Metaphor forces argument overtness

3.1 Metaphor and overtness

Our central claim is that metaphors require overt expression of the metaphor-driving argument, covering Rai and Chakraverty’s (2020) “Type II” metaphors of verb-argument combinations, but going beyond these in covering any functor-argument combination irrespective of word class. Optionality of arguments is only found with literal uses.

Let’s look again at middle. Literal middles exist in spatially or temporally convex^[6] objects, e.g., rooms, tables, days or years, as seen in Examples (1)–(4).^[7] Now consider metaphorical uses:

Trouble in (11) and hiyaya ‘heart’ in (12) are not convex objects but respectively encode a state and the seat of thoughts and emotions, and so lack literal middles. This incompatibility of terms forces a metaphorical interpretation. In all such cases, we find the metaphor-driving argument expressed overtly.

What we never see in our data are sentences like If there is trouble it seems Jones is inevitably in the middle, where the covert argument of middle resumes trouble. We have illustrated metaphor-driven argument overtness here with nouns, but it occurs also in combinations of other parts of speech that force a metaphorical interpretation. Here (13) and (14) are examples from English involving above and arrive, contrasting with the optionality seen in the literal Examples (8), (9).

The hypothetical case where a metaphorical sense occurs without the corresponding metaphor-driving argument – e.g., arrive occurring without the overt expression of what is metaphorically arrived at – does not occur.

As far as we know, the only previous account of the phenomenon is given by Reinöhl (2016: 147–158) who describes the distribution of overt and covert arguments combined with relational nouns and adverbs in Indo-Aryan. While some existing work on zero arguments contains relevant contrast pairs, e.g., return (to the camp) versus return *(to the task) (Fillmore 1986: 102), the phenomenon of obligatory metaphor-driven argument overtness has passed unnoticed. One reason for this oversight seems to be that most studies (e.g., many of those quoted in Fn. 2) focus on the question of which functors (specifically, verbs) allow for zero arguments. In contrast, we ask the question: if a functor in principle allows an anaphoric zero argument, when must that argument be overt?

The literature on lexical semantics recognizes that more context is usually needed – and quite unsurprisingly so – to evoke less common interpretations of terms (e.g., Cruse 1986: 72–73; Löbner 2012: 74–77). For instance, Löbner (2012) briefly discusses the need for more context to facilitate shifts to figurative interpretations resulting from metaphor or metonymy. The phenomenon, however, is definable much more precisely and with a much wider scope than has been noted before. There is absolute or near-absolute resistance of metaphor drivers to being covert – across parts of speech and languages – no matter how activated.

3.2 A closer look at the literal and the metaphorical

At the heart of every metaphor is a conflict in literal senses. Terms have literal, “citation” senses, those that come to mind when asked about the word in isolation, and these are the ones that typically end up as the primary definition in dictionaries (the “established” senses in Cruse 2000: 108; see ‘minimalist’ approaches in Recanati 2003).^[8] One testable characteristic of literal senses is that they survive in emphatic reduplication, as in There was a mouse – a MOUSE mouse – sitting next to my laptop! (i.e., as opposed to a pointing device).^[9] Often, the literal senses are semantically richer than metaphorical ones (e.g., to arrive generally signifies that a particular endpoint is reached, but only the literal use involves the added component of spatial movement).^[10]

When two terms with conflicting literal senses come together in a functor-argument relationship – e.g., when the semantic type of the argument violates selectional restrictions imposed by the functor – this conflict can be resolved through metaphorical interpretation (Asher 2011). Language users resolve such conflicts by reanalyzing the terms to build a coherent interpretation. Combined with trouble, middle loses its reference to an actual spatial or temporal center. At the same time, a schema of convex containerhood is imputed to trouble (e.g., Boroditsky 2000; Lakoff and Johnson 1980). We refer to the argument in conflict with selectional restrictions and therefore coercing reinterpretation as the metaphor-driving argument (see Rai and Chakraverty 2020: 5–6 for references on selectional preference resolution accounts of metaphor).^[11]

Close examination of corpus data reveals that there are transparently literal and transparently metaphorical expressions, but also unclear ones, where one could argue for either classification. Importantly, however, argument optionality seems to be found only with transparently literal cases. Potentially ambiguous cases such as the natural spoken-language example in (15) are generally found in our corpus with overt lexical or pronominal arguments. While a restaurant can be conceptualized as a convex space, the point here does not seem to be that he is literally in the restaurant’s center, but in a situation where he can’t speak at ease; thus, in the middle here conveys immersion rather than reference to a specific center.

Metaphors are widespread, some more conventionalized and others more ad-hoc. As Lakoff and Johnson (1980) have famously emphasized, metaphorical language is not restricted to poetical or flowery ways of speaking, but pervades ordinary language use (see also e.g., Steen et al. 2010). In some cases, metaphorical constructions have a special pragmatic force; compare, for example, in the middle of trouble with in trouble. In many other cases, however, the metaphorical construction adds no special force, as in to arrive at a conclusion, to take control, the head of the company, warm light, etc. The ordinariness of these conventionalized metaphors is highlighted by Lakoff and Johnson (1980) to the point that they call them “literal” (e.g., 1980: 53, and elsewhere). While we take their point, we do not adopt their usage, but retain the traditional language sense of “literal” as contrasting with “metaphorical”.

It might be tempting to equate “literal” with “most frequent” (e.g., Levshina 2018). There are, however, many examples where non-literal uses compete in frequency with literal ones (see Cruse 2000: 199). Consider for example English nouns such as obstacle, goal, target, or foundation. The one respect in which frequency may play an important role is when the metaphorical sense of a lexeme over time becomes a literal one because the original literal sense has been lost or almost lost. An example of this might be the English noun set-up, which originally had a literal use meaning ‘An object set up or upright, an upright. Also spec. a stand or display at a carnival, etc.’ (OED)^[12] besides its metaphorical use ‘The way in which something is organized, arranged, or constituted’ (OED). In contemporary English, only the originally metaphorical meaning remains in productive use and may hence now be characterized as the new literal use. In such cases, we would predict that metaphor-driven argument overtness may not hold (anymore). We would further predict that it may already break down as the literal use declines.^[13] In other words, obligatory argument overtness affects cases where there is a productive relation between a literal and a metaphorical interpretation.

Before we move on, it is worth noting some phenomena that do not fall within the scope of the present study. Firstly, cases where a construction as a whole functions as an idiom are not relevant when they do not involve a semantic conflict between the literal senses of functor and argument. Thus, kick and bucket are semantically fully compatible under their literal senses in modern English, even if only folk-etymologically, and it is only the construction to kick the bucket as a whole that has a figurative sense.

We also do not address metonymy. Cases of metonymy do not force argument overtness in the way metaphors do. For instance, it is widely believed that modal verbs grammaticalize by way of metonymical, not metaphorical processes (e.g., Bybee et al. 1994). Modals can elliptically take up an already-mentioned argument – Will you help me? Yes, I will . – which contrasts with the obligatory expression of metaphor-driving arguments.

3.3 Alternative accounts of metaphor

The distinction between literal and metaphorical is often paraphrased as the distinction between spatial or “concrete” and abstract, but this characterization is misleading. Even though metaphors from the spatial domain to an abstract one seem common, there are many cases where we see a different, even reverse mapping (see Kövecses 2017: 16). For instance, in Currently the world is limited to existing on a completely flat, homogeneous surface that extends forever in all directions (Belote 2008: 30), spatial size is conceptualized in terms of temporal duration. Metaphors building on non-spatial concepts are common also in the realm of cross-modal metaphors. The temperature terms warm and cold can indicate voice qualities (among other things). In other cases, metaphors remain in the same general domain, e.g., from sound to sound in the example awakened by the thunderclap of a large branch falling on the roof.^[14] In our data collection, both from corpora and in the experiment we conducted, we made sure to not only include the better established metaphors from space to non-spatial domains, but also others.

A mapping from space to time – as a sub-type of mappings from space to non-spatial domains – is particularly often invoked. For example, take a verb like to come whose literal sense involves a reference to space (“to move to or towards a person or place”, OELD).^[15] Because of this, your day will come is a metaphor (see Boroditsky 2000 on the metaphoricity of temporal come-constructions). Some concepts, however, are equally literal across multiple domains, some across both space and time. One such example is middle, whose primary sense is defined in the Oxford English Learner’s Dictionary as “the part of something that is at an equal distance from all its edges or sides”,^[16] illustrated both by spatial and temporal examples (e.g., in the middle of the road, in the middle of the night etc.) (see also Reinöhl 2016: 148–149). In fact, literal uses of middle extend to other abstract entities that have inherent quantifiable scales, e.g., temperature or income.

As laid out above, we operationalize the concept of functor-argument metaphors more generally in terms of a clash in selectional restrictions, rather than in terms of traditional comparison-based or substitution-based approaches to metaphor (see Rai and Chakraverty 2020 for an overview). In terms of the actual interpretation of a functor-argument metaphor in communication, we furthermore subscribe to an interaction-based view that takes into account context, world knowledge and discourse participants’ perspectives and attitudes that are updated in real time, and we thus reject the possibility of a monolithic semantic interpretation. Our focus in this paper is not, however, on such interactional aspects, but on the structural consequences of the clashes in underlying selectional restrictions. While recent years have seen a rise in computational approaches to metaphor-detection, we here rely on manual annotation given that there is so far no proposed method that approaches 100 % of what coders intuitively regard as metaphorical (see Rai and Chakraverty 2020 for a detailed overview). Several computational approaches moreover involve a step of human coding, ultimately falling back on intuition, as is the case with the Metaphor Identification Procedure that is intended as an objective and reliable method (see Pragglejaz Group 2007). Since our interest here is in capturing all metaphorical cases in our datasets in order to test whether metaphor-driven argument overtness really does hold as robustly as we claim, we therefore rely on our operationalization as outlined above, where senses found in isolation, viewed as prototypical by native speakers, are the reference points for what is and is not metaphorical (see further detail in Sections 4 and 5).

3.4 Lexical semantics and discourse context

Many of the metaphor drivers we have encountered have inherently abstract meanings, e.g., trouble. However, there are terms which can represent either a concrete or an abstract sense. One example is hiyaya ‘heart’ in ex. (12), repeated as (16). In principle, ‘heart’ could designate either the physical organ or the seat of emotions and cognition. Context makes it clear that the term here refers to the heart as an abstract entity.^[17] Thus, whether a term drives a metaphor depends not solely on its lexical semantics, but on the use to which it is put in context.

Another important domain of such conceptual duality arises with humans, who can be talked about as bodies or as abstract entities (e.g., social agents). Since humans spend a good deal of time talking about themselves and others of their kind, references to people constitute particularly common cases of concrete-abstract ambiguity (see also Reinöhl 2016: 158–160).

The importance of discourse context naturally extends to pronominals. In (17), we have an anaphoric reference to contradictions which clashes with the literal meaning of middle, and so the pronoun drives a metaphorical rather than a literal interpretation.

Whether or not an argument drives a metaphor thus depends on its interpretation in context. The crucial role of context in triggering overtness underscores that metaphor-driven argument overtness is a phenomenon of communication. It arises out of the effort that speakers make to model listeners’ minds, an insight informing our psycholinguistic model in Section 6.4.

3.5 When we already speak in metaphorical terms

If a functor takes two (or more) arguments, either one could drive a metaphor, i.e., be the first argument in the clause to establish the metaphorical interpretation. If the first argument in the clause does (e.g., the subject), then subsequent arguments are not needed overtly, since we have already been pushed into the metaphorical realm.^[18] Consider the following example:

Truth, as an abstract entity, cannot in fact literally, spatially lie anywhere. Since it pushes lie in the middle into metaphor, there is then no need to supply an additional metaphor-driving argument to middle. In more formal terms, since the first argument clashes with the selectional restrictions of the functor’s literal sense, the latter is reinterpreted with adjusted selectional restrictions in accordance with the semantics of the first argument. Since the second argument does not clash with these modified restrictions, it can remain covert (see Ellison and Reinöhl in prep., for a detailed formal account employing graphical semantic representations). We refer to such cases for short as the “metaphor-post-driving” or just “post-driving” condition (see Section 5.2), where argument optionality (instead of obligatoriness) is expected.

Given Example (18), one might think that the non-obligatoriness of the second argument is restricted to frequent collocations. However, the same phenomenon is found in Example (19) which does not involve a frequent collocation. The metaphor-driving constituent diseases of joints, muscles and gut is overt, while the container of middle remains covert.

One special case occurs with imperatives. They have an in-built argument concerning (usually) 2nd person speech-act participants. This paradigmatic zero remains even when these act as metaphor drivers, e.g., Drop it! or Leave it!. Given the metaphor-driving zero, the second argument can also remain covert in some languages where the syntax allows it, e.g., Tamil Viḍu!.^[19]

We return to the interaction between multiple arguments to one functor in the context of our corpus studies and of our sentence-completion experiment.

3.6 Outside information structure

Whether we express a referent covertly or overtly with a full noun phrase or pronoun traditionally lies within the purview of information structure. Indeed, this choice is often determined by factors such as how activated or topical a given referent is. However, information-structural constraints on overt or covert expression are restricted to the arena of literal language. In the realm of metaphor, there is no optionality.^[20]

We suspect that scrutiny of the literature on reference tracking will reveal cases where overt expression has been attributed to causes such as low topicality, while it was in fact forced by metaphor. We are aware of at least one such study (David 2016: 83–84) which explicitly attributes the necessary overtness of at the conclusion in We arrived at the conclusion to the argument’s insufficient topicality, rather than to the inherent semantic clash between functor and argument.

Arguments can only drive metaphors within their own simple clause (For spoken language, further research will likely revise this generalization to the intonation unit). A metaphor-driving element cannot combine with a functor in a different clause, even if immediately adjacent. We have already seen this phenomenon above (e.g., (11), (13), (14)) and it is also illustrated by the next two Examples (20) and (21). The referents in question are pronominally resumed in the immediately following clauses (for an example of resumption by means of a synonym, see (25)). Both examples illustrate that the domain of obligatory argument overtness is not the sentence (or the complex clause), but the simple clause.

While information-structural factors are irrelevant for the overtness of metaphor-driving arguments, highly activated referents best exemplify the contrast between literal and metaphorical language. It is precisely with given referents that zero expression may become possible in literal language, while there is no such effect in the case of metaphor. Throughout the paper, we therefore primarily choose examples for illustration which involve a highly activated referent, which is pronominally resumed. This example type also forms the basis for our sentence-completion experiment discussed in Section 5.

4.1 Indo-Aryan: spatial nouns and their possessors

Reinöhl (2016) studies the emergence of postpositions and postpositional phrases across the three millennia of attested Indo-Aryan history. She observes that spatial nouns over time occur more and more frequently with overt possessors and in metaphorical contexts, suggesting a connection between the two phenomena. On the topic of Sanskrit madhye ‘in the middle’ developing into Hindi mẽ ‘in’, she observes that metaphorical uses (subsumed within “non-basic grounds”) require overt possessors:

[A]ny usage of MADHYE relating to a non-basic ground, that is to a ground which lacks an intrinsic centre, requires its overt expression. Only MADHYE expressing IN MIDDLE in relation to a spatial or temporal ground with an intrinsic centre may be used without an overt possessor. (Reinöhl 2016: 152)

She goes on to list two types of evidence for this claim:

First, instances of MADHYE in isolation tell us about the possibilities when covert reference is possible. Second, instances where MADHYE takes a pronominal possessor give clues about where a ground which is inferable from context is nevertheless expressed overtly, even though covert reference is a syntactic possibility. (2016: 152)

Her results show that argument overtness depends on the literal-metaphorical distinction. Literal arguments allow for covert expression, while metaphor drivers require overt forms. This leads to the phenomenon of pronominals being overt even where referents are highly activated.

The study is based on a corpus consisting of the bulk of Vedic Sanskrit texts and a substantial number of texts in different varieties of Middle and Early New Indo-Aryan. All 104 cases of madhye occurring without an overt possessor refer to a physical or temporal container as illustrated in (22) and (23) (see also Example (3b)).

In metaphorical constructions, possessor arguments are overt, even when given, such as the 1st person speech act participant cast as a social entity in (24):

The Indo-Aryan evidence unequivocally supports obligatory metaphor-driven argument overtness. While the corpus is larger than many historical corpora of other languages or language families, it is small in comparison to the corpora of some modern world languages. Thus, we now turn to English.

4.2 English

We explore the obligatoriness of metaphor-driving arguments in English across several construction types. These include relational nouns taking possessors (Section 4.2.1), adverbs selecting spatial arguments (Section 4.2.2), and verbs of motion and posture having goals or locations (Section 4.2.3). The data is taken from the British National Corpus (BNC),^[21] which contains about one hundred million words. It covers modern British English in a broad variety of genres, spoken and written, and was compiled between 1991 and 1994.

The full BNC is available for download free of charge. We incorporated it into a MySQL relational database and developed SQL and PYTHON scripts to search for target words in tailored contexts defined by part-of-speech and punctuation. For each target word and context, we made a random selection of 100 examples for the study. All examples are provided in the supplementary materials.

Examples were classified into “literal” or “metaphorical” (in the first step, without taking into account which argument is the metaphor driver) by three research assistants and the authors based on our understanding of this difference outlined in Section 3.2, which builds on our operationalization of literal as the primary definition sense given for a word in isolation. If there was disagreement between raters, we coded examples as “metaphorical”, in order not to overlook potential counterexamples for our hypothesis (see also fn. 23). A residue of cases was marked as “unclear” where the context did not disambiguate literal from metaphorical (potentially including, but not limited to, twice-true metaphors, see Cohen 1976) or where it was hard to pin down the exact meaning. For example, in (25), arrive can be taken spatially or non-spatially and both interpretations are felicitous. The talks can be understood either as an abstract entity or literally as standing for the location where they take place (by metonymy).

A fourth category “other” was used for the few cases where automatic filtering failed to select only cases of relevance to the study, as in the following example with arrive at not before punctuation:

In very few cases in data from spoken conversations, instances are classified as “other” because the construction type cannot be determined with certainty, for example, when parts of the utterance are inaudible.

To assign an instance to one of the four categories, we worked with the immediate context of 128 words and with punctuation symbols after the target word. When more information was needed, we made use of the larger context accessible in the online version. Appendix A explicates the decisions for each target word individually, with several examples for illustration (As for slight differences in presentation between the downloaded and the online version of the BNC, please see details in Appendix A). Figure 1 gives an overview of the results which we discuss in detail in the next Sections; the exact queries can be found in Appendix A.

Figure 1:

BNC queries and distribution of results across the four classifications: “metaphorical”, “literal”, “other” and “unclear”.

For each target functor, we contrasted occurrences before nominal expressions (“with”, i.e., before nouns, adjectives, articles, or pronouns) with occurrences before punctuation (“without”). The former indicates the presence of an overt argument, while the latter its absence. Note that all cases of covert arguments rated metaphorical occur in sentences with a preceding, overt metaphor-driving argument in the same sentence, i.e., our “metaphor-post-driving” condition. We discuss these results in more detail below.

For all part-of-speech combinations, we trialed several more items than the ones seen in Figure 1 and selected only ones productively occurring in both literal and metaphorical uses. For some of the trialed target words, we found no optionality under a literal interpretation, and thus there is no contrast with obligatoriness forced by metaphor. For example, while fall from occurs literally and metaphorically (e.g., fall from a tree vs. fall from grace), it requires an argument in any case (John was climbing a tree. Suddenly, he fell *from). The bare verb likewise does not create meaning-invariant optionality, because it does not necessarily have a semantic argument slot for a source. Thus, John was climbing a tree. Suddenly, he fell. can mean that John fell or slipped on a branch, rather than that he fell from the tree.

4.3 Vera’a: verbs and their arguments

Research on verbal arguments in particular has traditionally focused on the impact of information-structural factors on their overt or covert realization (e.g., Huang 2000; Matić et al. 2014; van Craenenbroeck and Temmerman 2018). However, the following study of the Austronesian language Vera’a, described in Schnell (2011), shows that verbal arguments, like those of English verbs of movement discussed above, are also subject to metaphorically conditioned obligatoriness.

Vera’a lends itself to this study because substantial amounts of the spoken language corpus have been annotated for zero arguments (Schnell 2015).^[24] We investigated all annotated missing arguments in the text as1, regardless of semantic role or syntactic function; these came to 94 zero arguments from a total of 589 arguments, occurring in 213 clause units. While adding important, non-Indo-European evidence, this corpus study is explorative and awaits future validation based on a larger corpus sample. It here serves to demonstrate the methodology of how metaphor-driven obligatoriness can be tested in lesser-studied languages.

The zero arguments combine with verbs across a range of semantic fields, in particular verbs of movement, handling, speaking, and possessing, as illustrated in the examples below.

None of the zero arguments seem to occur with metaphorical meaning. We certainly do not claim a general ability to decide between literal and metaphorical senses in a language for which we lack native intuitions, even while metaphor has been shown, e.g., by grammaticalization studies (e.g., Sweetser 1990), to serve as a strategy across languages. However, all the examples with zero arguments in the Vera’a corpus involve basic actions such as ‘to take’, ‘to go’, or ‘to say’, so it seems safe to attest a literal sense in these cases. A literal interpretation is corroborated by a comparison of the glossed senses on the word level with the free translations of the whole utterances. Moreover, context showed literal interpretations to be coherent in each case. Stefan Schnell (p.c.) agrees with our assessment of these uses as literal.

In contrast to the optionality observed in examples such as (36) and (37), there is evidence that the argument is generally overt in cases that are good candidates for metaphor such as (38) and (39), an assessment shared by Schnell (p.c. and see Schnell 2011: 36 on the example quoted in (38)). Across all Vera’a texts, there are no comparable metaphorical uses of these verbs where the metaphor-driving argument remains unexpressed.

The same verbs le and mom found in (38) and (39) can occur without overt second argument if the latter is activated and when the meaning involves a literal action of transfer, as illustrated in the two following Examples (40) and (41).

In summary, then, this section has shown that evidence from corpora of Indo-Aryan, English and Vera’a show absolute support for metaphor-driven argument obligatoriness across a range of part-of-speech combinations.

5.1 Design

We conducted an online sentence-completion experiment in which English native speakers chose whether or not to provide functor nouns with overt arguments. The experiment explored how arguments were provided for 13 target functors. Framing each functor once as literal in a suitable context, and once as metaphorical using a different context, we constructed 26 distinct questionnaire items. Each item consisted of an initial context sentence which primes the argument, typically in sentence-final position. This is followed by a sentence containing the target functor. Potential arguments to the functor were provided in a drop-down menu, allowing the choice between a covert or an overt, i.e., pro-form, argument. For some target words, more than one type of pro-form is available in natural language; for this reason, some questions offer a total of two, and others three completion options.

An example of a target word in a literal context is shown in (42) with three options for completion, and the corresponding metaphorical case is shown in (43). The order of completion options was randomized for each participant.^[25]

The 13 metaphorical framings were split into two groups. In six sentences, the targeted argument is the driver of the metaphor. In this condition, an example for which is (43), we expect consistent argument overtness. In the remaining seven, a prior argument in the clause – the subject – acts as a metaphor driver, so that the targeted argument becomes what we call a post-driving argument. In these cases, we would predict covert arguments to occur as discussed in Section 3.5. In the following example, His relationships drives the metaphor, so our prediction is that fodder does not have to be provided with an overt argument:

Participants were randomly assigned different versions of the questionnaire. Each target functor occurs only once in each version – in a literal or a metaphorical frame. The different questionnaire versions correspond to three experimental conditions. The first two conditions are blocked in the sense that they contain either only literal frames (Condition 1) or only metaphorical ones (Condition 2). Condition 3 is mixed, containing similar numbers of both literal and metaphorical items. To rule out target word-specific effects, we created two versions for Condition 3, resulting in four questionnaire instances in all. The two questionnaires for Condition 3 were counter-balanced: a target functor framed metaphorically in the first was framed literally in the second, and vice-versa. For each participant, the 13 items of their questionnaire version were presented in a randomized order. The breakdown of the conditions is shown in Table 1.

The combination of mixed and blocked conditions allows us to check for inter-question structural priming. If structural priming was a problem, we would expect differences in the distribution of responses for the same question posed in a blocked versus mixed condition.

We collected metadata from the participants about their native variety of English, in case this might be a factor influencing the results. We also asked for their competence in other languages. Participants were given the option of supplying their gender and age. After the experiment, participants could comment on their experience, and we discuss some of their remarks below. The experiment was disseminated via institutional and private social media accounts.

5.2 Results

Our questionnaire was completed by 259 people. Of these, 9 were excluded as the participants indicated in the metadata that they either had been, or may have been, at a talk by the authors on the topic being investigated. The remaining participants came from a range of English dialects: 99 were native speakers of Australian English, 63 of a North American dialect, 60 of a dialect from the UK or Ireland, 18 of South African English, and 7 of New Zealand English, with other dialects of English sharing the remaining 3. The majority of participants (185; 74 %) identified as not speaking another language natively or native-like. Female-identifying participants made up 65 %, male-identifying as 33 %, with 1 participant each identifying as trans, non-binary, and as agender; 3 participants did not volunteer a gender. The mean age of the 227 participants who volunteered their year-of-birth was 42 (standard deviation 14.3).

Because participants were randomly assigned a questionnaire, we were not guaranteed identical numbers in each. Thus there were 67, 53, 70 and 60 respondents across the four conditions respectively. A basic statistical analysis of the blocked and mixed conditions showed no significant difference in the covert/overt responses overall, although it is near significant (χ² = 3.3856, df = 1, p = 0.06577). Breaking these results down by conditions shows larger χ² for metaphor-post-driving (χ² = 2.3366, df = 1, p = 0.1264) and literal (χ² = 2.4226, df = 1, p = 0.1196) than metaphor-driving conditions (χ² = 0.92369, df = 1, p = 0.3365), suggesting that these are more responsive to priming conditions.

The primary descriptive results appear in Figure 2 (see Appendix B2 for all results). The rows correspond to the three conditions: the top is the metaphorical condition where the targeted argument is the metaphor driver, the second where an earlier argument (the subject) acts as metaphor driver, so that the questioned argument is a (metaphorical) post-driving argument. The third row corresponds to the literal condition. The columns correspond to the two types of sentence completion: one with an overt or one with a covert argument. The crucial number of metaphorical uses without an overt argument – which our theory would predict to be zero – are shown in the top left corner.

Figure 2:

Descriptive results of sentence-completion study.

The results in Figure 2 show strong support for overtness. For literals – whether in a mixed or blocked condition – more than half of the responses drop the argument. By contrast, when the argument acts as metaphor driver, the overt choice is almost exceptionless. When another argument drives the metaphor, we find an intermediate level of argument overtness, consistent with our expectations for post-driving arguments.

Since there are no relevant differences between the blocked and mixed conditions, we aggregate the results from these conditions in what follows. Figure 3 contrasts the proportion of overt arguments we found in literal constructions compared to those where they act as metaphor drivers. The contrast is striking: only 41 % overt arguments in the literal cases, and over 99 % in the metaphor-driving cases.

Figure 3:

Literal and metaphor-driving arguments (in percentage).

Figure 4 adds the post-driving argument condition, in which participants preferred overt arguments 70 % of the time.

Figure 4:

Literal, metaphor-post-driving and metaphor-driving arguments (in percentage).

These results support the claim that metaphorical constructions require, rather than just prefer, overt drivers. Only when a preceding argument acts as driver – e.g., when the subject has already pushed us into the metaphorical realm – can the targeted argument be omitted.

While the results shown in the bar graphs are clear, we also evaluate them statistically. The question we wish to answer is whether the same process could have given rise to the overtness frequencies for all three categories: literal, post-driving and metaphor-driving, or whether the near-absolute overtness of metaphor-driving arguments reflects an entirely different process. In other words, the question is how strongly does our experimental evidence support metaphor-driven argument overtness.

We evaluate the strength of evidence using a simple model in which each category has its own value for the likelihood of responding with an overt argument. For a literal target, the likelihood of using an overt argument is l, for a post-driver target the likelihood is p, and for a metaphor-driving target the likelihood is m. We assume that the prior distributions – that is the probability associated with different values before considering any data – are flat for l, m and p. We then evaluate how strongly our evidence supports one hypothesis over the other. We first compare the values of primary interest, i.e., the literal category l and the metaphor-driving one m. We explore the strength of evidence that they differ by more than our criterion for ‘sameness’ – being within 0.05 (5 %) of each other.

The log₂ Bayes Factor favors LM-Different over LM-Same with a strength of 662.7. Kass and Raftery (1995) define log₂ Bayes Factors above 7.2 as indicating very strong evidence for one hypothesis over the other. Given the logarithmic scale, 662.7 – corresponding to an odds ratio of 3.19 × 10¹⁹⁹ – indicates extreme certainty that Hypothesis LM-Different is correct, i.e., that metaphor is a distinct and powerful factor forcing argument expression.

These results confirm that we have very strong evidence that the distribution of arguments in the literal and metaphorical cases reflect substantially different distributions, and hence are the result of different underlying processes.

Parametric chi-square statistics offer us similar conclusions. Pairwise comparison of the overtness/covertness distributions give us likelihoods of the data for metaphor-driving and literal having been generated by the same distribution as very low (χ² = 708.59, df = 1, p ≃ 0).^[26]

Just as LM-Different and LM-Same were two hypotheses about the proximity of the likelihoods of overtness in literal and metaphor-driving arguments in the experiment, we can define hypotheses LP-Different and LP-Same about the difference between the overtness likelihoods of literal and post-driver arguments, and similarly PM-Different and PM-Same for the difference in overtness likelihoods between post-driver and metaphor-driving arguments in the experiment. The log₂ Bayes Factor in favor of LP-Different over LP-Same and PM-Different over PM-Same are 121.2 and 197.9 respectively. The corresponding odds ratios are 3.0 × 10³⁶ and 3.8 × 10⁵⁹ respectively. The data provide very strong evidence (in the Kass and Raftery [1995] sense) that the effect strength – the difference between overtness likelihoods – is substantial. The chi-square comparisons of these conditions are also overwhelmingly significant (χ² = 187.0, df = 1, p ≃ 0 and χ² = 238.4, df = 1, p ≃ 0). The base R distributional functions for the chi-square distribution were unable to distinguish p from 0, in all comparisons between conditions.

The above evidence has shown that there is very strong evidence for a difference between the distribution of arguments found in literal and metaphorical constructions where the argument is a metaphor driver. Furthermore, the effect size is very large.

5.3 Discussion

Our hypothesis that metaphor-driving arguments are overt is borne out by the results of the experiment. The metaphor driver is expressed in more than 99 % of the cases, even though the referent was always strongly primed. The few aberrant cases, 6 out of 716, are spread across stimulus items (1× bottom, 3× middle, 1× warmth, 1× view), participants, varieties of English, and both blocked and mixed conditions, with no evidence for any specific determining factor.

Given the unnaturalness of this type of experiment, these results by far surpass our expectations. In view of the near-absolute robustness of the effect seen here, paired with the absolute robustness across the corpus studies, the few “aberrant” cases might not indicate that the phenomenon is not categorical, but rather that something else is going on. Perhaps the most straightforward explanation is performance error: one can easily imagine clicking on an unintended sentence completion in the drop-down menu. However, at least two other explanations are possible. There is the artificiality of asking someone to complete a sentence that someone else started. When a speaker expresses a message of their own, semantic coherence is more likely to be sustained. In experimental set-ups of this kind, participants may not always fully process the priming material before selecting a completion. This may especially be the case here because the context provided is minimal.

Another possible explanation can be found in the fact that, in four of the six cases, a literal interpretation is marginally accessible if additional context is imagined, which could account for the argument covertness. Thus, for the target words bottom and middle, literal interpretations are imaginable despite the intended metaphorical framing. For instance, providing the pronominal argument supports the metaphorical reading in Something weird is going on there. I just can’t get to the bottom of it. However, the argument-less version Something weird is going on there. I just can’t get to the bottom. forces a literal reading, where e.g., someone is trying to reach the bottom of some sort of a container, or a swimming pool that is unexpectedly deep. Such literal interpretations are not available for the other two target words warmth and view for which the two remaining instances of aberrant cases are found.

Given these potential confounding factors, it is surprising that the number of aberrant answers is so vanishingly small. Thus, we conclude that the near-uniformity of the overall results reflect the profound robustness of metaphor-driven obligatoriness.

The astute reader might wonder if a preference account could explain our experimental results. Under this account, there is not a hard requirement for the metaphor-driving argument to be overt – this option is just preferred. However, two kinds of evidence speak against this explanation. In the first case, our experimental results are exceptionally close to categorical, meaning that the preference is very strong. The second kind of evidence is what we saw in the corpus results. In spontaneous productions found in the BNC, we found no clear cases of covert metaphor-driving arguments. So, if the overtness of metaphor-driving arguments is a preference, it is a preference that enjoys very strict application in language production. When a preference is universally applied, how can we distinguish it from a hard rule? However, see our previous discussion of potential grey areas due to, e.g., lexicalization.

In the cases of literal use, we find a rate of 41 % overt arguments; a mix of covert and overt arguments conforms to the expectations outlined in Section 2. Differences of overt expression between target items (see Appendix B2 for details) may reflect factors from realms such as activation or animacy, to the extent that the small contexts given in the experiment provide such information. For instance, the number of overt argument encodings is relatively high with the functor edge. A possible explanation is that the referent to be resumed (crevass) is slightly further away, coming a little earlier in the context sentence in comparison to other stimulus items (see Appendix B1 for the full set of stimuli).

Post-driving arguments also can be covert or overt – which matches our prediction (see Section 3.5 above). The higher overtness rates in comparison to literal uses may reflect a dependency of post-driver argument overtness on the type of the metaphor driver. Thus, inherently abstract entities, such as truth, may better facilitate the covertness of a post-driver argument than inherently ambiguous terms, such as body parts or humans (see Section 3.4). If the driver is lexically ambiguous, the speaker may help the listener reach a metaphorical interpretation by overtly resuming a post-driver argument, since argument covertness is normally linked to literal use. Besides questions of lexical ambiguity, it is also conceivable that the difference between pronominal and a lexical encoding of the metaphor-driving argument may influence subsequent argument overtness.

The experimental results fully corroborate the results from the corpus analyses. It is noteworthy that these results match those of the natural language of the corpora despite being produced in less-than-natural conditions.

The results are particularly noteworthy in light of the inability of speakers to introspect and identify the impact of metaphorical language on what must be expressed. None of the comments submitted with the questionnaire indicated any understanding of the pattern. On the contrary, several participants highlighted their puzzlement over why they sometimes felt like supplying, and sometimes did not feel like supplying the argument in question. It is fascinating to see how the effect, being so robust, utterly escapes speaker consciousness. Here are some of the comments:

Other speakers mention intuitions without any insight into where these come from.

In conclusion, our language production experiment with English native speakers yields remarkable support for our claim that metaphor drivers must be overt. Strikingly, the effect raised no conscious awareness whatsoever on the side of the speakers. This explains why some members of our audiences at conference presentations claim that they might accept covert metaphor drivers, or even use them. Since they struggle to accurately access the factors determining their choice in natural production settings, their ability to acccurately predict what they might do is poor.

6.1 Other triggers of argument overtness

Metaphor is not the only trigger for frequent or obligatory argument overtness. We discussed syntactic constraints in Section 2, such as those on core verbal arguments in English. We have also mentioned information structure-related biases, where e.g., the lack of activation of a referent, or the need for a contrastive focus, promote overt expression, even if zero anaphora is a syntactic option. There are two ways of thinking about how the different constraints – syntactic, semantic, and information-structural – interact. Either we describe their interaction about an individual instance of a construction use or we focus in general on what determines argument overtness for a particular construction type in a language.

In the first scenario, at least one of the three sources of overtness tends to apply for the argument to be realized in that way. For instance, the English subject must normally – for syntactic reasons – be overt even when the construction is literal and even when the referent is highly activated. It does not affect the outcome if more than one condition applies, e.g., when the English subject is also a metaphor driver. The same applies to metaphoricity and information-structural biases as determining factors. Any one of them, or multiple ones in concert, may trigger argument overtness.

When generalizing over a construction type in a given language, a nested picture emerges. Thus, information-structural biases are only testable factors for argument overtness in the absence of a metaphorical or syntactic constraint. Metaphoricity, in turn, is only a testable conditioning factor in the absence of a syntactic constraint.

6.2 Argument overtness and verbal-complement overtness

While we are not aware of any study on metaphor-driven obligatoriness apart from the remarks in Reinöhl (2016) and subsequent work in Ellison and Reinöhl (2022), a related phenomenon has been brought to light by Rosemeyer and Grossman (2017). Investigating the grammaticalization of a Spanish auxiliary construction expressing anteriority, the authors note that acabar ‘to finish’ was in Old Spanish at first only accompanied by verbal complements expressing unexpected or “informative” events. They illustrate this phenomenon with the following English example: Eating is often unexpressed in a context such as John finished the pasta, as it is the event associated by default with pasta. In contrast, instagramming would be more likely to be expressed overtly, given that it is not the default action associated with food. Through “overtification” accompanying a process of grammaticalization, uninformative events also come to be more frequently expressed.

Rosemeyer and Grossman (2017) thus uncover the influence of “informativity” or “unexpectedness” on whether or not the verbal complement to acabar ‘finish’ is expressed. As with our selectional-restrictions approach, the “default” semantics (Rosemeyer and Grossman 2017: 525, drawing on Pustejovsky 1991, 1995) determine what is and what is not a natural match. The naturalness of the match in turn governs argument expression.^[27]

At this point it is unclear whether informativity-driven overtness – if we may call it that – applies as an absolute rule or whether it is sensitive to activation. Rosemeyer and Grossman (p.c.) checked 50 instances of transitive acabar without an overt verbal complement, none of which involved an informative event. However, it seems imaginable that Speaker B’s response in Example (54) would occur in a situation where the interlocutors are engaged in instagramming food, and Malte Rosemeyer agrees (p.c.). In other words, informativity-driven overtness may be sensitive to activation in contrast to metaphor-driven overtness.

Certainly, we may be misjudging the naturalness of (47), just as some of our audience misjudged the possibility of dropping metaphor drivers. If our intuition is correct, there is an important difference between the two phenomena. When comparing Speaker B’s response in (47) with the second sentence in Example (48), we believe that the former shows meaning invariance, i.e., it paraphrases No, I only just finished instagramming the pasta. By contrast, the sentence Finally, he arrived. in (48) does not show meaning invariance. Without an argument, the sentence is interpreted (possibly with a processing delay, see below) as the subject referent arriving at a contextually retrievable physical location; the metaphorical interpretation is not selected.

In summary, metaphor-driven and informativity-driven overtness appear to be closely related phenomena. Clarifying how the two relate to each other is an exciting direction for future research. Our findings lead us to believe that sensitivity to activation is plausible with informativity-driven overtness, but not with metaphor-driven overtness. This would suggest different underlying psycholinguistic mechanisms. For example, metaphor-driven overtness may be triggered in concert with lexical retrieval, as soon as a clash in default interpretations is identified. Informativity-driven overtness, on the other hand, might arise in the combination of lexical structures with contextual information. Finally, it also remains to be seen how the verbal complements discussed by Rosemeyer and Grossman (2017) relate to the argument types discussed in this paper given their differences in semantic and syntactic characteristics. An obvious difference is that the verbal complements’ nominal arguments serve to buttress conceptual structure; no parallel support structures exist for the arguments discussed in this paper.

6.3 From metaphor-driven overtness to syntactic obligatoriness

Metaphor-driven overtness can develop over time into a syntactic constraint: one that applies to all cases including literal ones. Reinöhl (2016) shows for Indo-Aryan that the crucial semantic shift which marks the transition of relational nouns and adverbs into postpositions begins with metaphorical uses. We briefly illustrate this with the development of Sanskrit madhye ‘in the middle’, via Middle Indo-Aryan majjhe/-i, into the Hindi postposition mẽ ‘in’. The mismatch between the literal sense in Example (49), ‘in the middle of his heart’, and the intended sense, an emphatic expression of containment (here: in the seat of emotion) without reference to an actual spatial center, leads over time to the loss of the semantic component ‘center’ in all contexts. In its new meaning ‘in’, the functor requires an overt dependent in all cases; thus, it has become a postposition.

It has long been recognized that metaphor plays an important role in grammaticalization processes (e.g., Bybee et al. 1994; Sweetser 1990). However, the impact of metaphoricity on argument overtness has until now only been described in Reinöhl (2016) to our knowledge. Non-metaphorical sources of syntactic obligatoriness include the process of “overtification” discussed in the previous section, where arguments are at first overtly expressed because of their informativity, but later as a general syntactic rule.^[28]

6.4 The psycholinguistics of metaphor-driven argument overtness

While there is a considerable amount of research into the psycholinguistics of metaphor (e.g., Boroditsky 2000, 2018; Gibbs 2008; Gibbs and Matlock 2008; Glucksberg 2003; Katz 2017), we are not aware of any study focusing on metaphor-driven argument overtness. Yet the consistency of the effect suggests that its roots lie in human language processing, rather than being language-specific. We here suggest possible production- and comprehension-related effects as a focus for future psycholinguistic research.

For language production, we need to account for why speakers infallibly provide an overt argument even if the corresponding referent is highly activated. Two scenarios can be envisioned. In scenario A, the bare functor goes into production first, keeping online processing costs low. Observing signs of incomprehension, the speaker quickly repairs the sentence by providing the argument. In scenario B, there is no repair post-production, but an overt argument is incorporated in sentence planning before the functor is articulated, such as suggested in forward modeling accounts of language production (see Pickering and Garrod 2013). There is evidence suggesting that this second scenario is the correct one.

Firstly, we would expect consistent evidence for a pause if we were dealing with a repair mechanism as in scenario A. However, we have not seen any evidence pointing to such an effect. In the corpus data, we find no evidence for pauses after the functor is articulated – if these had been present, they would have been indicated by commas or pause annotations. Secondly, if scenario A were correct, with speakers repairing post-production, we would not expect argument overtness to be total. Not all speakers pay continuous close attention to their hearers, and so may not pick up on signs of incomprehension. Furthermore, in monologues and in written language, there is no such feedback loop in most cases (with the exception of quoted direct speech). Thirdly, post-production repair seems likely to imply a certain level of speaker consciousness, i.e., at least some speakers would plausibly be able to analyze what they are doing and why without having the phenomenon explained to them. This is not what we find.

All evidence points to scenario B, the formation of the complete construction before articulation. Possibly, a paradigmatic choice is available when the construction is planned. Crucially, we believe that – even though we are dealing with syntactically compositional constructions consisting of, e.g., nominal heads and their possessors or verbs and their arguments – the units in question are larger, indivisible semantic wholes. Functor and argument, by the very nature of their relationship, have an intimate semantic connection. The terms are even more closely bound in metaphorical uses as their meanings co-adapt. In the example in the middle of trouble, middle does not refer to an actual center, and trouble is portrayed as if it had one. We propose that metaphorical constructions as gestalts are indivisible for the purposes of production, as their semantic structures are not achievable through an incremental, monotonic process of cumulative restriction.

We make the following predictions about comprehension-related effects. We pointed out in Section 6.2 that, if the argument is not provided, a literal interpretation is forced (see Example (57), repeated from (55)). The listener will prefer any interpretation rather than considering the sentence ill-formed. Research exploring the boundaries of “grammaticality” is therefore not relevant to identifying the causes and limits of metaphor-driven argument obligatoriness. Instead, studies of referent recovery costs, which tie the referent closely to expectation-in-context, yield relevant predictions. In examples such as Example (50), the listener will consider the second sentence well-formed, but they have to “search around” for a possible referent matching a literal interpretation. In these cases, therefore, we predict an N400 effect, as the referent is hard to identify (e.g., Frenzel et al. 2011; Kaan 2007) or hard to “retrieve” (Brouwer et al. 2017).