AI Maps the Emotional Fingerprints of Metaphors, Sarcasm and Similes

When someone says that time is a thief, or delivers a barbed compliment dripping with sarcasm, they are doing far more than decorating their speech. Figurative language is one of the most powerful tools humans possess for expressing complex emotions and abstract thoughts, yet the precise emotional signatures carried by different figurative forms have remained surprisingly poorly defined. A new study published in the journal Cognitive Computation by Kun Sun and Yun Wu of Tongji University and Rong Wang of the University of Tübingen set out to change that, using an arsenal of BERT-based emotion classification models to systematically map how metaphors, similes, idioms, sarcasm and irony relate to emotional content across thousands of texts.

The researchers confronted a persistent methodological gap in the literature. Previous studies of figurative language and emotion have often relied on participants’ self-reports, which are subject to individual variability and limited generalizability. Many focused on conventional rather than novel metaphors, or used a restricted set of short lexical stimuli such as single words like blue or up. Cross-cultural differences in how expressions such as feeling blue are interpreted further complicate both psychological assessment and computational modeling. Neuroscientific evidence does suggest that processing figurative expressions activates brain regions involved in both semantic integration and emotional regulation, but large-scale, objective mapping of the emotion-figurativeness relationship has been lacking.

To fill this gap, the team applied four Transformer-based emotion classifiers to two human-annotated datasets: figurative-nli, containing 6,250 unique texts drawn partly from Reddit and Twitter, and FLUTE, containing 7,530 texts generated with GPT-3 and verified by expert annotators. The models spanned a wide range of emotional granularity. The Cardiffnlp model, fine-tuned on Twitter data, scored texts on just three sentiment dimensions: negative, neutral and positive. The AnkitAI model, built on DeBERTa V3, distinguished six emotions including anger, fear, sadness, joy, love and surprise. The Hartmann model, based on DistilRoBERTa, added disgust and neutral to produce seven categories, while the SamLowe model, trained on the GoEmotions corpus of 58,000 Reddit comments, delivered fine-grained scores across 28 emotion dimensions, from grief and approval to admiration, amusement and curiosity.

Each model processed every sentence in both datasets, producing probability scores between zero and one for each emotional dimension. The researchers then used multinomial logistic regression, multinomial LASSO regression with L1 regularization, and random forest models to test whether emotion scores could predict figurative language type. This multi-method design was deliberate: LASSO excels with noisy, high-dimensional data because it shrinks redundant predictors to zero, while random forests capture nonlinear relationships and served as a nonlinear cross-validation of the LASSO findings. Concordant results across methods strengthen confidence in the identified emotion-language associations.

The results revealed both striking regularity and instructive variation. Sarcasm emerged as the most emotionally stable figurative type, showing a consistent association with negative affect across both datasets and all models. In the FLUTE dataset, anger was a strong positive predictor of sarcasm, with LASSO coefficients reaching 3.221 for the AnkitAI model, while joy acted as a deterrent with a coefficient of -1.910 for the SamLowe model. In the ablation study using large language models, Claude Opus 4.5, GPT-5.2 and Gemini 3.1 Pro all classified 80 to 95 percent of sarcastic utterances as negative, with elevated anger scores between 0.39 and 0.52. Sarcasm, it seems, rarely hides behind emotional neutrality.

Metaphors told a more context-dependent story. In the figurative-nli dataset, metaphors were consistently associated with intense negative emotions such as fear and grief, alongside neutral sentiment, and avoided positive or dynamic emotions like joy and surprise. Fear strongly predicted metaphor with a regression coefficient of 2.36. Yet in FLUTE, metaphors flipped toward positive emotions including love, joy, approval and optimism while steering clear of anger and sadness. The researchers interpret this polarity shift as evidence that metaphor is highly sensitive to context and communicative goals rather than inconsistent: social media metaphors often appear in emotionally intense or evaluative contexts, whereas FLUTE’s curated examples favor illustrative, affectively positive expressions. This emotional moderation aligns with conceptual metaphor theory, which holds that metaphors map abstract domains onto concrete ones, and the authors suggest that abstraction enables emotional distancing, allowing speakers to articulate difficult experiences in a cognitively manageable form.

Similes proved to be the most corpus-sensitive of all. In FLUTE they were strongly associated with positive, vivid emotions such as joy, love, surprise and approval, but in figurative-nli they skewed negative, driven by fear, grief and negative sentiment while occasionally co-occurring with admiration. The authors attribute this contrast to differences in dataset design: figurative-nli is built for natural language inference tasks and draws from evaluative social media discourse, while FLUTE prioritizes emotional clarity and illustrative comparison. Rather than reflecting a fixed emotional orientation, similes appear to serve as flexible instruments capable of highlighting emotional intensity whether in praise, critique or empathy. Their explicit comparative structure, unlike metaphor’s implicit mapping, fosters stronger emotional engagement while maintaining conceptual clarity.

Idioms and irony added further nuance. Idioms, which appear only in FLUTE, consistently aligned with positive and social emotions such as joy, surprise, approval and curiosity, while avoiding sadness, anger and grief, suggesting their role in lively, expressive communication. Irony, present only in figurative-nli, displayed the most complex emotional alignment of all, tied to anger and mixed sentiments while conspicuously avoiding fear and overt positivity. The authors link this ambivalence to appraisal theory, which frames emotions as responses to cognitive evaluations of events: sarcasm and irony often reflect conflicting appraisals, such as simultaneous disapproval and amusement, or tension between critique and social bonding. Irony’s cognitive demands, including perspective shifting and dual-layer interpretation, make it one of the most demanding figurative types to produce and comprehend.

The theoretical stakes extend beyond linguistics. The findings support usage-based theories of meaning, indicating that affective patterns reflect distributional properties of figurative usage in naturalistic corpora. They also carry practical weight for emotion-aware natural language processing and human-computer interaction. Recognizing the emotional fingerprints of different figurative forms could help sentiment analysis systems move beyond surface-level interpretation toward deeper pragmatic intentions, and could inform the design of conversational agents, with metaphors supporting reflective dialogue in mental health or educational contexts and similes and idioms enhancing the expressiveness of virtual assistants. The authors caution, however, that deploying sarcasm in sensitive settings such as mental health chatbots risks misinterpretation and harm, and that emotion-laden figurative language in automated systems raises ethical concerns around consent, cultural variation, bias and affective manipulation.

The study is candid about its limitations. The emotion classifiers were trained primarily on English-language data, potentially introducing cultural and linguistic biases that limit generalizability to non-Western figurative expressions. The classifiers may also struggle with forms like sarcasm and irony that rely on pragmatic cues beyond literal meaning, and the scope was restricted to five figurative types, leaving allegory, hyperbole and personification for future work. The authors emphasize that some findings, particularly the strong association between sarcasm and grief, require further validation, and that the dataset-specific patterns observed for metaphors and similes may partly reflect artifacts of corpus construction. Nevertheless, the convergence between fine-tuned BERT classifiers and zero-shot large language models strengthens confidence in the core conclusion: figurative language types are not interchangeable, and some, like sarcasm, carry more stable emotional signatures than others. As emotionally aware language technologies mature, understanding these affective profiles will be essential for building systems that grasp not just what we say, but how we feel when we say it figuratively.

Subject of Research: Computational analysis of emotional associations in figurative language types using BERT-based emotion classification models

Article Title: Affective Profiles of Figurative Language: Cognitive and Computational Insights

Article References: Sun, K., Wang, R., & Wu, Y. (2026). Affective Profiles of Figurative Language: Cognitive and Computational Insights. Cognitive Computation, 18(1), Article 115. https://doi.org/10.1007/s12559-026-10648-w

Image Credits: AI Generated

DOI: 10.1007/s12559-026-10648-w

Keywords: figurative language, emotion classification, BERT, sarcasm, metaphor, simile, idiom, irony, affective computing, natural language processing, conceptual metaphor theory, sentiment analysis