MetFuse: Figurative Fusion between Metonymy and Metaphor

Saptarshi Ghosh; Tianyu Jiang

cs.CLApr 14, 2026

MetFuse: Figurative Fusion between Metonymy and Metaphor

Authors: Saptarshi Ghosh, Tianyu Jiang

Organizations: University of Cincinnati

Abstract

Metonymy and metaphor often co-occur in natural language, yet computational work has studied them largely in isolation. We introduce a framework that transforms a literal sentence into three figurative variants: metonymic, metaphoric, and hybrid. Using this framework, we construct MetFuse, the first dedicated dataset of figurative fusion between metonymy and metaphor, containing 1,000 human-verified meaning-aligned quadruplets totaling 4,000 sentences. Extrinsic experiments on eight existing benchmarks show that augmenting training data with MetFuse consistently improves both metonymy and metaphor classification, with hybrid examples yielding the largest gains on metonymy tasks. Using this dataset, we also analyze how the presence of one figurative type impacts another. Our findings show that both human annotators and large language models better identify metonymy in hybrid sentences than in metonymy-only sentences, demonstrating that the presence of a metaphor makes a metonymic noun more explicit. Our dataset is publicly available at: https://github.com/cincynlp/MetFuse.

Figures & tables

Figure 1: Illustration of generating metonymy, metaphor and hybrid sentences from literal expressions.

Noun	Literal Input	Candidates	Scores	Metonymic Sentence
judge	The [mask] in Lincoln saw issues	law office	-4.25	The law office in Lincoln saw issues
		briefcase	-12.28
		Springfield	-11.96
Queen	The [mask] officially pronounced Turing pardoned in August 2014.	crown	-1.58	Buckingham Palace officially pronounced Turing pardoned in August 2014.
		monarchy	-11.58
		Buckingham Palace	-0.27

Table 1: Illustrations of transforming literal expressions into metonymic ones using an LLM to generate candidate substitutions and masked token probabilities to select the best replacement.

	Metonymy	Metaphor	Hybrid
General	38.8%	70.8%	49.2%
Ours	75.2%	84.0%	74.0%

Table 2: Percentage of sentences evaluated by humans to have the intended figurative expression in a sample of 250. General row are sentences from basic prompting. The sentences generated by our framework are significant better with figurative expressions.

	Metonymy	Metaphor	Hybrid
General	0.70	0.60	0.44
Ours	0.84	0.82	0.70

Table 3: Cosine similarity score between the original literal sentence and the generated variant using a sentence transformer. Our framework better preserves the semantic meaning of the sentence.

Literal Input	Metonymy	Metaphor	Hybrid
The researchers formed a rational statement of his question.	The laboratory formed a rational statement of his question.	The researchers sculpted a rational statement of his question.	The laboratory sculpted a rational statement of his question.
The reporter couldn’t have done too good a job on you.	The newsroom couldn’t have done too good a job on you.	The reporter butchered you in that interview.	The newsroom butchered you in that interview.
The Queen officially pronounced Turing pardoned in august 2014.	Buckingham Palace officially pronounced Turing pardoned in august 2014.	The Queen’s pardon for Turing thundered through history in August 2014.	Buckingham Palace’s pardon for Turing thundered through history in August 2014.
The police exposed the crime ring in 1956.	NYPD exposed the crime ring in 1956.	The police unearthed the crime ring in 1956.	NYPD unearthed the crime ring in 1956.

Table 4: Example of a literal expression and its respective figurative variations from MetFuse dataset.

	Metonymy	Metaphor	Hybrid
Fluency	3.30	3.64	3.61
Meaning	3.51	3.10	2.74
Creativity	2.95	4.01	4.25
Metonymicity	3.47	-	3.65
Metaphoricity	-	3.95	3.82

Table 5: Human score on five criteria from 250 samples from the MetFuse dataset.

	Train	Train + $\text{MTF}_{mty}$	Train + $\text{MTF}_{hyb}$
ConMeC	75.49	76.71 (+1.22)	79.33 (+3.84)
Pedinotti	68.42	66.92 (-1.50)	70.44 (+2.02)
RelocaR	67.33	69.99 (+2.66)	70.67 (+3.34)
WiMCor	81.67	82.33 (+0.66)	82.67 (+1.00)

Table 6: Test accuracy for downstream tasks on metonymy datasets. Each dataset is split in a 70-30 train-test ratio. Train = original training sample, Train+

\text{MTF}_{mty}

= original training sample augmented with metonymic examples from MetFuse, Train+

\text{MTF}_{hyb}

= original training sample augmented with hybrid examples from MetFuse.

	Train	Train + $\text{MTF}_{mtr}$	Train + $\text{MTF}_{hyb}$
VUA Verb	64.38	64.53 (+0.15)	66.55 (+2.17)
Flute	76.26	79.73 (+3.47)	78.47 (+2.21)
MOH-X	76.26	76.92 (+0.16)	77.43 (+0.51)
TroFi	60.42	63.30 (+2.88)	61.76 (+1.34)

Table 7: Test accuracy for downstream tasks on metaphor datasets. Each dataset is split in a 70-30 train-test ratio. Train = original training sample, Train+

\text{MTF}_{mty}

= original training sample augmented with metaphor examples from MetFuse, Train+

\text{MTF}_{hyb}

= original training sample augmented with hybrid examples from MetFuse.

Model	Metonymy			Hybrid
Model	Pre	Rec	F1	Pre	Rec	F1
GPT-OSS-20B	95.8	51.9	67.3	96.1	57.0	71.6
Qwen3-30B	79.3	92.5	85.4	79.9	96.3	87.3
Llama-3.1-70B	85.6	95.8	90.4	85.9	97.5	91.3
Gemini-2.5	92.4	95.4	93.9	92.5	97.0	94.7

Table 8: F1 score of positive metonymic sentences under two conditions: using metonymy only sentences as positive sentences, and using hybrid sentences as positive sentences. Literal sentences are the negative sentences.

	$sim(\text{N}_{lit},\text{N}_{mty})$	$sim(\text{N}_{lit},\text{N}_{hyb})$
GPT-OSS-20B	71.00	72.38 (+1.38)
Qwen3-30B	90.88	91.75 (+0.87)
Llama-3.1-70B	57.78	59.64 (+1.86)
BERT	65.42	65.62 (+0.20)

Table 9: Similarity score between the contextual embeddings of the noun tokens.

sim(\text{N}_{lit},\text{N}_{mty})

= similarity between the noun in the literal sentence and metonymic sentence.

sim(\text{N}_{lit},\text{N}_{hyb})

= similarity between the noun in the literal sentence and hybrid sentence.

Type	Original Sentence	Generated Sentence
Metonymy	(1) His guitarist realizes what he did and knocks him out.	His chord realizes what he did and knocks him out.
	(2) The reporters questioned, “why would a cricketer do this?”	The reporters questioned, “why would a bat do this?”
	(3) His father guided him in his early year.	His wisdom guided him in his early years.
	(4) The player was furious at the referee.	The athlete was furious at the referee.
	(5) The teacher encouraged her to apply for the position.	The staff at the school encouraged her to apply for the position.
Metaphor	(6) A dancer should watch her diet carefully.	A dancer should obsess over her diet carefully.

Table 10: Error analysis of the generated sentences using our framework. Words highlighted in blue in the original sentence are the target words that are being altered. The red words are the cause of the errors in the generated sentence.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

	Metonymy	Metaphor	Hybrid
Llama-3.1-8B	74	86	73
GPT-oss-20B	74	87	74
Qwen3-30B	75	86	75
Llama-3.1-70B	75	87	74
GPT-5	72	90	72

Appendix

Table 11: Number of sentences evaluated by humans to have the intended figurative expression in a sample of 100 when ran on our framework with different LLMs.

	$sim(\text{V}_{lit},\text{V}_{mtr})$	$sim(\text{V}_{lit},\text{V}_{hyb})$
GPT-oss-20B	70.19	70.28 (+0.09)
Qwen3-30B	92.83	93.07 (+0.24)
Llama-3.1-70B	57.32	52.88 (-4.44)
BERT	65.24	65.02 (-0.22)

Appendix

Table 12: Similarity score between the contextual embeddings of the verb tokens.

sim(\text{V}_{lit},\text{V}_{mtr})

= similarity between the verb and in the literal sentence and metaphoric sentence.

sim(\text{N}_{lit},\text{N}_{hyb})

= similarity between the verb and in the literal sentence and hybrid sentence.

Model	Similarity		Entailment
Model	Mtr	Hyb	Mtr	Hyb
GPT-OSS-20B	86.45	84.82	96.03	93.84
Qwen3-30B	84.79	83.34	94.54	94.31
Llama-3.1-70B	82.02	78.59	85.94	79.40
Gemini-2.5	86.31	82.13	93.24	90.96

Appendix

Table 13: Semantic similarity and entailment scores of metaphor (Mtr) and hybrid (Hyb) sentences with respect to their literal counterparts.

Panel of Figure 2 — Figure 2: Figure showing results of the downstream experiment. Blue bar shows the performance when trained on an existing metonymy dataset. Green bar shows the performance when trained on existing dataset + MetFuse metonymic samples, orange bar is existing datatset + MetFuse hybrid samples.

Surprisal	Lit	Mty	Mtr	Hyb
Noun token	9.01	12.79	9.65	12.81
Verb token	7.03	9.02	11.38	12.66

Appendix

Table 14: Surprisal scores for the noun and verb token in literal (lit), metonymy (Mty), metaphor (Mtr) and hybrid (Hyb) sentences. Bold indicates the token was altered to create the intended figurative expression.

Figurative Language Understanding

Figurative language understanding studies how people and AI interpret nonliteral meaning in metaphors, idioms, proverbs, puns, and related forms.

See all 43 papers

M $^3$ R-Bench: A Unified Benchmark for Evidence-Grounded Multimodal Metaphor Understanding
Aug 6, 2026Hong Jiang, Junnan Zhu, Jingwang Huang +9Figurative Language Understanding Multimodal Reasoning
School of Computer Science, Chongqing University, Chongqing, China · Institute of Automation, Chinese Academy of Sciences, Beijing, China · Institute of Advanced Intelligence and Computing (IAIC), Agency for Science, Technology and Research (A*STAR), Singapore +3
Interpretable Chinese Metaphor Identification via LLM-Assisted MIPVU Rule Script Generation: A Comparative Protocol Study
Mar 11, 2026Weihang Huang, Mengna LiuFigurative Language Understanding Interpretable ML
University of Birmingham Birmingham, UK · Guangdong University of Foreign Studies Guangzhou, China
Towards End-to-End Multilingual Metaphor Processing: Integrating Detection, Translation, and Evaluation
Aug 4, 2026Jiahui Liang, Lifeng HanFigurative Language Understanding Machine Translation
Centre for Linguistics, Humanities, Leiden University, NL · LIACS, Leiden University, NL · BDS, Leiden University Medical Centre, NL

Aug 6, 2026J/K move · Enter open · S save

M $^3$ R-Bench: A Unified Benchmark for Evidence-Grounded Multimodal Metaphor Understanding

Hong Jiang, Junnan Zhu, Jingwang Huang, Xiao Sun+8

School of Computer Science, Chongqing University, Chongqing, China · Institute of Automation, Chinese Academy of Sciences, Beijing, China · Institute of Advanced Intelligence and Computing (IAIC), Agency for Science, Technology and Research (A*STAR), Singapore+3

Figurative Language Understanding Multimodal Reasoning Multimodal Model Evaluation Multimodal Understanding

Open paper AskPDF arXiv

Multimodal metaphor understanding requires identifying cross-domain Target-Source mappings and inferring sentiment from visual and textual evidence. However, existing benchmarks largely evaluate these subtasks separately and lack evidence-grounded explanations, limiting their ability to assess coherent cross-modal reasoning. To enable unified and evidence-grounded evaluation, we introduce M

^3

R-Bench, a benchmark of 1,000 human-annotated image-text instances with joint labels for metaphor occurrence, Target-Source mapping, and sentiment, alongside stage-wise explanations for evidence identification, mapping establishment, and sentiment inference. Evaluations reveal that existing models frequently rely on textual shortcuts and produce Target-Source mappings inconsistent with visual evidence. To address this evidence-mapping mismatch, we propose M

^3

R-Reasoner, which combines curriculum-based reasoning supervision with task-aware reinforcement learning. Built on an 8B-parameter backbone, M

^3

R-Reasoner achieves 37.90% Joint Label Accuracy (JLA) metric, outperforming the strongest proprietary baseline by 19.50 percentage points. It further surpasses GPT-5.5 in image-text evidence and sentiment justification scores by 28.45 and 28.11 points, respectively. The dataset and code are available at https://github.com/hongshi4/m3r-bench-artifact.

Abstract

Figurative Language Understanding

Explore similar work

M3^33R-Bench: A Unified Benchmark for Evidence-Grounded Multimodal Metaphor Understanding

Interpretable Chinese Metaphor Identification via LLM-Assisted MIPVU Rule Script Generation: A Comparative Protocol Study

Towards End-to-End Multilingual Metaphor Processing: Integrating Detection, Translation, and Evaluation

M $^3$ R-Bench: A Unified Benchmark for Evidence-Grounded Multimodal Metaphor Understanding