Metaphors on the Move: AI Tracks How 19th-Century Poetry Reads in the Internet Age

The metaphor “capelli di fiamma” — hair of flame — was coined in an Italy where written language was still the preserve of an educated elite. More than a century later, most readers encounter written Italian not in printed verse but on screens, in blogs and web articles. Does that shift change how hard a literary metaphor is to understand? A team of psycholinguists at the University School for Advanced Studies IUSS in Pavia set out to answer precisely that question, and their findings reveal a surprising story about how genre, not time itself, reshapes the cognitive cost of figurative language.

Literary metaphors occupy a peculiar place in the science of language. Decades of experimental work have examined how we process everyday metaphors such as “that lawyer is a shark,” but the metaphors of poetry and fiction have received far less empirical attention. Theoretically, they are special. Relevance Theory holds that while everyday metaphors point toward a fairly direct interpretation, literary metaphors generate a wide range of so-called weak implicatures — Flaubert’s “his ink is pale” suggests writing that lacks contrast, may fade, will not last. That proliferation of meanings is thought to demand extra cognitive effort, producing what scholars call the poetic effect. Corpus studies back the intuition: metaphors make up roughly 11 percent of words in fiction, and they are more often novel, non-lexicalized pairings than metaphors elsewhere.

What had never been examined, the Pavia team led by Veronica Mangiaterra, Chiara Barattieri di San Pietro, Paolo Canal and Valentina Bambini noticed, is the temporal dimension. Many literary metaphors used in laboratory studies were coined by authors who lived long before the participants reading them. Whether that gap between creation and comprehension contributes to processing difficulty — and whether readers today face a heavier burden than readers of the author’s own era — had never been systematically tested. A quarter-century replication of the classic 1988 metaphor norms of Katz and colleagues showed ratings held steady over 25 years, but two centuries is a very different span, one over which Italian itself changed dramatically, shifting from an illustrious, courtly register toward informal, oralized modern styles.

To capture meaning as it existed in different eras, the researchers turned to diachronic distributional semantics, a branch of computational linguistics built on the distributional hypothesis: words appearing in similar contexts tend to have similar meanings. Using the Word2Vec algorithm, they trained vector space models on period-specific corpora totaling 124 million tokens, so that every word becomes a vector whose geometry encodes its usage at a particular time. Crucially, semantic similarity between two words can be computed as the cosine of the angle between their vectors, and this measure has well-documented psycholinguistic validity — lower cosine similarity predicts longer reaction times and modulates electrophysiological brain responses in literal language processing, making it a defensible proxy for processing cost in metaphor research.

The metaphor dataset itself was built by mining 19th-century Italian texts from Project Gutenberg for genitive constructions of the form “A of B,” which clearly display a metaphor’s two parts: the topic, the subject being described, and the vehicle, the term used figuratively. After part-of-speech tagging and filtering by semantic classes commonly found in figurative expressions — natural events, emotions, and the like — roughly 95,000 candidate strings underwent manual annotation, validated with strong inter-annotator agreement (Cohen’s kappa of 0.78). Combined with an existing collection, the final dataset contained 515 literary metaphors such as “cielo di perla” (sky of pearl) and “nebbia di malinconia” (fog of melancholy), now openly available in the Figurative Archive.

The computational centerpiece was the TWEC model — Temporally aligned Word Embeddings with a Compass. Because vector models are stochastic, training runs produce different coordinate systems that cannot be directly compared. TWEC solves this by first training an atemporal “compass” on the entire corpus and then using its context matrix to initialize time- and genre-specific models, anchoring every representation to the same spatial frame. The researchers built four such embeddings: 19th-century literary and nonliterary (drawn from Gutenberg literature, technical manuals, newspapers, diaries, and a corpus of Italian immigrant newspapers in the United States) and 21st-century literary and nonliterary (contemporary bestseller and award-listed literature versus a web-crawled corpus of news sites, blogs, and educational pages).

When the team modeled topic-vehicle semantic similarity across these spaces, the headline result was one of stability: overall, the semantic relationship between metaphor terms barely changed from the 1800s to today, suggesting no blanket increase in processing cost as metaphors aged. But the interaction between epoch and genre told a far more interesting story, and it was statistically robust. Metaphor terms grew farther apart — harder to connect — in 21st-century literary texts compared with 19th-century literature, while they grew closer together — easier to connect — in today’s nonliterary, web-based language compared with 19th-century nonliterary prose. In the 19th century, when both literary and nonliterary texts were elite products sharing a high, classicizing register, metaphors were processed similarly regardless of genre. Modern Italian has split that unity apart.

The authors link this divergence to broader stylistic trajectories. Modern Italian literature has drifted toward orality and simplification, a plainer register in which novel metaphorical associations are rarer and thus more striking and effortful. The language of the web, by contrast, is notoriously variable, hybrid, and creatively experimental, routinely combining registers and coining neologisms. In that environment, unexpected associations between distant concepts have become entrenched enough to be processed with reduced effort. The pattern appears language-specific: a companion proof-of-concept study applying the same method to English metaphors found only a genre effect, with no epoch change — consistent with English having reached its modern form earlier and with more stable stylistic conventions.

Word-level features added a second layer of nuance. Semantic neighborhood density — how many close neighbors a word has in the vector space — and vector coherence — the stability of a word’s meaning over time, computed as the cosine similarity between its 19th-century and 21st-century vectors — both shaped metaphor trajectories. In literary texts, both topics and vehicles showed high coherence, meaning their semantics barely shifted, while nonliterary words drifted substantially. In the statistical models, the vehicle’s neighborhood density predicted metaphor similarity most strongly in 21st-century nonliterary texts: dense vehicles like “flour” in “sky of flour” anchor entrenched, easily processed associations, whereas sparser vehicles like “inebriation” in “inebriation of light” allow the loose, open-ended readings prized in literature but costlier to compute. Meanwhile, in contemporary literary texts, processing demands tracked the coherence of the topic: a topic whose meaning has shifted — “lode” (praise), which once evoked gratitude and now evokes report cards and graduating cum laude — offers a weaker guide for working out what the vehicle is doing.

The authors are careful about what these measures do and do not show. Cosine similarity is a distributional proxy, not a direct record of a reader’s mental effort, and the study’s binary literary/nonliterary split conflates register, medium, and audience — the modern nonliterary corpus is, after all, web text, a fundamentally different communicative setting from 19th-century manuals. Future work combining temporal embeddings with behavioral and neurophysiological measures of metaphor comprehension, or probing the imagistic and sensorimotor dimensions that text-based models cannot capture, would help close those gaps. Yet the conceptual payoff stands: what makes a metaphor difficult is not a fixed property of the expression itself but a dynamic outcome of diachronic, genre-driven shifts in the language around it. The same “hair of flame” that burned brightly for a 19th-century reader now flickers differently in a novel than on a webpage — and the difference can be measured, one vector at a time.

Subject of Research: Diachronic computational modeling of processing demands for 19th-century Italian literary metaphors using temporal word embeddings

Article Title: Metaphors' journeys across time and genre: tracking the evolution of literary metaphors with temporal embeddings

Article References: Mangiaterra, V., Barattieri di San Pietro, C., Canal, P., & Bambini, V. (2026). Metaphors' journeys across time and genre: tracking the evolution of literary metaphors with temporal embeddings. Frontiers in Language Sciences, 5, Article 1828340. https://doi.org/10.3389/flang.2026.1828340

Image Credits: AI Generated

DOI: 10.3389/flang.2026.1828340

Keywords: metaphor, literary language, word embeddings, diachronic semantics, psycholinguistics, Italian language, corpus linguistics, semantic similarity, Web2Vec, TWEC, semantic change, web language