for her children.

Hale & Keyser (2002) argue for this representation by considering the alternation of verbs like love in an overt possession construction in its nominal form in (31a); here, the genitive pronoun referencing the ‘possessor’ of the emotion cannot corefer with the surface subject. Furthermore, with an intervening DP indirect object in a double object construction with give (31b), coreference is possible now only with the structurally more distant surface subject.

What’s more, as noted by Bale (2007), and discussed in section 2 above, this class of stative transitive verbs is similar to intentional verbs of killing and ingestive verbs in being infelicitous with subjectless presuppositions. Given this similarity, and the general cross-referencing of external arguments across these three verb classes, we take such an indexation mechanism to be the key property that explains the absence of subjectless presuppositions with again observed with these verbs. The challenge is to formalize this in an explicit compositional semantics that explains how these classes of verbs interact with again’s presupposition. We turn to this in the following section.

4 Analysis

5 Conclusion: On the compositional nature of agent entailments

We have argued here for an account of how verbs make reference to their agent (external) arguments, making use of evidence from the unavailability of subjectless presuppositions with again. Specifically, we argued that certain classes of verbs make reference to their agent arguments via an anaphoric index, which is bound a VP-external argument introducing head Voice that shifts the assignment function to one mapping the index to the argument it introduces. This semantic, binding-theoretic account is independently motivated by prior proposals that present evidence for the presence of reflexive-like behavior in the lexical semantics of verbs in the classes of interest. The account is able to explain the absence of subjectless presuppositions with the relevant classes of verbs, while nevertheless maintaining an analysis on which the agent (and other external arguments, such as experiencers) are introduced VP-externally by a higher functional head. The analysis further permits a unified understanding of the absence of subjectless presuppositions with again in the presence of VP-internal bound variables, including reflexive objects of verbs that otherwise permit such presuppositions.

Zooming out to the bigger picture, our analysis leads us to something of a middle road with regard to the compositional status of agents, and external arguments more generally. On the one hand, our proposal does allow verbs of certain classes to encode information related to their agent arguments in such a way that this has demonstrable semantic effects in interaction with other operators. This ability to interact with other operators, such as again, is reminiscent of, but ultimately distinct from, the fact that verbs may have entailments relevant to the agent of the event they introduce, of the sort discussed by Dowty (1989) and Wechsler (2005; 2020) (see also Müller & Wechsler 2014). For Dowty and Wechsler, agentive verbs must entail something about properties of their agents; for example, the difference between sing and whistle depends, among other things, on properties of the agent’s vocal chords during the event. A meaning postulate of the sort in (49) can capture this fact.

This information, however, seems to behave differently from the sort of verb-internal reference to the agent we have discussed in this article. In particular, it does not seem to interact with subjectless presuppositions with again in the same way: the agents may differ between asserted and presupposed singing events, for example, as (50) shows.

In this sense, then, the evidence we have discussed in this paper presents a stronger argument that verbs, at least those belonging to certain classes, must encode information about their agents in a way that is compositionally accessible to other operators, and thereby affect the felicity conditions those operators place on a sentence containing them.

On the other hand, however, we maintain that the presence of this information in the verb does not necessitate abandoning a Kratzerian approach to the composition of VPs with their external arguments via a VP-external functional head; if anything, the very existence of subjectless presuppositions with again seems to us to require such a compositional setup. Rather, if reference to the agent is encoded in an index on the verb, and argument-introducing functional heads can bind this index to the argument they introduce by shifting the assignment function their complement is evaluated with respect to, then the absence of subjectless presuppositions with verbs in the classes we have studied here can be accounted for without requiring that such verbs denote functions from their agent arguments. Indeed, on this approach, the agent, and external arguments more generally, can be introduced VP-externally in all cases, with the absence of subjectless presuppositions in certain cases arising from the presence of a necessarily bound index within the VP that again takes scope over.

In this way, our analysis occupies a middle ground between approaches that systematically sever agents from their verbs (Kratzer 1996) and those that argue that verbs must take their agent arguments as true arguments (Wechsler 2005; Bale 2007). Verbs of certain classes may track their agent argument via indices without denoting functions from that argument.

Funding information

Smith would like to acknowledge the support of European Research Council Consolidator Grant ERC-2017-COG769192, awarded to Andrew Koontz-Garboden. Yu would like to acknowledge partial support from the KU Leuven-funded C1-project Comparatives Under the Microscope (C14/20/041) awarded to Jeroen van Craenenbroeck, Guido Vanden Wyngaerd, and Dany Jaspers.

Acknowledgements

We thank Andrew Koontz-Garboden, Jens Hopperdietzel, Jialei Zhu, and the audience at the LAGB 2022 workshop ‘Agents: Grammar or Roots?’ at Ulster University, Belfast for their feedback on this work. We also thank three anonymous Glossa reviewers and Víctor Acedo-Matellán, thanks to all of whom this paper has been greatly improved.

Competing interests

The authors have no competing interests to declare.