The majority of psycholinguistic studies on metaphor focus on comprehension, leaving the important question of metaphor production poorly understood (Holyoak and Stamenković, 2018). Moreover, those few studies that examine metaphor production tend to focus on a variety of questions, such as pragmatic factors that promote metaphor production (Hussey & Katz, 2006); brain networks involved in metaphor production (Beaty et al., 2017); the role of working memory in metaphor production (Chiappe and Chiappe, 2007); individual differences in semantic networks in creative metaphors (Li et al., 2021); or the distinction between metaphor and simile production (Oka et al., 2022). While these studies characterize multiple facets of metaphor production, they do not address a more basic elementary question; namely, what kinds of concepts are chosen to produce good metaphors? That is, why is it apt to describe one’s lawyer as a shark, and perhaps less apt (though conventional) to describe a faulty car as a lemon? In this paper, we characterize the semantic representations that make for producing comprehensible novel metaphors.

Katz (1989) employed a vehicle selection paradigm to bring metaphor production under experimental control. In his task he asked participants to complete metaphor frames (e.g., Sociology is the ________ of sciences) by choosing the single best word from a list to serve as a vehicle (e.g., robin, hawk, USA, Switzerland, etc.). Two primary semantic variables of the selected vehicles were analyzed, word concreteness and semantic distance to the topic. Both concreteness and semantic distance effects were observed. Participants chose concrete vehicles more often than they did abstract vehicles, and the vehicles that were chosen were a moderate distance from the topic, rather than near or far. Therefore, Katz showed that a vehicle’s concreteness and its position among other concepts in semantic space both play a role in producing metaphors.

In this paper, we extend on the work of Katz (1989) by characterizing the semantic representations of topics and vehicles in metaphor production beyond semantic distance and concreteness. Again, a vehicle selection task was employed wherein participants were provided with abstract words that served as metaphor topics, and provided with a word-bank of concrete words to serve as vehicles. We focused on two semantic variables that are related to the variables studied by Katz (1989).

First, while Katz examined the semantic distance between topic and vehicle, we considered the semantic neighborhood density (SND) of the topic and vehicle words, which characterizes the proximity of a word and its neighbors in semantic space (Buchanan et al., 2001). Second, although Katz demonstrated that concrete words are preferred to use as metaphor vehicles, it remains unclear what aspects of concreteness are important. One aspect of concrete concepts is their motoric affordances (Gibson, 1979). Some concepts afford motoric interaction, such as bicycle, in as much as we interact with the referent motorically in many ways, whereas with others, such as butterfly we have less motoric interaction. Therefore, it is unclear if motoric interaction is the driver of concreteness effects in metaphor production. To investigate this, we considered body-object interaction (BOI), which characterizes the ease with which one can interact with a word’s referent (Siakaluk et al., 2008). Below, we describe these variables in greater detail and how they may pertain to metaphor processing.

The data from 11 participants were not included in the analysis, for one or more of the following reasons: they did not complete the study, created many similes despite the instructions given, or created extended metaphors. Thus, the analyses are based on 49 participants. The resulting participants created and interpreted nominal metaphors of the form A is B (although 5 of which each created a single simile – we included such participants because a single simile may not necessarily be strategic, but possibly accidental). A topic (high-SND vs. low-SND) by vehicle BOI (high vs. low) by vehicle SND (high vs. low) repeated measures ANOVA was conducted on the frequency of vehicles chosen from the four semantic conditions. A main effect of BOI was obtained, F (1, 48) = 31.297, p = <0.001, η_p² = 0.395; participants chose low-BOI vehicles (M = 5.1, SE = 0.104) 56.4% of the time whereas they chose high-BOI vehicles (M = 3.9, SE = 0.103) less frequently at 43.5% of the time; however, this main-effect was qualified by two interactions. First, a vehicle-BOI by vehicle-SND interaction was observed, F (1, 48) = 9.883, p = 0.003, η_p² = 0.171. As is depicted in Figure 1, the effect of BOI differs for high and low-SND vehicles: low-BOI vehicles were preferred for both high and low-SND items but this preference was greater for low-SND vehicles than high-SND vehicles.

Most importantly, a three-way, topic SND by vehicle BOI by vehicle SND, interaction was observed, F (1, 48) = 9.1, p = 0.004, η_p² = 0.159. As Figure 2 shows, the distribution of vehicle choices differs between high and low-SND topics. A simple-main effects analysis with vehicle BOI as the factor and topic and vehicle SND as the moderating factors revealed that, the low-BOI effect was present when the topic was high-SND and the vehicle was low-SND, F (1) = 40.74, p < 0.001 but not when the topic and vehicle were both high-SND, F (1) = 1.59, p = 0.21. Thus, for high-SND topics, participants relied more on low BOI – low SND vehicles to construct apt metaphors. We interpret this as an attempt by participants to reduce the overall semantic richness of the metaphor because the topic is already from a dense semantic neighborhood. Moreover, simple main effects analyses also revealed that when the topic was low-SND, the low-BOI effect was observed when participants selecting high-SND vehicles, F (1) = 8.27, p = 0.006, but no BOI effect was observed when both topic and vehicles were low-SND, F (1) = 2.17, p = 0.15. We interpret this to mean that when the topic is low SND, participants are drawn to low-SND vehicles, and higher semantic richness from BOI is tolerated. Moreover, when selecting a high-SND vehicle for a low-SND topic, participants preferred low-BOI words, again to reduce the richness from high-SND vehicles. Thus, when the topic is from a sparse semantic neighborhood, vehicle semantic richness does not matter as much as when the topic is from a dense semantic space.

In sum, participants did not choose the semantically richest vehicles when creating novel metaphors. Rather, our findings demonstrate that semantic richness is detrimental to metaphor processing, and that, on average participants aim to reduce semantic richness of the metaphor when selecting an apt metaphor vehicle to create a metaphor.

While Katz (1989) showed that concrete words serve as better metaphor vehicles than abstract words, our results characterize the nature of vehicle concreteness further. In general, our results showed that semantically less-rich concrete concepts are more amenable for metaphorical abstraction. In particular, we observed that sensorimotor and linguistic richness variables interact in novel metaphor production, such that participants reduced the overall semantic richness of their novel metaphors. This finding observed in the metaphor production task is consistent with recent work demonstrating that semantic richness is detrimental to novel metaphor comprehension. For example, Al-Azary and Buchanan (2017) constructed various novel metaphors varying on topic-concreteness and semantic neighborhood density of the topic and vehicle words, resulting in four conditions varying in semantic richness (abstract – low SND; concrete – low SND; abstract – high SND; concrete – high SND). They found that concreteness and semantic neighborhood density interacted, with semantically less-rich metaphors (low SND metaphors) being the most comprehensible, followed by abstract – high SND metaphors, and concrete – high SND metaphors, the semantically richest condition, being least comprehensible. To interpret their results, they argued that semantic richness, coming from topic concreteness and semantic neighborhood density, is detrimental to constructing metaphorical meaning, as such rich semantic representations are too specific and less malleable to take-on novel metaphoric associations. Moreover, irrelevant semantic properties, of which semantically richer concepts will have more of, must be inhibited during processing (Kintsch, 2000). Subsequent studied have reported similar findings in both offline and online tasks, and in special populations (Al-Azary et al., 2019, 2021). The present study extends findings of semantic richness to body-object interaction, and to metaphor production (rather than sole comprehension). We interpret the current results in a similar fashion to Al-Azary and Buchanan (2017); namely, semantically rich representations are too specified to take-on novel metaphorical meanings during metaphor processing. This was apparent here with the three-way interaction, as participants chose vehicles that were both low-BOI and low-SND when creating metaphors for high-SND topics.

The finding that low-BOI words served as better metaphorical vehicles on average (and especially when topic and vehicle words were high-SND) may be interpreted as evidence against embodied cognition, as such words are less amenable to motoric simulations than high-BOI counterparts. However, one must be cautious in such an interpretation, as the low-BOI words used in our study are nonetheless concrete concepts, and as such, can evoke sensory imagery. For example, low-BOI concepts such as lighting or tiger can trigger embodied simulations wherein one creates scenarios involving such concepts, even though one may not directly motorically interact with those words’ referents. Moreover, experimental work has demonstrated that bodily-actions related to concrete concepts are activated during metaphor processing (Al-Azary and Katz, 2021). Therefore, our results do not adjudicate between theoretical positions that necessitate a role for sensorimotor simulations in nominal metaphor comprehension (e.g., Gibbs and Matlock, 2008) or those that eschew sensorimotor simulations in favor of abstract representations (e.g., Glucksberg, 2008). However, our findings of low-BOI effects suggest that first-person motoric interaction, which BOI characterizes, is not the primary factor driving concreteness effects in metaphor production.

The semantic richness variables we considered – body-object interaction and semantic neighborhood density – characterize distinct sources of experience. Body-object interaction is learned through embodied experience whereas semantic neighborhood density reflects disembodied linguistic experience. The fact these two distinct variables interact in metaphor processing lends support to theoretical views that posit the presence of embodied and disembodied semantic representations that conjointly influence cognition (Paivio, 1971; Barsalou et al., 2008; Dove, 2011). Moreover, it has been argued that metaphor is a tool for building abstract representations from concrete concepts (Jamrozik et al., 2016). According to this “sensorimotor shedding” framework, novel metaphors are thought to be processed by sensorimotor simulations, and after repeated use, the sensorimotor properties begin to “shed,” making way for more abstract meanings to acrue. For example, consider lemon, a conventional word that metaphorically refers to something of little value due to its faultiness (e.g., my car is a lemon). Initially, lemon would be processed concretely, evoking multimodal simulations of taste, smell and feel, and after repeated exposure in linguistic contexts, develops an abstraction that is the primary representation when used metaphorically (see Al-Azary and Katz, 2021 for experimental support for this hypothesis). We speculate that low-BOI concepts, having less motor richness, have a “head start” in shedding their sensorimotor properties, and for this reason, are more amenable for metaphorical abstraction, as demonstrated in our study.

Other work has demonstrated that metaphorical potential is captured in a word’s relationality (Jamrozik et al., 2013). Accordingly, words such as marriage are more metaphorical than entity words, such as knife. Our work, however, demonstrates that among entities, there is considerable variation in metaphorical potential, which is predicted by word-level semantic representations such as semantic neighborhood density and body-object interaction. However, future work ought to characterize further the nature of metaphoricity in concrete concepts. For example, although concrete concepts are often defined in relation to imagery, imageability and concreteness are related yet distinct dimensions (Paivio et al., 1968). That is, some words are imageable, yet low on concreteness (such as ghost). Moreover, our focus was on motoric interaction, but other modalities in concrete concepts may play a role in metaphor, such as concepts with salient auditory (e.g., thunder) or motion (e.g., bullet) properties (see Cardillo et al., 2017 for modality specific metaphors). Furthermore, other measures of semantic richness have yet to be explored in metaphor processing experiments, such as the number of semantic features (McRae et al., 2005; Buchanan et al., 2019). Some concrete concepts are semantically rich because they have a high number of features, such as toad whereas others, like guppy, have fewer features and are semantically less-rich. More broadly, however, we acknowledge that pragmatic context can, and often does, make virtually any concept metaphorical, and can also shape a word’s intended metaphorical meaning (Gibbs et al., 2011; Haught, 2013). It should be the goal of future research to investigate the boundary conditions of semantic effects on metaphor processing, and how semantic representations interact with pragmatic contexts.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving humans were approved by the University of Western Ontario Research Ethics Board. The studies were conducted in accordance with the local legislation and institutional requirements. The participants provided their written informed consent to participate in this study.

HA-A conceived and designed the experiment, collected and analyzed the data, and wrote the first draft of the manuscript. ANK played a supervisory role in every step of the process and supported the project with grant funding. All authors contributed to the article and approved the submitted version.

This research was supported by Natural Sciences and Engineering Research Council of Canada Operating Grant NSERC 06P007040 to ANK.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.