Metaphorical meaning is not a flat mapping between concepts, but a complex cognitive phenomenon that integrates multiple levels of interpretation. In this paper, we propose a stratified model of metaphor processing that treats meaning as an onion: a multi-layered structure comprising (1) contextual information, (2) conceptual blending analysis, and (3) pragmatic analysis. This three-dimensional framework allows for a richer and more cognitively grounded approach to metaphor interpretation in computational systems. At the first level, metaphors are annotated through contextual metadata. At the second level, we model conceptual combinations, linking components to emergent meanings. Finally, at the third level, we introduce a pragmatic vocabulary to capture speaker intent, communicative function, and contextual efects, aligning metaphor understanding with pragmatic theories. By unifying these layers into a single formal framework, our model lays the groundwork for computational methods capable of representing metaphorical meaning beyond surface associations-toward deeper, more context-sensitive reasoning.
Metaphors pervade human communication and cognition, extending far beyond mere linguistic decoration. As cognitive tools, they grant privileged access to implicit knowledge structures that might otherwise remain hidden1][. By mapping relationships between concepts, metaphors serve as bridges that both reveal and reshape our conceptual frameworks—think of how we s aspyetnhd,astavwee,or waste time, implicitly assuming it is a finite resource.
Despite rapid progress in natural language processing, computational metaphor analysis continues
to face five intertwined challenges rooted in the very knowledge structures metaphors invoke:
1. Data scarcity and representational gaps. Datasets accounting for many metaphorical phenomena
are scarce, and building new ones is (i) resource-intensive, and (ii) hindered by frameworks that
go no further than simple domain mappings.
2. Contextual insensitivity. While the Conceptual Metaphor Theory (CMT) developed by Lakof and
Johnson [
ISSN1613-0073 5. Limits of current computational works. Even large language models (LLMs), the state of the art in metaphor processing, struggle to distinguish deep relational mappings from mere associations, particularly in complex or multimodal metap5h,o6r]s. [
Given this context, in this paper we aim to articulate a theoretical proposal that may serve both as a fruitful direction for future research and as a foundation for subsequent empirical work on thi issue. Specifically, we put forward an operational framework designed to support the processing of metaphorical meaning in a way that can efectively represent and interweave both conceptual and pragmatic aspects.
The starting point for this framework lies in the observation that, for analytical purposes, implici linguistic meaning can be treated as operating on a diferent level from that of explicitly encoded, surface-level meaning. Conceptual and pragmatic meanings are not directly encoded in the expression itself; rather, they are inferred through processes of contextualization and implicature. Consider, fo example, the sentencwee are wasting our time: the conceptual interpretation—where notions such as ifnite resources, waste, and time emerge, and where an analogical operation groups features of these concepts—is not directly expressed by the literal sentence. Similarly, the utterance’s potential pragmati function—i.e., the way that the utterance is intended—is also not explicitly encoded. This distinction reflects a view of meaning as a multi-level phenomenon—not necessarily intended as a model of how language inherently works, but as a potentially efective way of structuring meaning for computational processing—where explicitly encoded meaning and interpretation can be distinguished from conceptual and pragmatic ones, which remain implicit. Although stratified or hierarchically structured models of meaning are well established in the literature on Pragmatics (e.g. 7in]),Gcroimcepu[tational metaphor analysis and pragmatics have yet to establish a systematic connection—a link that could substantially advance research in both domains.
Therefore, from this theoretical perspective we propose a framework focused on the implicit level of meaning that aims to study metaphorcsogansitive tools in context, which do not exist as an abstract operation, separate from its usage or from concrete linguistic experience. A simultaneous cognitive and pragmatic analysis would in fact support the view of metaphor as a tool that efectively works—that is, it achieves communicative eficacy—when it is perceived as conceptually appropriate and functionally aligned with the speaker’s pragmatic goal, in addition to serving as models that reflect the cognitive structuring of experience, shaping both reasoning and a1]c.tIinonth[is view, the implicit knowledge conveyed by metaphors can be processed more fruitfully in computational systems if conceived as a multi-layered entity, like an onion. Each of our onion layer can be ontologically represented, providing a structured and functional approach to modeling metaphorical knowledge that aligns with our analytic goals. If the outermost layer correspond to the level of contextual information mapping, moving inward reveals the level at which the analogical conceptual operations take place, and deeper still, the pragmat intention that motivated the entire utterance. By connecting CMT’s cognitive account of metaphor with SAT–inspired pragmatics, our stratified framework aim to capture both what metaphors map and what they accomplish in communication. This unified model lays the groundwork for richer datasets, standardized evaluation, and computational systems that provide a clearer account of how to unlock implicit knowledge through metaphor.
The remainder of this paper is structured as follows. In S2e,cwteiodnescribe the related works. Section3 introduces our three‐dimensional model of meaning processing, illustrating how conceptual and contextual layers can be systematically integrated. In4,Sweceteioxpnlore the consequences of this multi‐layered view for computational processing of metaphors, discussing both implementation strategies and evaluation methods. Finally, Se5cstuiomnmarizes our contributions and outlines directions for future work.
In this section, we describe the theorical background for our proposal and discuss the challenges related to existing works on metaphor processing.
CMT, developed by Lakof and Johnson, posits that metaphors map a source domain onto a target domain via systematic correspondences, enabling abstract reasoning through familiar experiential structures2][. Conceptual Blending Theory (CBT) extends this view by introducing a generic “blend” space that selectively inherits elements from both input domains according to a blending criterion or key property, such aysellow when we say “golden hair”, evoking that golden is yellow and shiny like the hair8[]. CBT is regarded as a valid computational approach also by the Categorizatio9n]. theory [ This account sees metaphors as category statements where the source acquires a categorical meaning, more abstract than its literal meaning. For example, “golden” in “golden hair” denotes the category of “shiny, yellow things”. In this view, conceptual blending can be used to extract the abstract meaning of the source and combine it with the meaning of the target. These frameworks emphasize that metaphor comprehension relies on shared background knowledge (or frames), which Fillmore’s frame semantics formalizes by associating lexical items with structured role–filler expec1t0]a.tTiohnesc[omplementary relationship between CMT, CBT, and frame semantics highlights that metaphorical meaning emerges not merely from lexical similarity but from dynamic frame activation and role alignment within a community’s commonsense knowledge4][. Beyond commonsense or prototypical knowledge, recent theories of metaphors have noted the lack of inclusion of personal and sometimes contextual aspects that influence knowledge acquisition and interchange. In fact, the experiential dimension of metaphor has traditionally been downplayed, with research focusing primarily on metaphors as a mental and individual achievement. Researchers have so far paid little attention to context and the collaborati production of metaphoric langua11g]e. [
The communicative aspect of metaphor is fundamental to view metaphor as a multidimensional
phenomenon [
Connected to inter-bodily multidimensional accounts of metaphor is the metaphor resistance phenomenon, only recently studied, and the various reasons why it happens. For instance, people resist metaphors if they lack explanatory power or for a preference for alternative metaphorical concept with respect to normative ones. However, without a comprehensive metaphor study it is not possible to know why some metaphors aren’t picked 1u7p].[Thus, we aim to shed light on these theoretical studies to account for a multidimensional view of metaphor that can also reflect in a new strand of computational metaphor processing studies.
Computational representations of metaphor have leveraged structured resources such as MetaNet and Framester, which align conceptual metaphors with FrameNet frames an1d8]r.oTlhese [Amnestic Forgery Ontology further integrates MetaNet into Framester, providing a rich graph of source–tar get frame pairs, example sentences, and hierarchical relations among met19a]p.hOonrsto[logical formalisms based on the Blending Onto1loegnycapsulate the four-space blending networks of CBT, 1Available ahtttps://github.com/dersuchendee/BlendingOnt.ology permitting explicit encoding of input spaces, generic spaces, blends, and their mapping relations. Despite these advances, existing SWRL-based and rule-driven approaches often lack scalability and fail to account for tacit, context-dependent knowledge, limiting their applicability to open-ended o multimodal metaphor interpretat2i0o,n21[].
Recent years have seen new dataset for computational metaphor processing. Among others, the VU Amsterdam Metaphor Corpus (VUA) has become a standard benchmark for metaphor de2t2e]c.tion [ However, such corpus only marks metaphoric tokens and does not specify the source and target domains behind each metaphor. To fill this gap, smaller domain-annotated datasets have emerged, such as the one by Gordon et al2.3[], which annotates metaphorical tokens, source and target conceptual domains. A more comprehensive corpus is Metane2t4][, a semantic wiki with conceptual metaphors that has been employed and extended in the Framester knowledge 1h8u].b L[ippolis et al. introduce the Balanced Conceptual Metaphor Testing Dataset (BCMTD), the first dataset that contains metaphors from the medical domain to test systems’ generalizab2i5l]i.tAyt[ the same time, it is claime2d6][ that in spite of the attention that metaphor has received over the centuries, and more recently within the cognitive paradigm, we still lack explicit and rigorous procedures for its identification and analysis, especially when one looks at authentic conversational data rather than decontextualized sentence Furthermore, more recently, doubts have been expressed about the legitimacy of extrapolating too readily from language to cognitive structure, and distinctions have been drawn between claims about whole linguistic communities or idealised native speakers, and claims about the minds of single individuals.
Recent work in theory-driven metaphor processing integrates CMT and CBT into model architectures
and annotation schemas. Mao et2a7l].a[nd Tian et al2.8[] demonstrate that embedding theoretical
constraints into training objectives improves metaphor detection performance. For interpretation,
unsupervised and neural methods extract source and target domains or link attributes be2t9w, een them [
30], but typically depend on single-word annotations or pre-specified t3a1r]g.eVtissu[al metaphor
datasets such as MetaCLUE and ELCo provide multimodal challenges, yet public resources remain
scarce 3[
FurthermoreL,ieto et al[3.5] recently presented a system called CMLEaTble to perform metaphor
generation and classification by applying a formal operationalization of the CBT. The coCrLeisof MET
a reasoning framework specialized in human-like commonsense concept combination: the
Typicalitybased Compositional LogicC(TL) first presented in 3[
Pragmatics, as the study of the meaning of linguistic signs in context—that is, in their actual use—deals primarily with implicit linguistic knowledge: the type of meaning it investigates, the pragmatic message, is not encoded in any direct way in the literal utterance. One can say something that literally means one thing while actually intending something entirely diferent. For instance, saying to someone on the subway, “You’re standing on my foot”, means likely not wanting to describe the situation to them, but rather asking them to move. The utterance may contain only hints as to how the pragmatic meaning should be interpreted, such as a specific tone of voice, and it depends on a variety of extra-textual and extra-linguistic elements, including context, linguistic conventions, and socio-cultural norms.
Utterances that contain metaphors can, of course, be described in pragmatic terms; however, pragmatic approaches to metaphor depend on the diferent approaches to the interpretation of metaphor itself. In that of two leading figures in the pragmatics literature and of SAT such as Grice a3n7d,3S8e]atrhler[e is the idea that the metaphorical interpretation presupposes and is derived from the literal interpretati through a “decoding” of a secondary level of meaning, as if metaphors and other figures of speech were some exceptional or supplementary use of language. The assumption of a clear-cut distinction between the literal interpretation and the metaphorical interpretation of an utterance is instead abandone the approach to metaphor of Sperber and Wi3l9s,o4n0][, the theorists of Relevance Theory (RT). RT, whose strength lies in its understanding of how language works through its emphasis on the inferential comprehension of non-literal meanings from contextual cues and assumptions, and on the cognitive principle that human communication aims at maximal relevance with minimal processing efort, treats metaphor as a pragmatic tool conveying implicatures and implicit evaluations. The cognitive approach to the study of metaphors has been increasingly followed after the rise of CMT, and works have focused on the many and varied efects that metaphor has on cognitive processes41(e,.4g2.,])[. However, in computational pragmatics the field still appears to be very open: the pragmatic processing and understanding of metaphor—why it is used, in what context, and with what efect—remains only partially addressed, despite recent advances of contextual neural models. Focusing on cognitive aspects may also obscure others, such as the performative dimension of utterances, which is emphasized in SAT. Speaking of utterances as speech acts, in fact, means considering language and linguistic utterances not merely as expressions of mental operations, like articulations of thoughts, but as actions in themselves. As actions, while pursuing their speaker’s intentions, they produce efects in the world, whether intended or not, and carry out acts such as describing, ordering, pleading, or, through conventional formulas, marrying two people or sentencing someone.
A central aspect in explaining pragmatic meaning is the intention attributed to an utterance, the illocution—the way a speaker intends the sentence they utter, for instance as a statement or a request. One of the main goals of computational pragmatics inspired by SAT is to study how to automatically assign anillocutionary act to an utterance, a task framed as a problem of context dependence but complicated by several factors. First, there is the challenge of formalizing intentional, conventional, o otherwise contextual aspects that are extra-linguistic, along with all the choices that such formalizati entails. Second, there is no deterministic relationship between clause types and illocutionary force: imperative clauses are not invariably commands, interrogative clauses are not always queries, and declarative clauses are not necessarily assertions. Finally, any illocutionary classification, however widely shared, will inevitably fall short of capturing the compositionality of the intentions at play i natural language. Depending on factors such as the power dynamics between speaker and addressee and their communicative goals, an utterance may emerge as a complex blend of illocutionary forces, and the same applies in the case of a “metaphorical spee2c.hAascMt”ichelli, Tong and Shutov4a3][ observe, the literature on metaphor intention remains fragmented: there is still no systematic and comprehensive account of intentions behind metaphor use, nor an operationalized framework that enables their annotation in linguistic data. Yet explaining the communicative role of metaphors in terms of taxonomies of intentions, as they suggest, risks reducing intentions to isolated categories. Moreover, although metaphor scholars do not share a common notion of intention, it is typically formalized as a prior intention—that is, a representation in the speaker’s mind of their communicative goals, particularly in approaches that address metaphor through Theory of Mind models based on beliefs and intentions. As Gibbs [44] points out, conceiving of intentions as individual mental states makes them opaque, since agents are not always aware of the causes of their behaviour; this, in turn, can lead to computational abstractions that lack real explanatory value for illocutionary intention. Studying metaphors as speec acts, instead, means treating intention not as a mental state, but as a feature attributed to linguistic a 2Understood here not as a distinct illocutionary category, but as a speech act that is in some way metaphorical. in line with the philosophy of action outlined by Austin in his linguistic4a5n, a46ly].seTsh[is means that intentions are not the reasons speakers might give for using certain metaphors, but rather the intentions expressed by the utterance itself, such as whether it is interpreted as a request rather than statement. Consequently, it is not immediately necessary to presuppose mental states beyond linguistic analysis. From a computational perspective, this entails modelling intention as an inferred property of the communication rather than as a presupposed mental condition—certainly a challenge, but also a potentially valuable contribution to computational pragmatics. If metaphor-containing utterance are speech acts, they also produce efects on the communicative context, part of which derive from metaphor as a cognitive and stylistic device. The pragmatic efectiveness of a metaphorical speech act—how well the utterance serves the communicative intention and aligns with the interlocutor’s context and sensitivity—is shaped by the stylistic tone conveyed through rhetorical figures and by the strength of the conceptual evocation created by the metaphor itself. Within a given illocutionar intention, a metaphor can either reinforce or attenuate that intention: a command may sound more polite, a request more engaging, or a threat more forceful; conversely, if the metaphor is perceived as inappropriate, its expressive or persuasive force may be diminished. This suggests that, beyond illocutionary intention, it is also important to conspiedrelorcutthioenary level (in Austin’s terms, the efects producedby saying something) of a metaphorical speech act, which concerns the efect an utterance or image has on the listener (e.g., convincing, frightening, provoking thought). These efects can be cognitive—often more dificult to study—but also more immediate at the emotional-psychological level, since, as figures of speech, metaphors can aim at linguistic persuasion, including the evocation of emotional responses.
We now present the main thesis of our research proposal and the three-dimensional framework for processing metaphor meaning. This model provides a foundational vocabulary that, once debated within the community, can lay the groundwork for future computational implementations. Its primary aim is to ofer a comprehensive interpretation of metaphor, treating it simultaneously as a cognitive instrument and as a speech act.
(a) Image of the song “Smoking kills” by Dopesmokea.
(b) Anti-tobacco campaign from(c) Ad by The Leith Agency for the online campaigNno Smoke ASH UK (Dec 1994).a Revolution.a
ahttps://adsspot.me/media/prints/ ahttps://open.spotify.com/intl-it/ ahttps://digitaladvocacycenterwhera.sh-action-on-smoking-and-healthtrack/6XL0aFfNf6PzhyzPddRArR com/en/gun-with-cigarette-bullets/bullet-788f341dba7b
The external content layer includes metadata of the communicative object under analysis: domain, provenance, connections to existing knowledge bases, etc. It also includes annotator metadata.
Consider the anti-smoking images in Figu1.rFeirst of all, this layer would include domain classification (public health advertising), provenance references (campaign organization, publication date, media outlet where it appeared), and frame connections linking to established conceptual metaphor databases such as Metaneth’sarm is destruction. The annotator metadata would capture the interpreter’s geographical background, cultural context (attitudes toward smoking and firearms), and demographic information, recognizing that metaphor interpretation is inherently subjective and culturally situate This foundational layer ensures that subsequent cognitive and pragmatic analyses can be properly contextualized within their social, temporal, and interpretive framewor2kssh. oFwigsuarne example of annotations for the metaphorical image shown in1Fbi.gure
Domain: “Public health” Provenance: ● Author: “NoSmokeRevolution” ● Media platform: “Facebook” ● Publication date: November 15, 2013
Annotations: “A gun with cigarettes as bullets.”
This layer includes the cognitive context of the element, and in particular what enables the metaphorica mapping by taking into account only the two domains employed in the metaphors. Analyses that detect source and target concepts in a sentence or image are part of this layer, as well as the approaches that try to blend the two concepts to generate a representation of the metaphorical meaning. The representation of source and target can provide a rich description of the respective domains, including frame roles that describe related concepts and the semantics of their relations, including why they map (the blending principle).
The context of this layer is given at least by the annotation of source and target domains and is fully represented according to the Blending Ontology (see2S)ewcittihonthe following elements: (i) Blendable (source and target domains); (ii) Blending (the blending principle); (iii) Blended (the actual blend).
To generate a representation of the metaphorical meaning that emerges from the blending, we need a computational mechanism to realize the combination of source and target. Consider again one of the anti-smoking advertisement such as Fig1u.rTehe source domain ishooting, while the target domain is smoking. According to Blending Theory, the blendables are the two input spaces: the weapon frame (with roles like shooter, ammunition, target, harm) and the smoking frame (with roles like smoker, cigarettes, user, health consequences). The blending principle that enables this conceptual integration is lethality—both bullets and cigarettes cause death, though through diferent temporal mechanisms. This creates an emergent blended space where cigarettes inherit the immediate, violent danger typically associated with ammunition, while smoking adopts the intentional, direct harm associated with shooting. The blended space produces the integrated concept where each cigarette becomes a bullet the smoker ifres at themselves, creating a self-destructive cycle. Frame roles map systematicsahlolyo:tterhe maps onto thesmoker, ammunition ontocigarettes, theact of shooting onto theact of smoking, andfatal injury ontosmoking-related disease, unified by the overarching principleloetfhality that bridges the temporal gap between immediate and gradual self-harm.
Many attempts have been made to develop an algorithmic solution for conceptual combinations, each with its limitations. The Structure Mapping Engine4(S7M]) Eis[a well-known computational approach that takes into account a broad description of source and target domains. Given the knowledge graphs representing the objects and the relations involved in the two domains, the SME finds isomorphic subgraphs that yeld the mappings. While powerful, the SME requires a rich, formal description of the two domains. A less demanding approach, directly inspired by the Categorization theoCrLy, is MET [35]. Based on a cognitively inspired logic for conceptual combination, this system can automatically generate a prototypical representation of the metaphorical mapping (the blend) from the prototypica representations of source and target conceptCsL. McoEnTsists in a three-step pipeline. The first step builds a structured representation of the metaphors to be analyzed, highlighting source and targe The second step generates a prototypical representation of both the source concept and the targe concept. Each concept is represented by a prototype, i.e. a small set of typical features that can be automatically extracted from Concep4t8N]e.tTh[e third step is the conceptual combination. CTLhe T logic combines the source and target prototypes to generate an abstract representation of the metapho The combinations generated by MCELTwere generally accepted by human judges as capturing relevant aspects of the intended metaphorical meaning.
METCL can be seen as a tool for knowledge graph completion. Indeed, manually curated resources like MetaNet are inevitably incomplete and sufer from under-representation of the wide metaphor phenomenon3. The ability of the system to automatically generate a representation of a given metaphor allows to cover a wider spectrum of expressions. In Lieto3e5]t, aLlL.M[s were used to classify metaphorical sentences into the MetaNet ontology classes, showing the benefits of extending the ontology with the representations generated bCyL.MOEuTr proposal for Layer 2 is to use a conceptual combination system like MECTL as the reasoning mechanism. It can also be provided as a basis for the LAG-based approach by Lippolis et a2l5.][(see Section2) that addresses multimodal metaphorical raw data and generates knowledge graphs based on implicit knowledge3. sFhigouwrsethe METCL implementation of this layer.
Source: “bullets” ● war ● nail targets ● roadblock ● poker slang ● ...
Target: “cigarettes” ● unhealthy ● bad ● gas station ● carcinogens ● ...
METCL
Combination: “cigarettes-bullets” ● unhealthy ● bad ● war ● nail targets ● ...
Compared with the first two layers, the third remains largely under‑investigated from a computational perspective. Our aim is to discuss a categorization that would allow for its efective implementation, also with a view to the composition and annotation of a dataset. The idea, in fact, is to propose a human annotation campaign on a dataset—such as one consisting of metaphors found in images, as in examples discussed so far—and then to assess the behavior of an automatic system on the same task.
As seen in Section 2.5, works on the communicative role of metaphors in a pragmatic sense tend to explain it in terms of the intentions or discourse goals they are meant to achieve, proposing taxonomies of intentions43[], that often assumes intentions as mental states. For the analysis of intention as an illocutionary act, we reuse the standard categories of SAT, focusing on defining the role of metaphorical linguistic acti4o.n 3In particular, the MetaNet project is still under constant improvement and enrichment after more than a decade from its beginning. More information is availabltehoenwebsite of the MetaNet Pro.ject 4The general classification of illocutionary acts (or language functions) has been a recurring topic in the philosophy of language literature, but here we are not concerned with defending a particular classification. We adopt an approach that would allow us to flexibly assess the composition of forces within an utterance containing a metaphor, focusing on the example ofdirectives—the linguistic category of requests or orders typically derived from imperative sentences and found across most languages and linguistic cult4u9r].esA[ccording to certain theoretical frameworks, such as those proposed by Popa-Wyatt, one could also distinguish primary and secondary illocutionary acts, sometimes nested within each other, as, for instance, in the case of an ironic act embedded inside a metaphori5c0a],lbounteg[iven the preliminary nature of this operational framework, we focus on annotating only the forces perceived as primary.
As discussed, a perlocutionary analysis aligns with evaluating not only what a metaphor means, but also what it does, treating utterances as actions that produce efects in the communicative context and in listeners. We propose capturing these efects through a first, immediate psychological evaluation, and extend this approach across modalities, including the recognition of tone of voice in speech as a cue to pragmatic intention.
We therefore propose the following categorization: • Attitude: The speaker’s evaluative stance toward the object of the metaphorical analogy (not toward individual discourse elements separately). This dimension reflects whether the attitude conveyed is positive, negative, or neutral, and can influence the tone and intention inferred from the utterance. • Illocutionary Act: The act that expresses how speakers intend their utterances to be understood.
Following classifications such as Searle’s taxonomy, illocutionary acts can include assertive, directive, commissive, expressive, and declarative types. Communicative acts often involve a composite of multiple illocutionary forces, but our case study focuses specifically on elements that can be traced back to directive illocutionary acts as the principal pragmatic force perceived by annotators.
– Directive Kind: The definition of the specific type of illocutionary act according to a range of assertivfeorce. Directive acts are those in which the speaker attempts to get the addressee to perform an action, but this force can vary in intensity: they may take the form of requests, commands, suggestions, prohibitions, or pleas. We propose to model the type of directive along a continuum of assertiveness to capture this variation in illocutionary strength. • Perlocutionary Act: The act concerning the efects an utterance or image has on the listener or annotator, here analyzed as psychological or emotional. This includes both intended and unintended responses, which we propose to capture as annotators’ evaluations.
– Perlocutionary Efect : The emotions evoked by the metaphor in the annotator, categorized according to the Emotion Frame Ontol5o1g]y. T[hese emotional responses may vary and are optional: in some cases a metaphor may evoke no particular emotion. – Eficacy : A measure (on a Likert scale from 1 to 5) of the metaphor’s efectiveness, namely its capacity for persuasion relative to the presumed illocutionary intention, its appropriateness to the perceived context, and the annotator’s personal sensitivity. • Other Contextual Elements – Pragmatic/Visual Clues: Indicators that suggest the speaker’s or communicator’s attitude toward the metaphor’s referent, e.g. use of specific colors. – Pragmatic/Visual Tone of Voice: The overall communicative tone (e.g., humorous, dramatic, ironic), which contributes to how the metaphor and its communicative intent are perceived.
As an example, consider again the image shown in Fig1ub.reThe speaker’s evaluative stance can easily be perceived by an annotator as negative toward the analogy’s object: smoking is framed as a lethal threat and self-destructive act. This negative attitude is conveyed both conceptually (via t metaphor) and visually (via design choices). The utterance, although visual, can function as a directive illocutionary act because it is embedded in the context of anti-smoking public campaigns and visually represents a cigarette as a bullet. This metaphor implies a communicative function—discouraging smoking behavior by equating it with an act of self-inflicted violence. From this function, the speaker’s intention can be inferred as emerging from the pragmatic force and evaluative stance of the act itself, as to prevent smoking behavior by highlighting its deadly consequences through the conceptual mapping ofsmoking is shooting oneself. As for the directive kind, the discouraging attitude against smoking places the metaphorical act in a position that can be perceived by an annotator as the highly assertive end of the directive spectrum, referring to a prohibitive force similar to a command. This is achieved not through explicit verbal instruction but via the emotional and conceptual weight of the metaphorica image. The perlocutionary efects that can be experienced by an annotator can include emotions such as concern, discomfort, and shock as viewers process the visual equation of cigarettes with ammunition. The metaphor’s eficacy can be perceived as potentially high due to the stark, unambiguous nature of the weapon imagery and its cultural associations with death and violence. Visual clues reinforce the negative attitude through the absence of color, and the conceptual mapping between shooting a gun and smoking a cigarette is underlined by the presence of smoke. The overall visual tone of voice is dramatic and alarming, designed to interrupt habitual thinking about smoking through visceral impact rather than rational argument. This example of analysis is outlined4i.n Figure
Attitude: negative Illocutionary Act: directive
● Directive Kind: highly assertive Perlocutionary Act: “stop smoking” ● Perlocutionary Effect: “fear”, “concern”, “shock” ● Efficacy: high (4/5) Visual Clues: “absence of color”, “smoke”
Visual tone of voice: “dramatic”, “alarming”
Our proposal holds several implications for computational metaphor processing. In this section, we outline them and propose future research that can implement the approach into an empirical framework, going back to the five problems described in Sect1io.n
Datasets. Metaphor datasets should not only account for simple domain mappings within sentences,
but also allow for the annotation of conversational and multimodal data, capturing dimensions beyond
the purely cognitive. To date, no single metaphor-related dataset annotates attitude, speech acts an
clues of utterances, although initial eforts are being made in this direction, for example in representing
intentions4[
Knowledge representation. Current ontological frameworks for metaphor representation, as discussed in Sectio2n, have made important strides by encoding source–target mappings and blend spaces. However, they tend to treat metaphor as a phenomenon abstracted from discourse and pragmatic intent. In contrast, our proposal emphasizes a multi-layered model of metaphor meaning processing that integrates contextual, conceptual, and pragmatic levels, requiring a representational shift tha could, in future work, be formalized ontologically. Metaphor representation should therefore evolve toward stratified ontological models that include a pragmatic layer capturing the illocutionary an perlocutionary acts of metaphorical utterances, grounded in discourse context, communicative intent, and emotional efects. Such a layered representation would also support updates or reinterpretations of metaphorical meaning in conversational settings—for instance, through reasoning mechanisms capable of tracking metaphor reinterpretation and resistance, grounded in community norms and individual diferences.
Possible application of the framework in computational metaphor processing. The proposed framework has direct implications for computational processing. By aligning processing tasks with the three dimensions, we can design systems that generate, interpret, and apply metaphors in a more context-aware, human-aligned, and socially beneficial manner. For instance, both the LAG approach [25] and the METCL [35] approach can be employed in a complementary way to assess the conceptual content and combination layer for metaphor processing. We argue tChLaatnMdELTAG are not only compatible but potentially integrable and mutually reinforcing. Specifically, LAG could leverage the prototypical property lists provided byCMLtEoT weight RDF triples, yielding more precise definitions and reducing LLM hallucinations. Conversely,CMLEcoTuld exploit LAG’s capacity to identify missing frames in MetaNet or to align typical features with ontological categories, thus enabling a share RDF layer for interoperability. For instance, LAG might detect objects in an image, wCLhile MET could suggest which visual attributes instantiate their prototypical properties. For what concerns t pragmatic layer, the lack of datasets currently hinders computational implementation4),(see Section starting from a gold-standard annotated dataset, but LLM-based approaches such as LAG can be used to extract pragmatic features, which can in turn be analyzed.
For what concernmsetaphor generation, current methods often rely on shallow associations or fixed templates, leading to output that lacks conceptual coherence or pragmatic fit. By incorporating the layered representation we propose, generation systems—particularly those embedded in interactive environments such as chatbots or digital companions—can produce metaphors that are not only structurally sound, but also functionally appropriate to the communicative context. mLeiktaepwhiosre-a,ware recommender systems could translate figurative queries (e.g., “a sofa with a cozy vibe”) into concrete product attributes, letting users find items that match their afective intent while keeping the system’s reasoning transparent.
Metaphor understanding, too, can benefit from the layered representation we propose. By encoding both conceptual mappings and pragmatic implications, systems can disambiguate metaphorical usage more accurately, especially in real-life settings. LLMs, when coupled with structured semantic representations, may serve as efective metaphor interpreters. In such a neurosymbolic setup, the explicit surface form of expressions is parsed, mapped to conceptual domains, and then interpreted through pragmatic filters reflecting the speaker’s goal, emotional tone, and discourse situation. This facilitates better performance in metaphor detection, paraphrasing, and explanation tasks, especially in underexplored genres like dialogue, narratives, or scientific texts.
Metaphors for social good. Metaphors are not neutral; they carry social, cultural and emotional weight, and using or detecting them responsibly has implications across several socially relevant domains (see Section2), for instance employing metaphor-aware computational tools in education, science communication, doctor-patient interactions and hate speech.
Metaphors are cognitive tools deeply embedded in human reasoning and communication. In this work, we assess current challenges in metaphor processing and propose a three-layered framework that integrates both conceptual and pragmatic implicit knowledge. This onion-like stratification allows computational models to capture not only what metaphors mean but what they do, illuminating both their cognitive structure and their communicative function. In doing so, the framework aspires to bridge disciplinary boundaries, situating metaphorical analysis at the intersection of cognitive science, computational modeling, and pragmatics, thus ofering a conceptual scafolding that accommodates both ontological and epistemic variability.
Although the framework is currently theoretical, we demonstrate how it can serve as both a fruitfu direction for future research and a foundation for empirical work on this issue, for instance, in designing computational systems that are metaphor-aware, context-sensitive, and socially responsible. By formalizing metaphor as a structured, multi-dimensional phenomenon, we lay the foundation for new datasets, evaluation metrics, and neurosymbolic methods that better reflect the richness of metaphor in natural discourse. This formalization highlights the interplay between conceptual structure and pragmatic function, opening avenues for interpretive flexibility in subsequent computational implementations.
The main current limitation of this framework is the lack of computational implementation. It is also necessary to develop a robust, shared line of research that integrates CBT and pragmatics theories on metaphor interpretation. This efort should include discussion of a common vocabulary for the construction of future datasets, and, on the computational side, a coherent formalization pragmatic aspects such as intentional and contextual factors. Looking ahead, we envision extending this framework to new modalities and domains and implementing it in computational experiments.
This work was supported by the PhD scholarship “Discovery, Formalisation and Re-use of Knowledge Patterns and Graphs for the Science of Science”, funded by CNR-ISTC through the WHOW project (EU CEF programme - grant agreement no. INEA/CEF/ICT/ A2019/2063229).
In the preparation of this work, the authors used GPT-4 and Grammarly in order to: Grammar and spelling check. After using these tools, the authors reviewed and edited the content as needed and take full responsibility for the publication’s conten