Metaphor Decomposition Score

• Metaphor Decomposition Score is a family of scoring approaches that quantify how accurately systems decompose metaphors into key semantic components such as source, target, and intended meaning.
• It employs diverse frameworks—from contrastive basic versus contextual meaning to explicit Source–Target–Meaning and tenor–vehicle–ground decompositions—to support tasks like metaphor detection, visual generation, and translation.
• Empirical studies show its discriminative power in tasks including visual metaphor generation and machine translation, while highlighting limitations like evaluator bias and domain specificity.

Metaphor Decomposition Score denotes a family of scoring approaches that quantify how a system decomposes a metaphor into its constitutive semantic structure and how well that decomposition aligns with the intended figurative meaning. The term is used explicitly in visual metaphor generation as a scalar that evaluates a Source–Target–Meaning decomposition before image synthesis (Koushik et al., 26 Aug 2025). Closely related formulations appear in metaphor detection, where metaphoricity is treated as contrast between contextual and basic meanings; in metaphor understanding, where chain-of-thought prompts induce latent variables such as question under discussion and shared properties; in machine translation, where figurative quality is decomposed into Metaphorical Equivalence, Emotion, Authenticity, and Quality; in Chinese metaphor recognition, where tenor, vehicle, and ground are identified with confidence scores; and in cognitive modeling, where structural metaphor mechanisms are scored through the Minimal Cognitive Grid (Li et al., 2023, Prystawski et al., 2022, Wang et al., 2024, Wang et al., 2024, Donvito et al., 2 May 2026).

1. Theoretical bases of metaphor decomposition

A central foundation is the Metaphor Identification Procedure, which defines the metaphoricity of a lexical unit through the contrast between its contextual meaning and its basic meaning. For a lexical unit in a sentence, the procedure determines its contextual meaning, determines its basic meaning as the “more concrete; related to bodily action; more precise; historically older” sense, and classifies the use as metaphorical if the contextual meaning contrasts with the basic meaning but can be understood in comparison with it. The example “This project is such a headache!” is analyzed by contrasting the basic meaning of “headache,” “a continuous pain in the head,” with the contextual meaning “a thing or person that causes worry or trouble; a problem” (Li et al., 2023).

A second line of work grounds decomposition in Conceptual Metaphor Theory and requires explicit Source–Target–Meaning structure. In this formulation, the source is the concrete, familiar domain, the target is the abstract or less-understood domain, and the meaning is the intended conceptual mapping or entailment. The decomposition is treated as a prerequisite for faithful visual metaphor generation, because a visually coherent image can still be conceptually wrong if the underlying Source–Target–Meaning analysis is wrong (Koushik et al., 26 Aug 2025).

A third formulation appears in cognitive-psychology-inspired prompting for metaphor understanding. Here the latent variables are not only subject and object, but also question under discussion and relevant properties. In the question-under-discussion formulation, metaphor comprehension involves inferring what implicit question is being answered, which property of the source answers that question, and which corresponding property should be attributed to the target. In similarity-based prompting, the structure is simpler and centers on the target, the source, and a selected shared feature (Prystawski et al., 2022).

A fourth formulation is componential rather than contrastive. Chinese metaphor recognition uses tenor, vehicle, and ground as the operative decomposition. The tenor is the subject, the vehicle is the comparative element or source domain, and the ground is the shared property or metaphorical meaning linking tenor and vehicle. The paper emphasizes that tenor may be absent from the surface string and must sometimes be inferred, as in the example where “瀑布” is the inferred tenor for “飞流直下三千尺，疑是银河落九天” (Wang et al., 2024).

These frameworks do not define a single universal score. They instead specify different decomposition targets: basic versus contextual meaning, Source–Target–Meaning, subject–object–property or question-under-discussion structure, tenor–vehicle–ground, and explicit structural mechanisms for analogy and categorization. This suggests that “Metaphor Decomposition Score” is best understood as a task-dependent formalization of metaphor structure rather than a single standardized metric.

2. Formalizations and scoring functions

The most explicit use of the term appears in visual metaphor generation. There, the score is produced by a judge model using a specialized prompt that asks whether the proposed source correctly identifies the concrete concept, whether the target correctly identifies the abstract concept, and whether the meaning accurately captures the intended connection. The judge returns a single score from 0.0 to 1.0 inside <decomposition_score> tags, and this scalar is used as one reward component in the overall reward function (Koushik et al., 26 Aug 2025).

Formally, the decomposition reward is written as

$$r_{\text{decomp}}(M_{\text{text}}, S, T, M) = \operatorname{parse\_float}\big( J(\text{MetaphorDecompositionPrompt}(M_{\text{text}}, S, T, M)) \big)$$

and the multi-faceted reward is

$$\mathcal{R}(I, P, D) = \sum_{k=1}^{K} w_k\, r_k(I, P, D).$$

The decomposition reward and CLIP score each receive weight 0.20, while the remaining metrics receive weight 0.10 (Koushik et al., 26 Aug 2025).

In metaphor detection via explicit basic meanings modeling, the paper does not introduce a scalar explicitly named Metaphor Decomposition Score. It instead encodes contextual meaning and basic meaning in vector form, combines them through linear layers, and produces a prediction score $\hat{y}\in(0,1)$ interpreted as the probability that the target word is metaphorical. It also measures contrast directly through cosine similarity. The contextual embedding is

$v_{S,t} = h_t,$

the basic meaning prototype is

$$v_{B,t} = \frac{1}{|B_t|}\sum_{S_b \in B_t} h_t^{(b)},$$

and the final prediction is

$$\hat{y} = \sigma\left(W^\top [h_{\text{BMIP}}; h_{\text{AMIP}}; h_{\text{SPV}}] + b\right).$$

The paper’s structured explanation states that a natural decomposition score could be

$\text{MDS}(w_t,S) = 1 - \cos(v_{S,t}, v_{B,t}),$

with larger values indicating greater contrast and thus higher metaphoricity. This is an interpretation built from the reported cosine-contrast analysis rather than a named score in the original model (Li et al., 2023).

In machine translation evaluation, no metric is explicitly called Metaphor Decomposition Score, but the proposed evaluation protocol decomposes figurative quality into Metaphorical Equivalence, Emotion, Authenticity, and Quality. The structured explanation then defines either a vector

$$\mathbf{M}^{(i)} = \bigl( s_{ME}^{(i)}, s_{EMO}^{(i)}, s_{AUTH}^{(i)}, s_{Q}^{(i)} \bigr)$$

or a scalar composite

$$S_{MDS}^{(i)} = w_{ME} \, s_{ME}^{(i)} + w_{EMO} \, s_{EMO}^{(i)} + w_{AUTH} \, s_{AUTH}^{(i)} + w_{Q} \, s_{Q}^{(i)}.$$

This is again a decomposition-based formulation constructed from the paper’s evaluation dimensions (Wang et al., 2024).

In Chinese metaphor recognition, the paper provides the scoring primitives rather than a named decomposition score. A DeBERTa model assigns confidence scores

$$p_i=\text{softmax}(W\mathcal{M}^{\text{DeBERTa}}(T_i)+b)$$

and is trained with

$$\mathcal{R}(I, P, D) = \sum_{k=1}^{K} w_k\, r_k(I, P, D).$$0

The structured explanation proposes possible joint confidence formulations such as a product or geometric mean over tenor, vehicle, and ground confidences, but those are presented as an overview built from the paper’s confidence-based design rather than as an explicit metric in the original shared-task system (Wang et al., 2024).

In the Minimal Cognitive Grid, the closest approximation to a decomposition score is the structural score

$$\mathcal{R}(I, P, D) = \sum_{k=1}^{K} w_k\, r_k(I, P, D).$$1

where the six constraints cover one-to-one mapping, parallel connectivity, systematicity, inferential projection, categorization, and property selection. The normalized structurality index is

$$\mathcal{R}(I, P, D) = \sum_{k=1}^{K} w_k\, r_k(I, P, D).$$2

Within that framework, this score functions as a quantitative measure of how structurally a system decomposes and processes metaphors (Donvito et al., 2 May 2026).

3. Computational workflows and representation design

The architectures underlying these scores differ substantially. In BasicBERT, the core decomposition branch is BasicMIP, which compares the contextual meaning of a target word with a basic meaning prototype derived from literal training instances. Two additional branches are AMIP, which compares contextual meaning with aggregated or frequent meaning, and SPV, which compares the target’s contextual meaning with sentence-level meaning. The basic meaning is not taken from an external sense inventory; it is computed by averaging contextualized embeddings across literal occurrences of the target word in the training set. For targets without literal annotations, the model falls back to a decontextualized representation (Li et al., 2023).

The visual metaphor framework begins with an LLM that analyzes the input metaphor and emits source, target, intended meaning, and a visual prompt. A judge model then evaluates the Source–Target–Meaning decomposition. In the training-free pipeline, this decomposition score is computed once per metaphor and reused across prompt-refinement iterations. In the GRPO-based training pipeline, multiple candidate completions are sampled, each candidate includes a full decomposition plus prompt, and the decomposition score enters directly into the scalar reward used to optimize the LLM parameters (Koushik et al., 26 Aug 2025).

The Chinese multi-stage prompting system separates scoring and reasoning. Stage I trains a small DeBERTa model to obtain confidence scores for answer candidate generation. Stage II clusters questions with k-means, uses the elbow method to choose $$\mathcal{R}(I, P, D) = \sum_{k=1}^{K} w_k\, r_k(I, P, D).$$3 for both subtasks, and generates chain-of-thought demonstrations with “Let’s think step by step.” The final heuristic-enhanced prompt combines demonstrations, the test question, and the confidence-ranked answer candidates. The LLM is thus exposed both to explicit candidate probabilities and to representative reasoning traces about tenor, vehicle, and ground (Wang et al., 2024).

Psychologically informed chain-of-thought prompting implements another kind of decomposition pipeline. Each prompt contains an instruction, ten examples with paraphrase options and rationale paragraphs, and a test instance. The question-under-discussion prompt explicitly identifies the implicit question, the compared entities, a relevant property of the source, and the corresponding property of the target before selecting a paraphrase. The similarity prompt identifies a property of the object, attributes that property to the subject, and then chooses the paraphrase. These rationales act as latent-structure scaffolds rather than as direct numerical scores (Prystawski et al., 2022).

The common computational pattern is explicit intermediate structure. Whether the structure is vector contrast, Source–Target–Meaning, tenor–vehicle–ground, or constraint satisfaction, the score is built on a decomposition that is represented before the final decision or generation step. This contrasts with purely end-to-end outputs that produce a classification, paraphrase, or image without exposing the internal metaphor analysis.

4. Empirical behavior and discriminative power

In explicit basic meanings modeling, empirical results show that the method outperforms the state-of-the-art method significantly by 1.0% in F1 score. On VUA18, BasicBERT reaches F1 = 79.0, compared with FrameBERT’s 78.8 and MelBERT’s 78.5; on VUA20, it reaches 73.3, compared with 73.0 and 72.3. Ablation shows that removing BasicMIP drops VUA18 from 79.0 to 78.3 and VUA20 from 73.3 to 72.2. For targets with literal annotations, VUA18 reaches F1 = 81.1 and the inter-annotator agreement is around 0.8, which the authors describe as reaching the theoretical upper bound (Li et al., 2023).

The same paper reports especially sharp contrast separation in cosine similarity. For contextual versus frequent meaning, cosine similarity is 0.516 for metaphor and 0.642 for literal. For contextual versus basic meaning, it is -0.082 for metaphor and 0.809 for literal. The gap is therefore much larger for BasicMIP than for AMIP. This is the strongest empirical argument in that work for a decomposition score based on basic-versus-contextual contrast (Li et al., 2023).

In visual metaphor generation, average decomposition scores on the held-out test set differ substantially across systems. The training-free configuration with Gemma-3-27B + Janus-Pro-7B + Qwen-VL-32B reports Decomposition Score 0.8668, CLIP 0.2960, and MA 0.8760. The training-free configuration with Gemma-3-12B + SD-3.5 + Qwen-VL-7B reports Decomposition Score 0.7212. GRPO variants with Gemma-3-4B report decomposition scores in the range of approximately 0.6480–0.7076. In zero-shot comparison, GPT-4o reports Decomposition Score 0.8072 (Koushik et al., 26 Aug 2025).

The same study shows that the highest decomposition score is associated with the highest MA score, best target-presence, and a strong CLIP score, but not with overall human preference. Participants preferred GPT-4o overall, while the training-free pipeline led open-source methods and edged Imagen on abstract metaphors. The paper attributes the remaining gap to aesthetics and sampling rather than to decomposition alone (Koushik et al., 26 Aug 2025).

In Chinese metaphor recognition, the staged design yields large gains. The LLM-only baseline reaches accuracy 0.73. With DeBERTa-unfinetuned candidates, accuracy rises to 0.766. With DeBERTa-finetuned candidates, it reaches 0.964. With both candidates and demonstrations, the full method reaches 0.98 on validation and official averages of about 0.959, 0.979, 0.951, and 0.941 across the shared-task tracks and subtasks (Wang et al., 2024).

In machine translation evaluation, metaphorical expressions display different traits from literal ones. Full-equivalence for metaphorical expressions is only about 56.6%, while non-equivalence plus mistranslation is about 30%. For literal expressions, literal equivalents are about 82.9%. The paper also reports that metaphorical translations obtain lower Fluency, Intelligibility, Fidelity, Authenticity, and Overall scores than literal translations, while restricting analysis to Full-equivalence metaphor cases raises scores substantially (Wang et al., 2024).

These results support a recurring pattern: decomposition-sensitive evaluation is most discriminative when the model has access to explicit structural evidence, such as literal annotations, Source–Target–Meaning analysis, candidate confidence scores, or human figurative labels.

5. Applications across research areas

In text metaphor detection, decomposition scoring functions as a diagnostic of metaphoricity. The basic claim is that literal and metaphorical usages can be separated by modeling basic meanings explicitly rather than approximating them with aggregated meanings. This makes the score useful not only for binary classification but also for analyzing why metaphor and literal cases diverge in representation space (Li et al., 2023).

In visual metaphor generation, the decomposition score is an upstream semantic-alignment signal. It is computed at the text level and does not inspect the generated image directly. Its role is to validate that the system has correctly identified what is concrete, what is abstract, and what conceptual relation should be visualized. Image-level metrics such as CLIPScore, source presence, target presence, meaning alignment, and BERTScore similarities then evaluate whether the image expresses that decomposition (Koushik et al., 26 Aug 2025).

In metaphor understanding, chain-of-thought prompting uses decomposition as an interpretability device. The target task is paraphrase selection for metaphors of the form “subject is object,” and the primary metric is mean appropriateness score over four candidate paraphrases with human-defined appropriateness levels 1–4. The paper’s structured explanation proposes extending this setup into a decomposition score by combining paraphrase appropriateness, latent-variable correctness, and reasoning quality, but this is a downstream synthesis rather than a metric introduced in the original paper (Prystawski et al., 2022).

In machine translation, metaphor decomposition becomes multi-dimensional quality assessment. Metaphorical Equivalence measures whether figurativity and conceptual alignment are preserved, Emotion measures affective preservation, Authenticity measures whether the metaphor sounds natural in the target language, and Quality aggregates Fluency, Intelligibility, Fidelity, and Overall quality. This formulation is particularly important because mainstream MT metrics are described as paying little attention to figurative quality (Wang et al., 2024).

In Chinese metaphor recognition, decomposition is the task itself. One shared subtask takes a symbolic sentence and requires the system to extract TENOR, GROUND, and VEHICLE; the other takes a tuple of those components and requires metaphor generation. Because the dataset contains 34,463 metaphorical sentences annotated with TENOR, VEHICLE, and GROUND, the setting is well suited to component-level scoring and confidence analysis (Wang et al., 2024).

In cognitive modeling, decomposition scoring is used to rank the cognitive plausibility of computational systems. The Minimal Cognitive Grid evaluates whether a system implements metaphor-relevant mechanisms such as systematicity, categorization, and property selection, and combines structurality, generality, and performance match into a cognitive plausibility score. In this setting, the score is not about task accuracy alone; it is about whether the system’s internal processing aligns with cognitive theories of analogy and metaphor (Donvito et al., 2 May 2026).

6. Limitations, interpretation, and points of dispute

A first limitation is the absence of a single standardized meaning for the term. Only the visual metaphor generation paper explicitly names a metaphor decomposition score. Other works provide contrast metrics, component confidence scores, multi-dimensional figurative metrics, or structurality indices that can serve analogous purposes, but they are not presented as one shared benchmark (Koushik et al., 26 Aug 2025, Li et al., 2023, Wang et al., 2024, Wang et al., 2024, Donvito et al., 2 May 2026).

A second limitation concerns annotation and coverage. In explicit basic meanings modeling, some target words lack literal instances, so the basic prototype degenerates to a decontextualized embedding; the benefit of BasicMIP is strongest when genuine basic prototypes exist. The model also uses a single basic prototype per word, so words with multiple literal or bodily senses are collapsed into one averaged representation (Li et al., 2023).

A third limitation is evaluator dependence. In the visual metaphor framework, the decomposition score is itself produced by a LLM acting as a linguistic critic. The authors note that this can inject biases and that the judge’s understanding of metaphor is not perfect. They also note that the score does not capture aesthetics, even though aesthetics strongly affect human preference (Koushik et al., 26 Aug 2025).

A fourth limitation is domain and language specificity. The machine translation framework is based on EN–ZH and EN–IT with MOH-derived sentences, mainly verb metaphors, and the paper notes that it does not cover the full range of metaphor types such as extended metaphors, idioms, novel creative metaphors, and multi-word expressions. The Chinese shared-task system is likewise tied to dataset format, option structure, and prompt design (Wang et al., 2024, Wang et al., 2024).

A fifth limitation is the gap between performance and structural explanation. The Minimal Cognitive Grid ranks LLMs low on structural metaphor decomposition, with $$\mathcal{R}(I, P, D) = \sum_{k=1}^{K} w_k\, r_k(I, P, D).$$4 and normalized $$\mathcal{R}(I, P, D) = \sum_{k=1}^{K} w_k\, r_k(I, P, D).$$5, even though LLMs may perform competitively on some downstream tasks. This indicates that high behavioral accuracy does not by itself imply explicit structural metaphor processing (Donvito et al., 2 May 2026).

Two common misconceptions follow from these limitations. One is that a high decomposition score necessarily implies superior end-user preference; the visual metaphor results show otherwise, because human judgments were strongly affected by visual style and sampling. Another is that decomposition scores are necessarily image-based or necessarily scalar. In fact, the literature includes text-only judge scores, vector-contrast analyses, component confidence scores, categorical evaluation schemes, and weighted structural indices (Koushik et al., 26 Aug 2025, Wang et al., 2024, Li et al., 2023).

Taken together, these works treat metaphor decomposition not as a single canonical metric but as a principled way to expose and evaluate the internal structure of figurative interpretation. The dominant technical question is not whether a metaphor score exists, but which decomposition is theoretically privileged for the task: basic versus contextual meaning, Source–Target–Meaning, tenor–vehicle–ground, latent pragmatic variables, figurative quality dimensions, or explicit structural constraints.