Verbum – Analecta Neolatina XXVI, 2025/1
ISSN 1588-4309; https://doi.org/10.59533/Verb.2025.26.1.7
Abstract: In this paper, we propose a taxonomy of verbs based on their compositional behaviour regarding Agency. The taxonomy comprises three primary classes: agent-selecting verbs, non-agent-selecting verbs, and verbs that are lexically underspecified regarding Agency. We describe the semantic and syntactic characteristics of each class using evidence from corpora and considering the potential influence of coercive phenomena in semantic composition, which may be prompted by contextual cues such as agent-oriented or patient-oriented adverbials (intentionally, accidentally). Our analysis focuses on Italian, but the conclusions generally apply to other languages as well.
Keywords: agent, subject, semantic role, verb classes, context, coercion
Semantic relations were first introduced in generative grammar during the mid-1960s and early 1970s. According to the traditional framework, these relations are determined by the verb; a verb subcategorizes its arguments and assigns them a specific role (such as Agent, Patient, Experiencer, and so on). However, as Wechsler (2005b) noted, this perspective needs to be re-evaluated, as contextual factors may also influence how a participant in the situation described by a verb is interpreted. Consider, for example, the sentences in (1) where the subject of the verb colpire, whose English equivalent hit is classified as an agent-selecting verb in Dowty (1989), is interpreted as agentive in (1a) and non-agentive in (1b), depending on the nature of the direct object filler:
In other words, in (1a) and (1b), the direct object’s type assigns the semantic role to the Subject: if it is a ball, the Subject is intended as an Agent; if it is a crossbar, it is intended as a non-Agent.2
The goal of the paper is to establish a verb taxonomy based on the premise that semantic roles are not assigned exclusively by verbs. The taxonomy consists of three primary classes: agent-selecting verbs, non-agent-selecting verbs, and unmarked verbs. Agent-selecting verbs lexically specify an agent Subject argument; for non-agent-selecting verbs, the opposite holds, and unmarked verbs are lexically underspecified concerning the role they assign to their Subject referent.
It is not easy to define the class membership for each verb, as within our framework, besides assigning Agency in the case of unmarked verbs (as in the case of the direct Object in (1)), contextual cues may change a verb’s classification, due to semantic type coercion (Pustejovsky & Ježek 2008): this is the case, for example, of participant-oriented adverbials, specifically agent-oriented such as deliberately or patient-oriented, such as accidentally, which we examine below.
In the upcoming sections, we will first provide background information on the concept of Agency and address methodological concerns (Sections 2, 3, and 4). We will then proceed with analysing the semantic and syntactic characteristics of the suggested classes, which will be supported by studying empirical evidence from corpora (Section 5). In the analysis, we will consider the following contextual factors: semantic type of argument fillers, adverbial modification, and modification by purpose clauses. Finally, we will present our concluding remarks in Section 6. Our analysis focuses on Italian, but the conclusions are generally applicable to other languages. A few examples in English are used for illustration purposes.
In theoretical linguistics, starting from Gruber (1965) and Fillmore (1968), the notion of Agent has traditionally been defined in terms of the role a participant plays in the event expressed by the verb. At the syntax-semantic interface, a linking rule has been defined: if the semantic structure of a sentence contains an Agent (the animate entity that activates and controls the action), it normally corresponds to the Subject in the syntactic structure: the Object complement can never be an Agent (Salvi 1988). As discussed in Wechsler (2005b), there are two main versions of the participant role approach. In the first version, the verb is associated with an atomic relation and a list of arguments labelled with roles: for example, eat (agent, patient). In the second version, the verb is decomposed into a structure built from more basic relations such as cause, go, stay (Jackendoff’s 1990 localist approach), and become (Dowty 1991). In this latter case, semantic roles are derived from argument positions in the verb’s decompositional representation. For example, in the Role and Reference Grammar (RRG) framework (Van Valin & Lapolla 1997), the Agent is the first argument of the abstract operator DO (also informally known as big DO, originally proposed as an abstract higher predicate in Ross (1972)). This operator is necessarily associated with a logical structure containing the primitive do (or act), which is intended as the universal primitive for situations in which a participant does something, either intentionally or unintentionally. This is represented in (2) below, where the abstract operator DO is in capital letters, while the primitive do is in small capitals:
As regards the content of the notion of semantic role, Dowty proposes to analyse it as a cluster of entailments associated with the verb meaning. For example, “if the sentence ‘x builds y’ is true, then it is necessarily also true that x performs purposeful actions, that as a result of these actions an artefact y comes into existence, and so on” (Dowty 1969: 75). In this view, the Agent role type may be defined as the set of entailments that are common to all the individual semantic roles of the Subject argument of various verbs that are identified as Agent arguments, including:
Lexical entailments, as in (3), may be converted into properties, such as [±rationality], [±volitionality], [±sentience], and so forth.
It should be noted that defining what properties make up the concept of the Agent role is a topic of much debate in the linguistic literature (see Huyghe & Wauquier 2020, for an overview). For example, scholars have varying opinions on whether Agents need to be animate and intentional. The reason for this is that alongside intentional animate beings, there exist biological and chemical agents, natural forces (such as wind, see What the wind did was to blow the tree down, Cruse 1973), and instruments (such as sophisticated mechanical devices and softwares) that are not animate nor intentional, but once they are triggered by an Agent, operate independently and can perform actions alone, utilising their energy. The topic is particularly controversial today when abstract or artificial agents, such as computers, robots, and similar devices, act on our behalf, as when a computer calculates the best route (see What the computer is doing is calculating the correlation coefficient, Cruse 1973).
Different solutions have been proposed to address these issues, such as adding roles – next to the Agent role – like Effector (“the participant that brings something about, but there is no implication of its being volitional or the original instigator”, Van Valin & La Polla 1997)3 or Cause (“actor that may be animate or inanimate and that initiates the event, but that does not act with any intentionality or consciousness” Petukhova & Bunt 2008).4 Another solution is to view agency as a scalar concept, differentiating between prototypical and non-prototypical agents.5
Regardless of this debate, in all the approaches above, it is generally assumed that thematic relations are assigned by the verb, i.e., a verb such as eat assigns the role of Agent to its Subject.
The view that Agent is the role of an event participant instead of an inherent property of an entity is, however, not uncontroversial. For example, in ontological studies, an Agent is often analysed as an ontological category instead of a participant. In the DOLCE ontology (Gangemi et al. 2010), Agents are defined as “physical objects endowed with intentions, beliefs, and desires”. According to this latter interpretation, an entity does not need to be involved in an action and be assigned a role to qualify as an Agent. Instead, an Agent is defined based on its inherent cognitive capacities and, more specifically, its ability to have intentions. In this view, an Agent is an “intentional or cognitive entity”. This definition entails that every person is necessarily an Agent since agentivity (the capability of dealing with objects or states of the world) is an essential property of human beings (and according to many, of other animates like animals; see (22) below on this point).
For the purposes of this study, we will adopt a linguistic perspective and assume that Agency is a role that certain kinds of entities take on when they participate in specific types of events reported by verbs. Specifically, we will assume that verbs selecting for an Agent role in the Subject position are associated with the following mandatory lexical entailments:
In this view, a potential Agent becomes an actual Agent (i.e., takes on the Agent role) when he/she voluntarily performs an action driven by a well-defined purpose.
As referenced in Section 1 (ex. (1)), Wechsler (2005b) and Pustejovsky (2010), among others, pointed out that theories attributing the property of assigning Agency entirely to the verb are not satisfactory, as many verbs may be construed as agentive or non-agentive in the syntax. For example, Wechsler notices that the non-agentive verb disappear may be construed as agentive in the context of (5a), whereas the opposite holds for enter (5b):
Along the same line, the agentive It. verb lasciare ‘leave’ may be construed as agentive in (6a) and non-agentive in (6b).
Studying the coding of intransitive Subjects in Tsova-Tush (an ergative language spoken in the Caucasus, in which ergative marking generally conveys agentivity while the form with absolutive is always interpreted as non-agentive), Holinsky (1987) notes that intransitive verbs fall into three main classes.7 Some intransitive verbs take ergative (Agent) marking, while others take absolutive (Patient) marking, but most intransitive verbs can take either one, depending on how they are used. For example, in Tsova-Tush, the verb expressing the meaning ‘to lose one’s footing and fall’ falls into the third class. If one uses the ergative suffix, it means ‘slide.’ If, however, one uses the absolutive suffix, it means ‘slip.’ In other words, Tsova-Tush exploits its case marking system to signal that sliding is a controlled (agentive) event, whereas slipping is not.
Although there is consensus among scholars that the concept of Agency is best understood as a derived notion, there is still controversy regarding how the derivative account of Agency should be modeled, particularly concerning the interaction between semantic and pragmatic factors. Holinsky, for instance, argues that the evidence coming from Tsova-Tush suggests taking into account the interaction of the semantics of the sentence with general principles of conversation, according to which, for example, speakers tend to interpret a human doer as an agent, unless there is information to the contrary in the sentence.
In our discussion, we expand on previous studies that recognize that Agency can be determined either lexically or compositionally and introduce a new argument to model the possible construals of Agency. Specifically, we argue that semantic coercion phenomena supplement the standard compositional principle in building a sentence’s semantics. We define semantic coercion as the mechanism by which a word or expression imposes a specific interpretation on another word or expression with which it combines. Semantic coercion has been frequently used to account for grammatical verb–argument combinations that nevertheless exhibit a mismatch between the type selected by the verb and the type of the argument (Pustejovsky 2005), as in hear the bell, where the verb selects a sound as direct Object, and the bell is instead an artefact. This mismatch is resolved by positing that the verb imposes the selected type on the argument type, as illustrated in the transformation of the interpretation from ‘hear the bell’ to ‘hear the sound of the bell’ (Pustejovsky & Ježek 2008).8 In our framework, we examine how semantic coercion is induced by adverbial modification, specifically agent-oriented adverbs (adverbs that presuppose/are oriented towards the Agent) such as intenzionalmente ‘intentionally’, and patient-oriented adverbs (adverbs that presuppose/are oriented towards the Patient) such as inavvertitamente ‘inadvertently, unintentionally’.9
It is important to note that when coercion is integrated into the theoretical framework, defining class membership for verbs with respect to Agency becomes more complex because contextual cues, in addition to assigning/not assigning Agency with verbs underspecified for Agency, may trigger coercion in verbs which are specified either for Agency or for absence of Agency. For example, the patient-oriented adverb inavvertitamente ‘inadvertently, unintentionally’ may assign non-agency to the unmarked verb uccidere ‘kill’ (7a) or coerce the agent-selecting predicate azionare ‘activate’ from intentional to unintentional in (7b):
It follows that a verb exhibiting flexible behaviour concerning Agency in the syntax may be analysed in at least three different ways:
As we will see in Section 5, for verbs of class iii., Agency is always assigned compositionally. For verbs of class i., the agentive interpretation is the norm and the non-agentive one is coerced, while for class ii. verbs, the opposite holds. In the rest of the paper, we examine the properties of the classes in detail.
As clarified above, the goal of the research is to propose a three-output classification of verbs concerning Agency, and to contribute to characterising the semantic and syntactic behaviour of these classes based on empirical data and assuming that semantic type coercion may be active in meaning composition. This is achieved by examining a set of sentences obtained by querying the Italian Web 2020 (itTenTen20) corpus through the Sketch Engine online platform (Kilgarriff et al. 2004). In particular, we use the platform’s Concordance and Word Sketch functions. The Concordance function returns examples of use of the target word being queried. In contrast, the Word Sketch function returns a one-page summary of the word’s distributional behaviour (i.e., collocates and surrounding words) organized into grammatical relations, such as words that serve as an object of the verb, words that serve as a subject of the verb, words that modify the verb, etc. The data we obtained through this analysis contain three kinds of linguistic expressions:10
We verify whether the type of Subject is animate or not. When it is inanimate, we assume it is not an Agent.11 In the following, we present and discuss the results of our investigation.
In this Section, we examine in detail the syntactic and semantic properties of verbs belonging to the classes introduced above as they emerge from our empirical investigation, considering the theoretical framework outlined in Section 3.
Agent-selecting verbs are verbs that, in the Subject position, select an animate rational entity that acts volitionally with a goal. We will also refer to these verbs as verbs with lexical agents for current purposes. Examples are given in (9):
From a semantic point of view, these verbs denote an event that cannot come about spontaneously. That is, they require an Agent that performs the action. This is shown by the test in (10) a. and b. below, where both sentences are ungrammatical:
This semantic property (lack of spontaneity) affects agent-selecting verbs’ syntactic and argument flexibility in several ways. For example, unlike unmarked verbs (see Section 5.3), transitive agent-selecting verbs do not exhibit inchoative alternations (11a–b).
A lexical Agent can never be demoted or backgrounded except in passive constructions (12a–b).
Another syntactic constraint for verbs that lexically entail a volitional Subject acting with a predefined purpose, such as assassinare ‘murder’, is the impossibility of the instrument used by the Agent to surface as a Subject, contrary to unmarked verbs such as uccidere ‘kill’:
In (13), the instrument (bullet) can fill the Subject slot of uccidere ‘kill’ in (13b) because this verb does not lexically entail an Agent in the Subject position (not class i.), while assassinare ‘murder’ does (class i.).
Regarding adverbial modification, adding an agent-oriented adverb with these verbs is generally odd, as it adds information already lexically specified in the predicate.
Sometimes, however, the adverb is allowed and emphasises the intentionality already entailed in the verb, as for mentire ‘lie’ in (15):
The insertion of a patient-oriented/agency-cancelling adverb is usually deviant semantically, as it contradicts the information made available by the predicate, namely that the action cannot occur spontaneously:
In some cases, however, agent-selecting verbs do license patient-oriented adverbials. For example, installare ‘install’ and masticare ‘chew’ can occur in contexts such as in (17).
Under the analysis proposed here, the adverb in (17) coerces the interpretation of the predicate from intentional to non-intentional by cancelling the lexical entailment of Agency in installare and masticare.12
Mangiare ‘eat’ and bere ‘drink’ are generally classified as agent-selecting verbs. This holds also when the referent of the Object argument is non-conventional food for the Subject, as in (18), where the Subject’s participation is voluntary.
In (19), on the other hand, the situation seems different, as the event of drinking is accidental:
These constructions are frequently accompanied by patient-oriented adverbs such as accidentalmente ‘accidentally’, involontariamente ‘involuntarily’, inavvertitamente ‘inadvertently’.
According to Kittilä (2005: 388–389), in such cases, “the agent’s participation cannot per se be considered involuntary. One cannot eat or drink something completely accidentally; the given action is always volitionally instigated and controlled. The accidentality manifests itself only in that the event’s target deviates from what it was supposed to be, and the overall intentionality of the event can be regarded as somewhat lower”.
According to our account, the type of Object acts as a functor and coerces the Subject from volitional to non-volitional. Verbs like eat license an agentive Subject only when the referent in the Object position is an artefact made for eating. Note that by contrast, a manner of eating verb such as divorare ‘devour’ does not behave like mangiare (i.e., it cannot be coerced) and appears to entail Agency in all its uses in its ‘ingest’ meaning (class i.).
Agent-selecting verbs may also be found with purpose or benefactive clauses that specify the inherent goal of the Agent in performing the action, as in Luca corre regolarmente per mantenersi in forma ‘Luca regularly runs to keep fit’. With benefactive clauses such as per la moglie ‘for the wife’ and per il figlio ‘for the son’ in (21) a. and b., the intended purpose is implicit. For example, in (21a), Luca’s final goal is that his wife reads the book, while in (21b), it is that his child plays with the new toys.
Agent-selecting verbs may occasionally exhibit non-human entities in Subject position, such as substances in (5)–(6b). In such cases, we assume that to satisfy the verb’s selectional requirements, the Subject referent is interpreted/seen as an entity endowed with intentions and that a process of coercion occurs in Subject-verb composition. A special case is that of animals, as in (22). While classifying animal behaviour as intentional is conceptually intricate (Heyes & Dickinson 1990), animals’ intentional agency is increasingly substantiated in both the natural and social sciences. Under this latter interpretation, no coercion occurs in (22a–b).
Non-agent selecting verbs are verbs that lexically select Subjects that do not act volitionally towards a goal. They express events characterized by spontaneity. Examples are:
This is a very heterogeneous class ranging from patient-selecting verbs such as morire ‘die’, which denote events in which the Subject referent is inactive and undergoes a change of state, to verbs like partorire ‘give birth (to)’ and segnare (un gol) ‘score’ in (24), denoting events in which the Subject referent is actively engaged in the event that is taking place but cannot be considered an Agent as the event is “happening to them” (see Quello che è successo alle due donne è stato che partorirono nello stesso momento ‘What happened to the two women is that they gave birth at the same time’).
In (24a), the two women are engaged in the activity but cannot be held responsible for the outcome. Similarly, in (24b), the player instigates the action, but the result is not under his/her control. As we anticipated in Section 3, following Van Valin & Wilkins (1996), this role can be called Effector and defined as the dynamic participant doing something in an event.
As regards adverbial modification, the addition of an agent-oriented adverbial such as deliberately with these verbs, is generally odd:
There are some exceptions, as in (26), where the adverb volontariamente ‘intentionally’ (and perhaps the expression controllandolo ‘controlling him’) may be analysed as coercing the interpretation of the predicate cadere and its Subject from non-agentive to agentive.
Contrary to expectations, patient-oriented adverbs, such as accidentally and inadvertently, modify non-agent-selecting verbs quite commonly.13 In this scenario, the adverb emphasizes the lack of an Agent, which is already implied by the verb’s meaning.
As noted in Cruse (1973), an agentive-like interpretation can be imposed in context (i.e., coerced, in our terms) also through the presence of purposive constructions, as in (28):
Non-agent selecting verbs may display flexible behaviour concerning Agency, depending on the type of Subject. For example, in the context of (29a), sparire ‘disappear’ is interpreted as non-agentive, while in (29b) (see also (5a)), the default interpretation is agentive.
Under the analysis proposed here, sparire is a non-agent-selecting verb (class ii.)., which may be coerced to agentive by contextual factors, such as the presence of a human Subject as in (29).
Unmarked verbs are lexically underspecified for the role of their Subject referent. They express events that may or may not occur spontaneously. Because of this property, they exhibit systematic alternation between agentive and non-agentive interpretations; the context assigns one or the other. Equivalent terms are context-dependent or underspecified verbs. Consider again the verb colpire ‘hit’ in the following contexts:
Under the analysis we propose, the predicate is underspecified in the contexts in (30), and no coercion applies. As explained in Section 3, in (30a), both the player and the ball induce an agentive interpretation in the underspecified predicate colpire, whereas the crossbar in (30b) induces a non-agentive one. Finally, in (30c), the default interpretation is non-agentive. A storm is a self-propelled, natural event capable of independent motion that can cause an eventbut is not endowed with intentions.
Because of their underspecification, unmarked verbs such as uccidere ‘kill’ tend to admit both agent-oriented (deliberatamente in (31a)) and patient-oriented adverbs (accidentally in (31b)):
Finally, causative/inchoative verbs tend to fall into this category. As is well known (Haspelmath 1993), the event expressed by these verbs may be construed in the language as an event caused by an Agent or as an event that comes about spontaneously. An example is affondare ‘sink’:
In this paper, we have demonstrated that the Agent role is not solely assigned by the verb. We have proposed a three-fold classification of verbs related to Agency: agent-selecting, non-agent-selecting, and unmarked verbs. We have described these classes’ main semantic and syntactic properties based on empirical analysis of data from corpora. Although all these types of verbs can be interpreted contextually as having both an agentive and a non-agentive interpretation, we argued that only unmarked verbs (class iii.) are underspecified and assign the thematic role of Agent based on context rather than lexically. In the other cases (class i. and ii.), the non-agentive or agentive interpretation results from a coercion mechanism in the semantic composition between the verb, its arguments, and other contextual cues such as adverbial expressions. The analysis reveals that several verbs, commonly regarded as agent-selecting verbs in the literature, may display non-agentive interpretations. This highlights the need for a semantic role analysis approach that takes into account a more comprehensive characterization of contextual structure and compositional processes in the semantics.
We are thankful to two reviewers for their insightful comments on an earlier version of this paper.
Cruse, D. A. (1973): Some thoughts on agentivity. Journal of Linguistics 9: 11–23.
Dowty, D. (1989): On the Semantic Content of the Notion of ‘Thematic role’. In: G. Chierchia, B. H. Partee & R. Turner (eds.) Properties, Types and Meanings. Dordrecht: Kluwer Academic Publisher. 69–129.
Dowty, D. (1991): Thematic proto-roles and argument selection. Language 67: 574–619.
Gangemi, A., N. Guarino, C. Masolo & A. Oltramari (2010): Interfacing WordNet with DOLCE: Towards OntoWordNet. In: N. C. Huang et al. (eds.) Ontology and the Lexicon. Cambridge: Cambridge University Press.
Gruber, J. (1965): Studies in Lexical Relations. Doctoral dissertation. Cambridge, MA: MIT. [Gruber, J. (1976): Lexical Structures in Syntax and Semantics. Amsterdam: North Holland.]
Fillmore, C. J. (1968): The case for case. In: E. Bach & R.T. Harms (eds.) Universals in Linguistic Theory. New York: Holt, Rinehart and Winston. 1–88.
Haspelmath, M. (1993): A Grammar of Lezgian. Berlin: Mouton de Gruyter.
Haspelmath, M. (1993): More on the typology of inchoative/causative verb alternations. In: B. Comrie & M. Polinsky (eds.) Causatives and transitivity. Amsterdam & Philadelphia: John Benjamins. 87–120.
Heyes, C. & A. Dickinson (1990): The Intentionality of Animal Action. Mind & Language 5: 87–103.
Holinsky, D. A. (1987): The case of the intransitive Subject in Tsova-Tush (Batsbi). Lingua 71: 103–132.
Huyghe, R. & W. Wauquier, (2020): What’s in an agent? Morphology 30: 185–218.
Jackendoff, R. (1990) Semantic Structures. Cambridge MA: The MIT Press.
Ježek E. & R. Varvara (2015): Instrument Subjects without Instrument Role. In: H. Bunt (ed.) Proceedings of the Eleventh Joint ACL-ISO Workshop on Interoperable Semantic Annotation (ISA) at International Conference on Computational Semantics (IWCS 2015). Queen Mary University, London. 46–54.
Kilgarriff, A., P. Rychlý, P. Smrž & D. Tugwell (2004): The Sketch Engine. In: G. Williams & S. Vessier (eds.) Proceedings of the XI Euralex International Congress, July 6–10, 2004. Lorient: France. 105–111.
Kittilä, S. (2005): Remarks on involuntary agent constructions. Word 56(3): 381–419.
Petukhova, V. & H. Bunt (2008): LIRICS semantic role annotation: Design and evaluation of a set of data categories. In: Proceedings of the Sixth International Conference on Language Resources and Evaluation (LREC 2008). Marrakech: Morocco. 2830.
Pustejovsky, J. (1995): The Generative Lexicon. Cambridge MA: The MIT Press.
Pustejovsky, J. (2012): Co-compositionality in grammar. In: M. Werning, W. Hinzen & E. Machery (eds.) The Oxford Handbook of Compositionality. Oxford: Oxford University Press. 371–384.
Pustejovsky, J. & E. Ježek (2008): Semantic Coercion in Language: Beyond Distributional Analysis. Italian Journal of Linguistics 20: 181–214.
Ross, R. J. (1972): Act. In: D. Davidson & G. Harmon (eds.) Semantics of Natural Language. Dordrecht: Reidel. 70–126.
Van Valin, R. D. & R. J. Lapolla (1997): Syntax. Cambridge: Cambridge University Press.
Van Valin, R. D. & D. P. Wilkins (1996): The case for ‘Effector’ Case Roles, Agents and Agency Revisited. In: M. Shibatani & S. A. Thompson (eds.) Grammatical Constructions. Their Form and Meaning. Oxford: Clarendon Press. 289–322.
Wechsler, S. (2005a): Thematic Structure. In: K. Brown (ed.) The Encyclopedia of Language and Linguistics. 2nd ed. Amsterdam: Elsevier. 645–653.
Wechsler, S. (2005b): What is right and wrong about little v. In: M. Vulchanova & T. A. Åfarli (eds.) Grammar and Beyond – Essays in honour of Lars Hellan. Oslo: Novus Press. 179–195.
By default reading, we mean the most likely reading when no further context is provided.↩︎
It is possible that pragmatic implicatures and commonsense knowledge are at play in the interpretation of these examples; we will return to this idea below.↩︎
The term Effector was first proposed by Van Valin & Wilkins (1996) to characterise humans that activate an event involuntarily and/or with no intended goal. It roughly corresponds to the notion of ‘involuntary’ or ‘unintentional agents’ proposed in the cognitive-typological literature (see Haspelmath 1993 and Kittila 2003).↩︎
To our knowledge, no semantic role set in linguistics has introduced the notion of Artificial Agent so far, nor is such a role present in lexical resources that employ role sets to define frame/event participants, such as FrameNet and VerbNet for English.↩︎
Note that adopting a prototypical view for the Agent role is not the same as adopting proto-roles (Dowty 1991): only two roles are foreseen in the latter account (Proto-Agent and Proto-Patient), while in the former this is not the case.↩︎
Examples are from Wechsler (2005b).↩︎
A synopsis of Holinsky’s account may be found in Van Valin & Lapolla (1997: 118–119).↩︎
Concerning Agency, semantic coercion has been specifically exploited to account for the agentive interpretation of verbs such as kill with human subjects (subject-induced coercion, according to Pustejovsky 2012).↩︎
In our framework, coercion may account for several linguistic phenomena, including figures of speech that unlock the speaker’s creativity, such as personifications. However, we do not explore this perspective in this paper.↩︎
The Sketch Engine (Kilgarriff et al. 2004) is available at https://www.sketchengine.eu. The Italian Web 2020 corpus is an all-purpose Italian corpus covering the largest possible variety of genres, topics, text types and web sources. Data was downloaded in October–December 2019 and December 2020. The corpus consists of more than 12 billion words.↩︎
See, however, Ježek & Varvara (2015) for a discussion of Instrument Subjects coerced to Agents.↩︎
Van Valin & Wilkins (1996: 308) also observe that look at seems to co-occur with patient-oriented adverbials, suggesting that its Agent is not an Agent like that of murder. Compare: Mary accidentally looked at his neighbor’s test and was accused of cheating vs. *Larry inadvertently murdered his neighbor.↩︎
This behaviour contrasts with that of agent-selecting verbs, which tend not to license agent-oriented adverbs.↩︎