Gilead and colleagues distinguish two specific reasoning mechanisms geared toward distinct levels of abstraction. First, they assume that a theory-based account is required for realizing inference on abstract levels. Second, they propose mental simulation as a means of extrapolating detailed imagination into possible futures. In contrast, we argue that mental simulation provides a flexible unified mechanism that supports a spectrum of (embodied) abstraction and inference driven by the schematic structure of action and event representations. Our proposal is based on a large body of neuroscientific and linguistic evidence arguing for the gradient nature of conceptual abstraction. The proposed type of simulation recruits (multi-)modal areas distributed across the brain as well as higher-level areas associated with more abstract and schematic concepts (Binder & Desai Reference Binder and Desai2011).

Evidence of recruitment of brain areas in language use supports this view of mental simulation as involving flexible underlying representations and mechanisms across different levels of abstraction: action-related words recruit sensorimotor representations (Pulvermüller Reference Pulvermüller2018). This process can be understood as grounding a mental simulation. Early experiments have established that this recruitment spans different levels of abstraction (Boulenger et al. Reference Boulenger, Hauk and Pulvermüller2008), a fact not mentioned by Gilead and colleagues. In a recent fMRI experiment, Wang et al. (Reference Wang, Wu, Ling, Xu, Fang, Wang, Binder, Men, Gao and Bi2018) further showed how processing of abstract words recruits distinct high-level language-related areas, as well as widely distributed (multi-)modal brain regions. They conclude that abstract concepts recruit embodied representations but also connect to higher-level linguistic networks. Findings from Desai et al. (Reference Desai, Conant, Binder, Park and Seidenberg2013) support this view through analysis of brain activations for different levels of abstraction. In comparing activations for language related to concrete action, metaphorical meanings and idioms, they found differential activations of motor areas for action language and metaphors – suggesting an effect of mental simulation in both. However, idioms showed only graded activation in these areas, suggesting some degree of abstraction from sensorimotor systems. Desai et al. conclude that idioms undergo a process of conventionalization in which their representations shift from more concrete motor representations toward more schematic representations. These neuroscientific studies suggest that sensorimotor representations contribute directly to the formation of abstract concepts and are differentially recruited in mental simulations. Further evidence suggests their role in simulation can be modulated by grammatical features, for example, as found by Liu and Bergen (Reference Liu and Bergen2016) in a study on how progressive and past tense affects language comprehension. They compared how speakers process sentences describing concrete versus abstract actions and found a location-sentence compatibility effect only for the progressive case. This result is consistent with the function of progressive tense as focusing on the internal details of the unfolding event, which requires a (more or less) detailed mental simulation. In contrast, the perfect tense shifts the focus from the means of action toward end-conditions after executing that action. Importantly, the effect is independent of using abstract or concrete verbs and reflects the shared schematic structure of actions.

We use the term “schematic structure” here to mean a categorical mental representation that generalizes over instances (Johnson Reference Johnson1987). Schemas capture regularities in sensorimotor systems that organize experience and interactions with the environment. In language comprehension, it is assumed that words and larger phrases activate schemas that add perceptual or motor components to a mental simulation of the content. These activations range from simple reinstantiations of detailed, modality-specific experiences up to abstract structures that capture basic conceptual knowledge. For instance, a source-path-goal schema is an abstract representation structuring our understanding of directed motion (Bergen & Chang Reference Bergen, Chang, Östman and Fried2005). Crucially, schemas are assumed to be “categorical” and rely on “predication” (as defined by the authors) across all levels of abstraction, because they allow for flexible recruitment in mental simulation across different levels of abstraction. Consider, for example, an action-schema for grasping: First, the use of parameterization (e.g., position where to grasp) is a form of abstraction. Second, a grasp-schema can subsume different types of grasping (pinch-grasp etc.) that share a generic event structure whereas involving different effectors or typical graspable-objects. Third, schemas can relate to each other (e.g., a grasp-schema requires an actor and a graspable object), implying abstraction over role-fillers. This view of schematic structure fulfills Gilead and colleagues' definition of abstraction as subsuming distinguishable schemas and actions. Moreover, the practical utility of such schemas for understanding abstract language has been nicely demonstrated by a computational model for metaphorical inference in the domain of economics (Narayanan Reference Narayanan1999). In this account, the ability to parameterize mental simulations in the source domain served as the basis for further inference in the target domain.

A schematic event representation is as well supported by neuroscientific evidence: a gradient of abstraction for such representation is assumed along midline structures of the brain (Stawarczyk et al. Reference Stawarczyk, Bezdek and Zacks2019). This includes the default network and, in particular, posterior cingulate cortex as well as medial prefrontal cortex, which is in agreement with the results discussed above (Wang et al. Reference Wang, Wu, Ling, Xu, Fang, Wang, Binder, Men, Gao and Bi2018). Recruitment of these event representations has been found independent of sensory modalities, stressing their abstract nature (Zadbood et al. Reference Zadbood, Chen, Leong, Norman and Hasson2017). Furthermore, comprehension of different script-based stories leads to distinct sequential schematic event patterns (Baldassano et al. Reference Baldassano, Hasson and Norman2018).

Overall, there is now substantial evidence showing an overlap in brain activation for abstract and concrete concepts, not only for actions (Desai et al. Reference Desai, Reilly and van Dam2018). We focused here on actions and how mental simulation provides a general inference mechanism across levels of abstraction, either through recruiting detailed sensorimotor representation or through relying on schematic simulation. (As a further example, Chwilla & Kolk [Reference Chwilla and Kolk2005] showed a priming effect for words that are related only through a shared schematic situation, suggesting that a schematic simulation can facilitate more efficient access of information for abstract information.) In particular, we reviewed results that show, first, how embodied representations are differentially recruited in language comprehension of abstract content, and, second, how this is complemented by the recruitment of more abstract schematic action structures that are engaged in mental simulation and that support abstract inferences. This broader view of mental simulation contrasts with the authors' view of the limitations of a simulation-based account.