Stratified Metaphor Processing Model

• Stratified metaphor processing is a layered framework that decomposes metaphor comprehension into levels including content analysis, conceptual mapping, and pragmatic intentionality.
• It integrates diverse computational methods—geometric, probabilistic, multimodal, and mechanistic—to systematically analyze metaphor interpretation.
• Empirical findings underscore the model’s ability to differentiate lexical, structural, and pragmatic features, highlighting challenges in benchmark evaluations and multimodal integration.

A stratified model of metaphor processing treats metaphor comprehension as a multilayer phenomenon rather than as a single source–target substitution. In recent computational work, the term denotes a family of frameworks in which metaphor is decomposed into interacting levels such as content analysis, conceptual mapping or blending, contextual reasoning, pragmatic intentionality, and, in some accounts, task or diagnostic interfaces (Cappa et al., 14 Jul 2025, Ye et al., 5 Oct 2025, Mu et al., 2019). Multimodal and mechanistic variants extend this layered view to hierarchical target–source identification, schema transfer, and cross-depth transformer dynamics (Zhang et al., 5 Jan 2025, Xu et al., 1 Feb 2026, Chakrabarti et al., 20 May 2026).

1. Intellectual background and scope

The stratified view arose from dissatisfaction with flat accounts of metaphor. One line of work grounds metaphor in Conceptual Metaphor Theory, where source domains transfer structured properties to target domains; another emphasizes pragmatic theories in which metaphor interpretation depends on violations of conversational expectations, speaker meaning, and broader context (Kramer, 4 Feb 2025, Mu et al., 2019). A third line formalizes metaphor as recursive pragmatic inference, as in Rational Speech Act models, where literal decoding, goal-driven utterance choice, and pragmatic interpretation are distinct inferential stages (Carenini et al., 2024).

Within this literature, “stratified” does not name a single canonical architecture. Rather, it denotes the recurrent observation that metaphor processing engages different representational levels. “Meanings are like Onions” states this most explicitly by proposing three layers—content analysis, conceptual blending, and pragmatic intentionality (Cappa et al., 14 Jul 2025). “Unveiling LLMs’ Metaphorical Understanding” reaches a similar conclusion indirectly: although it does not present an explicit layered architecture, its three experimental axes naturally induce a stratified model comprising concept mapping, a metaphor–literal repository, syntactic sensitivity, contextual reasoning, and a prompt/task interface (Ye et al., 5 Oct 2025). “Learning Outside the Box” similarly organizes metaphor identification into lexical, local syntactic, and discourse strata, with paragraph context functioning as a higher-order pragmatic layer (Mu et al., 2019).

This convergence suggests that a stratified model is best understood as an analytic framework for separating different sources of metaphor competence and failure. It distinguishes lexical association from structural alignment, structural alignment from discourse use, and discourse use from pragmatic effect, rather than assuming that one score or one representation exhausts metaphor understanding.

2. Recurrent strata of metaphor processing

A plausible synthesis of the literature distinguishes several recurrent strata. These strata are not always all present in one model, but they recur across symbolic, probabilistic, neural, and multimodal systems.

Content analysis or lexical-content stratum refers to the initial representation of the metaphor-bearing artifact. In the onion model, this layer records domains, provenance, frame links, and annotator metadata; in discourse-based metaphor identification it appears as the lemma-only level; and in LLM work it also includes retrieval-like activation of stereotypical metaphor–literal associations (Cappa et al., 14 Jul 2025, Mu et al., 2019, Ye et al., 5 Oct 2025). This layer answers what is present in the input before deeper reinterpretation begins.

Conceptual mapping or blending stratum handles the relation between source and target domains. In some systems this is modeled geometrically, as alignment with a conceptual plane; in others it is realized through source–target domain extraction, blending principles, or prototype combination. Hierarchical multimodal models make this stratum explicit by treating target identification as a lower-level problem and source identification as an upper-level problem conditioned on the target (Ye et al., 5 Oct 2025, Zhang et al., 5 Jan 2025, Cappa et al., 14 Jul 2025).

Syntactic and structural stratum captures word order, POS configuration, argument structure, and relational form. In discourse-level metaphor identification, adding subject and direct object embeddings consistently improves over lemma-only models. In LLM probing, syntactic irregularity acts as a cue for metaphoricity, but often more as anomaly detection than as deep structural comprehension (Mu et al., 2019, Ye et al., 5 Oct 2025).

Contextual and discourse stratum uses sentence, paragraph, genre, and world knowledge to disambiguate metaphor. This stratum is central to pragmatic accounts and is operationalized computationally by paragraph embeddings, discourse windows, or genre-specific corpora. It becomes decisive when literal and metaphorical readings cannot be separated by local lexical cues alone (Mu et al., 2019, Mangiaterra et al., 14 Feb 2026).

Pragmatic intentionality stratum captures speaker attitude, communicative function, illocutionary force, perlocutionary effect, and tone. The onion model treats this as the innermost layer, introducing categories such as Attitude, Illocutionary Act, Directive Kind, Perlocutionary Effect, Efficacy, and Tone of Voice (Cappa et al., 14 Jul 2025). This layer explains not only what a metaphor means but what it does.

Task-interface and diagnostic strata appear in recent LLM and multimodal systems. “Unveiling LLMs’ Metaphorical Understanding” isolates prompt framing as methodologically crucial because multiple-choice evaluation can fail even when models produce plausible paraphrases. Visual Metaphor Transfer introduces a hierarchical diagnostic agent that attributes failures to prompt-level, component-level, or abstraction-level errors (Ye et al., 5 Oct 2025, Xu et al., 1 Feb 2026). These strata do not belong to metaphor semantics in the narrow sense, but they strongly affect measured performance.

3. Formal computational realizations

Several formalisms make the layered structure explicit. A geometric realization appears in concept-mapping work on LLMs. Given reference paraphrases $R_1$, $R_2$, an additional literal sentence $S$, and a model interpretation $M_i$, sentence embeddings define a conceptual plane $\gamma_i$ and an interpretation plane $\beta_i$. The planes are estimated with SVD,

$A = U \Sigma V^T$

and

$A_r = U_k \Sigma_k V^T_k,$

and conceptual irrelevance is then measured by perpendicular distance $d_p$ from $M_i$ to $R_2$0 and planar cosine similarity $R_2$1 between $R_2$2 and $R_2$3 (Ye et al., 5 Oct 2025). In this formulation, the highest semantic stratum is not mere metaphor detection but alignment with the intended abstract domain.

A probabilistic-pragmatic realization is provided by the RSA model for metaphor understanding. Its literal listener is

$R_2$4

its speaker is

$R_2$5

and its pragmatic listener is

$R_2$6

Here the strata are literal semantics, communicative goals, pragmatic production, and pragmatic interpretation. The model learns $R_2$7 and achieves $R_2$8 overall correlation with human interpretation distributions, rising to $R_2$9 for vehicle-inherent metaphors and falling to $S$0 for non-vehicle-inherent ones (Carenini et al., 2024). This suggests that typicality-based Bayesian models capture a substantial but bounded portion of metaphor processing.

A hierarchical multimodal realization appears in CPMMIM. The lower-level problem identifies the target domain and the upper-level problem identifies the source domain conditioned on that target:

$S$1

Specialized to metaphor mapping identification, the model optimizes source extraction subject to the target-domain solution, with $S$2 and $S$3 supplied by Chain-of-Thought prompts (Zhang et al., 5 Jan 2025). The same work combines BART, ViT-L/16, cross-modal attention,

$S$4

and gated fusion,

$S$5

$S$6

thereby separating perceptual, interaction, reasoning, and decoding stages (Zhang et al., 5 Jan 2025).

A structural-symbolic realization is provided by TINT. Here metaphor comprehension is a mapping between coslice categories, with a base-of-metaphor functor

$S$7

and a natural transformation

$S$8

This yields a layered account in which associative networks form the base, local meanings are represented as coslice categories, and coherent metaphor interpretation emerges through functorial mapping and naturality constraints (Iwaki et al., 11 Apr 2026). In this framework, the intermediate structural layer is not optional: relation-based algorithms outperform object-based ones in data fitting, systematicity, and novelty.

4. Empirical findings and diagnostic controversies

Empirical work on LLMs shows that the higher strata remain fragile. Spatial analysis of model interpretations estimates that roughly 15–25% of interpretations exhibit substantial conceptual irrelevance. In the same study, multiple-choice interpretation hovers around 45–51% accuracy, indicating that option selection is an unstable probe of conceptual mapping. The same paper also reports 65–80% overlap between with-context and without-context metaphor–literal generation, with more than half of cases with overlap ratio = 1, which it interprets as evidence for a context-insensitive metaphor–literal repository (Ye et al., 5 Oct 2025).

Syntactic probing strengthens this diagnosis. In metaphor detection, POS shuffle often yields higher accuracy than original, while random shuffle usually collapses performance. For example, GPT-4 rises from 34.73 on original sentences to 43.74 under POS shuffle, and GPT-4o rises from 28.89 to 36.87; by contrast, random shuffle drops GPT-4 from 34.73 to 12.93 and GPT-4o from 28.89 to 7.78 (Ye et al., 5 Oct 2025). This supports the claim that many models respond to overt irregularity as a cue for metaphoricity, rather than reconstructing deeper compositional structure.

A major controversy concerns benchmark validity. “Metaphor and LLMs: When Surface Features Matter More than Deep Understanding” shows that performance on metaphor-oriented NLI and QA tasks tracks lexical overlap and sentence length more strongly than metaphoricity. Under CoT prompting, Qwen2.5–72B reaches an average of 91.22 on the original metaphor sets but 85.82 on literal paraphrase versions with the same labels, and the authors attribute this gap to higher lexical overlap and shorter sentences in the original datasets (Sanchez-Bayona et al., 21 Jul 2025). This directly challenges the common assumption that strong benchmark scores imply robust conceptual metaphor processing.

A related diagnostic reframes the field in terms of difficulty strata. “Finding Challenging Metaphors that Confuse Pretrained LLMs” argues that most VUA metaphors are easy for modern models and defines hard metaphors by low overlap ratio $S$9 in sense-representation space. The resulting Hard Metaphor Dataset contains 21k examples over 82 words and 110 metaphorical senses. On this subset, machine translation drops by 16–18%, NLI by 6–7%, QA by 2–3%, and metaphor identification recall by over 14% (Li et al., 2024). This suggests that a stratified model should distinguish not only processing levels but also levels of metaphor difficulty.

5. Multimodal, agentic, and mechanistic extensions

Stratified modeling is especially explicit in multimodal work. CM3D introduces 6,108 Chinese advertisements annotated with target and source domains, and CPMMIM turns metaphor mapping into a two-stage hierarchy: target-domain identification followed by source-domain identification (Zhang et al., 5 Jan 2025). On this task, CPMMIM improves target-domain performance over the best baseline from 32.13 to 36.23 Accuracy and from 53.28 to 58.52 Human Evaluation, and source-domain performance from 25.38 to 28.47 Accuracy and from 46.89 to 50.49 Human Evaluation (Zhang et al., 5 Jan 2025). Ablation shows that removing the first CoT stage reduces target H-E from 58.52 to 53.24 and source H-E from 50.49 to 47.70, while removing the second CoT stage reduces source H-E to 48.52 (Zhang et al., 5 Jan 2025). This is strong evidence that lower and upper layers are functionally coupled.

Visual Metaphor Transfer expands the idea into a multi-agent architecture. Its central intermediate representation is a Schema Grammar

$M_i$0

where $M_i$1 is subject, $M_i$2 carrier, $M_i$3 relational invariants, $M_i$4 violation/conflict points, and $M_i$5 emergent meaning (Xu et al., 1 Feb 2026). A perception agent extracts $M_i$6, a transfer agent preserves generic-space invariance while adapting the schema to a new subject, a generation agent realizes the target schema, and a diagnostic agent assigns failures to prompt-level, component-level, or abstraction-level errors. The evaluation framework uses Metaphor Consistency, Analogy Appropriateness, and Conceptual Integration, and reports 98.2% agreement between LLM-ensemble judgments and human ratings in a preliminary study (Xu et al., 1 Feb 2026). This work makes the diagnostic stratum explicit and closed-loop.

Mechanistic interpretability extends stratification into transformer depth. Using residual trajectories

$M_i$7

a contrast direction $M_i$8, wavelet responses, and conditional scale entropy

$M_i$9

CSE analysis finds that metaphorical tokens produce higher spectral breadth than literal tokens at contiguous layer positions across GPT-2 Small, GPT-2 Medium, GPT-2 Large, LLaMA-2 7B, and GPT-oss 20B (Chakrabarti et al., 20 May 2026). Significant active zones fall in early-to-mid relative depth, such as layers 5–13 in GPT-2 Medium and 5–11 in LLaMA-2 7B, and the effect survives cluster-based permutation correction (Chakrabarti et al., 20 May 2026). Because CSE is invariant to uniform scaling of updates, the result isolates structural coordination across depth rather than raw update magnitude. This suggests that a mechanistic stratum of metaphor processing can be localized as a depth band of reinterpretive reconfiguration.

6. Diachronic variation, limitations, and open problems

Temporal and genre-sensitive work adds macro- and meso-strata to the model. Using TWEC embeddings trained on 124 million tokens, “Metaphors’ journeys across time and genre” tracks 515 nineteenth-century literary metaphors across nineteenth- and twenty-first-century literary and nonliterary Italian corpora (Mangiaterra et al., 14 Feb 2026). Topic–vehicle cosine similarity is used as a proxy for processing cost. The overall temporal effect is null, but genre interacts strongly with epoch: mean similarity is 0.31 in nineteenth-century literary language, 0.27 in twenty-first-century literary language, 0.30 in nineteenth-century nonliterary language, and 0.33 in twenty-first-century nonliterary language (Mangiaterra et al., 14 Feb 2026). The base mixed model reports no main effect of Epoch, a main effect of Genre with $\gamma_i$0, $\gamma_i$1, $\gamma_i$2, and an Epoch × Genre interaction with $\gamma_i$3, $\gamma_i$4, $\gamma_i$5 (Mangiaterra et al., 14 Feb 2026). This suggests that a stratified account of metaphor must include not only micro-level semantic operations but also historically variable semantic spaces.

Word-level properties further refine this picture. In the extended model, Topic SND has $\gamma_i$6, $\gamma_i$7, $\gamma_i$8, Vehicle SND has $\gamma_i$9, $\beta_i$0, $\beta_i$1, Topic VC × Genre × Epoch has $\beta_i$2, $\beta_i$3, $\beta_i$4, and Vehicle SND × Genre × Epoch has $\beta_i$5, $\beta_i$6, $\beta_i$7 (Mangiaterra et al., 14 Feb 2026). In that formulation, metaphor difficulty depends on time, genre, topic stability, and semantic neighborhood density, not just on a timeless source–target distance.

The current literature also makes the limits of existing strata explicit. The geometric LLM study notes limitations of SVD-based planes, English-only scope, use of models “as is,” and the focus on single-sentence metaphors (Ye et al., 5 Oct 2025). The RSA model captures human distributions well when intended meanings capitalize on properties inherent to the vehicle concept, but it identifies “the more creative nuances of metaphorical meaning, not strictly encoded in the lexical concepts,” as a challenge for machines (Carenini et al., 2024). The onion model itself is programmatic: it proposes an integrated content–blend–pragmatics framework but does not yet provide an end-to-end implementation (Cappa et al., 14 Jul 2025).

Open questions therefore remain at every level. One set concerns representation: how to represent source and target domains explicitly, how to disentangle lexical repository activation from syntactic anomaly detection and true conceptual mapping, and how to define objective functions that reduce concept-irrelevant interpretations (Ye et al., 5 Oct 2025). Another concerns benchmarking: how to build evaluations that are not dominated by lexical overlap, sentence length, or model-specific easy cases (Sanchez-Bayona et al., 21 Jul 2025, Li et al., 2024). A third concerns integration: how to connect symbolic blending, probabilistic pragmatics, multimodal schemas, and mechanistic transformer analysis into a unified stratified theory without collapsing one layer into another. The cumulative record suggests that metaphor processing is best treated not as a single capability but as a stack of partially separable operations whose interactions are still only partly understood.