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

1 Introduction

Research on metaphors, both from the perspective of cognitive linguistics (Lakoff and Johnson, 1980) and pragmatics (Sperber and Wilson, 2012), has largely contributed to depicting metaphors as a pervasive phenomenon in language and cognition, rather than just a tool for poets. Similarly, empirical research as pursued within Experimental Pragmatics, psycholinguistics, and neurolinguistics has focused on everyday manifestations of metaphors. Nonetheless, metaphor remains a stylistic hallmark of poetic and narrative language (Steen et al., 2010). Understanding the mechanisms through which we comprehend metaphors in literature captures, therefore, a relevant aspect of metaphor processing. To date, however, experimental research on literary metaphors remains a niche topic.

Theoretically, literary metaphors hold a special status compared to everyday metaphors. According to Relevance Theory, while everyday metaphors point toward a straightforward interpretation (e.g., ‘That lawyer is a shark', meaning ‘The lawyer is aggressive'), literary metaphors are held to generate a wide range of weak implicatures (e.g., the metaphor ‘His ink is pale' by Gustave Flaubert, meaning that ‘His writing lacks contrast, may fade, will not last'). Processing this wide range of assumptions is supposed to require an additional cognitive effort that generates the so-called poetic effect of literary metaphors (Pilkington, 2000). On the quantitative side, corpus-linguistics studies found that metaphors are well represented in fiction, where they constitute 11% of the words (Steen et al., 2010), and that metaphors in literary texts are, more frequently than in other genres, non-lexicalized, i.e., with a novel association between topic and vehicle (Goatly, 1997).

From the experimental point of view, the few available studies on literary metaphors led to heterogeneous results. Katz et al. (1988) collected ratings from human participants for literary and everyday metaphors, finding no difference along ten psycholinguistic dimensions and suggesting that they do not differ in nature. Conversely, other studies reported that literary metaphors are less familiar, more open-ended, and more difficult to understand than journalistic metaphors (Semino and Steen, 2008; Steen, 1994). In addition, Bambini et al. (2014) found that ratings are modulated by context, and specifically that when literary metaphors are presented together with their context, people tend to consider them more meaningful and less concrete, less difficult, and less familiar, compared to when presented without context. Consistently, electrophysiological evidence showed that literary metaphors trigger a sustained brain response, possibly linked to the manipulation of multiple meanings (Bambini et al., 2019), supporting the theoretical claims that literary metaphors evoke multiple weak implicatures and require greater cognitive effort compared to non-literary metaphors.

A less considered factor in the study of literary metaphors is the temporal dimension: literary metaphors used in empirical studies were created by authors living well before the participants who took part in the experiments investigating the processing of these expressions. Whether this temporal gap between the time of production of literary metaphors and the time of their processing contributes to the cognitive effort involved in interpreting metaphors, and whether readers today need an additional effort compared to readers contemporary to the time of metaphor production, has never been investigated. A preliminary answer may come from the replication of the Katz et al. (1988) study by Campbell and Raney (2016). They recollected ratings for the same metaphors used in Katz et al. 25 years later and found that judgments were consistent over time. However, 25 years is a limited time span, while the time in which metaphors were created and contemporary readers might be centuries apart.

The issue of the temporal gap becomes particularly relevant when we consider that both language and society have deeply changed over the last two centuries, potentially modifying the conceptual and contextual factors underlying metaphor interpretation. Focusing on Italian, studies on the history of language noted that in the 19th century, written language was still an élite product. Despite some innovative thrusts, Italian was characterized by a style close to the courtly and illustrious tradition both in literature (Beccaria, 1993; Serianni, 1993) and in journalistic and nonfiction language (Marazzini, 2002; Masini, 1977). In the 21st century, instead, language has evolved, particularly in literary contexts, toward a strong shift to oral and informal styles (Coletti, 2022). Moreover, in the 21st century, much of the exposure to written language occurs on the web, as highlighted, for instance, by recent reports on news consumption that indicate that web-based sources are used by 68% of the Italian sample, compared to 12% for print media (Newman et al., 2025). The language of the web, or netspeak (Crystal, 2006), has its own specificities: it is based on informal speech, combining terms from different registers and neologisms (Cerruti and Onesti, 2013), and allows users to experiment with new forms of language creatively (Fiorentino, 2018; Goddard, 2015).

1.1 A distributional approach to metaphor evolution

To understand how literary metaphors were perceived at the time of their creation, distributional semantics techniques can be of help, as they allow us to construct representations of meaning mimicking the semantic networks of readers of different epochs and to compare the status of the metaphors in the two representations. Two key features of these techniques make them suitable for answering our research question: i) their ability to approximate human representation of meaning and ii) the possibility to build models starting from historical corpora and therefore represent meaning at specific time points. Distributional semantics is based on the theoretical intuition, the so-called distributional hypothesis (Harris, 1954), that states that words occurring in similar contexts have similar meanings. Implementing this hypothesis, Vector Space Models (VSMs) can be trained on large text corpora to learn co-occurrences of words. In VSMs, words are represented as vectors, whose coordinates are derived from their co-occurrences in the corpora. Words whose vectors are closer in the VSM tend to be closer in meaning. In these models, the semantic similarity between two words can be operationalized as the cosine of the angle between the vector of word₁ (v_*w*1) and the vector of word₂ (v_*w*2), as defined by the following formula:

Cosine Similarity=cos(vw1, vw2)= vw1. vw2||vw1|| . ||vw2||

Since their inception, distributional semantic approaches have been closely tied to psychological and psycholinguistic research, given that VSMs can model various dimensions of human semantic knowledge, capturing how word meaning is acquired, processed, and stored in the brain (Bhatia et al., 2019; Günther et al., 2016; Jones et al., 2015). Crucially, cosine similarity has demonstrated psycholinguistic validity for predicting human responses: in literal semantic processing, lower cosine similarity between words is associated with longer reaction times (Günther et al., 2016) and modulated electrophysiological responses (Dudschig et al., 2025; Frank and Willems, 2017), supporting its use as a proxy for processing cost. Distributional semantics tools can be a valid application to quantify also aspects of figurative language processing (Reid and Katz, 2018) and interpretation (Coulson and Matlock, 2001; Kintsch, 2000). In these applications, a widely studied measure is the semantic similarity between the two terms of a metaphor. Well before the diffusion of computational tools, the relation between topic, namely the subject of the metaphor (e.g., “lawyer” in ‘That lawyer is a shark'), and vehicle, namely the term used metaphorically (e.g., “shark”), was recognized as crucial to metaphor comprehension and appreciation, with rating-based semantic similarity being linked to a number of dimensions, such as metaphoricity, ease of interpretation, and goodness of nonliterary and literary metaphors (Katz et al., 1985; Marschark et al., 1983). The computational operationalization of semantic similarity via word embeddings has been largely applied to metaphor research (Bolognesi and Aina, 2019; Brglez and Vintar, 2025; Utsumi, 2011). Applications ranged from automatic metaphor identification (based on defining a cosine similarity threshold below which expressions are considered metaphorical, see Mao et al., 2018; Shutova, 2015; Su et al., 2017) to modeling metaphor processing costs (where cosine similarity is considered an approximation of human judgments). For instance, McGregor et al. (2019) found that contextual embeddings, i.e., dynamic representations of words based on surrounding contexts, can efficiently model human ratings of metaphoricity, meaningfulness, and familiarity, and Winter and Strik-Lievers (2023) showed that semantic similarity between the topic and the vehicle of synesthetic metaphors mirrors their degree of metaphoricity and creativity. Converging evidence comes from the Figurative Archive (Bressler et al., 2026), a recent resource collecting approximately 1,000 metaphors together with human ratings, where significant correlations between semantic distance and both familiarity and metaphoricity were reported. In particular, metaphors with less semantically similar topics and vehicles are considered less familiar and more difficult to process. Moreover, lower semantic relatedness as operationalized through word embeddings predicts slower processing times in metaphor comprehension tasks (Westbury and Harati, 2022), and distributional models using cosine-based measures consistently approximate human metaphor interpretation (Al-Azary et al., 2025). While we recognize that computational similarity scores and psycholinguistic processing costs are not equivalent and that the former is an indirect approximation of the latter, the converging evidence from both behavioral and neural measures supports the use of cosine similarity as a proxy for processing demands in the present study.

From word embeddings, other measures relevant for language and metaphor processing can be derived. Among those, semantic neighborhood density (SND), i.e., the number of words that are similar to the target one in terms of meaning, was shown to play a role in language processing, being able to predict performances on psycholinguistic tasks, such as lexical decision and word naming (Buchanan et al., 2001). In metaphor research, metaphors composed of low-SND words were considered more comprehensible than metaphors with high-SND words (Al-Azary and Buchanan, 2017). Moreover, the semantic neighborhood density of the vehicle allows to discriminate between different types of metaphors: literary metaphors have vehicles with higher SND compared to nonliterary metaphors (Reid et al., 2023).

Semantic similarity has found applications in historical linguistics and semantic change research. The distributional hypothesis can indeed be adapted to the diachronic perspective: changes in a word's co-occurrences reflect changes in its meaning (Hilpert, 2008). Hence, by looking at how word co-occurrences change through time, it is possible to track the evolution of meaning in time. Operationally, to create time-characterized word embeddings, i.e., embeddings representing the word's meaning at each specific time point, it is necessary to train VSMs on corpora of different epochs. However, given the stochastic nature of VSMs, every time a training process is initialized, different spatial coordinates are used, making it impossible to compare word vectors across spaces. For the VSMs to be comparable, they need to have the same spatial coordinates, i.e., to be aligned. Different procedures (Gulordava and Baroni, 2011; Kulkarni et al., 2015) have been proposed to train aligned VSMs, among which the Temporally aligned Word Embeddings with a Compass (TWEC) model (Di Carlo et al., 2019).

By employing aligned time-locked semantic representations of a word, it is possible to track its semantic shift over time by computing the cosine similarity between the embedding at time₁ (v_*wt*1) and at time₂ (v_*wt*2), a measure called vector coherence (Cassani et al., 2021; Hamilton et al., 2016). As a result, a word whose meaning has shifted over time would exhibit a lower semantic similarity between v_*wt*1 and v_*wt*2 (i.e., lower vector coherence) compared to a word whose meaning has remained consistent, which would exhibit a higher semantic similarity between v_*wt*1 and v_*wt*2 (higher vector coherence). To date, most studies in diachronic distributional semantics have focused on identifying semantic changes of single words (Cain and Ryskin, 2025; Charlesworth et al., 2022; Garg et al., 2018; Hamilton et al., 2016; Xu and Kemp, 2015), by examining variations in vector coherence across epochs. Only recently these approaches have been applied to investigate how the relation between different words changes over time (Jenkins et al., 2025), a perspective that is crucial for tracing the evolution of complex expressions, such as metaphors.

1.2 The present study

In this study, we aimed to explore whether Italian literary metaphors are associated with different processing demands for today's readers compared to 19th-century readers, contemporary to the time of metaphors' original creation. To do so, we innovatively extended diachronic VSMs from word-level to the case of multi-word expressions, namely metaphors. Specifically, we trained VSMs on historical corpora from the 19th century and on contemporary corpora, mimicking the linguistic input available to present and past readers. Given that (i) the evolution of Italian is closely intertwined with textual genres, and (ii) different interpretative attitudes and processing modes may be activated depending on the literariness of a text (Steen, 1989), we incorporated the genre dimension by training separate models for literary and nonliterary corpora within each epoch. Then, we compared the semantic similarity between topics and vehicles of a set of literary metaphors in each diachronic VSM, taking this measure as a proxy of the evolution of metaphors' processing demands. Moreover, we examined whether metaphors' evolution is further shaped by lexical-semantic features of individual topics and vehicles, such as their stability over time (vector coherence), their semantic neighborhood density, and their frequency.

We expected that the temporal dimension, when assessed over a sufficiently long span of time, would significantly influence the processing costs of metaphors. In particular, we hypothesized that metaphors would yield higher elaboration demands in present-day readers than in readers of the past, who shared the conceptual and contextual environment in which the metaphors were originally created. Also, given that in the past both literary and nonliterary texts were intended for educated readerships and exhibited similar stylistic features (Aprile, 2014), we expected no difference across genres in the past. In the 21st century, instead, literary language and web-based language constitute distinct varieties of Italian, which led us to expect differences between the two genres in the processing demands of metaphors.

2 Methods

2.1 Metaphor dataset

To extract literary metaphors, a set of 19th-century Italian literary texts was retrieved from Project Gutenberg (https://www.gutenberg.org/), an initiative started in the 1970s, intending to collect digital versions of books that have never been copyrighted or whose copyright has lapsed. Defining what constitutes a literary metaphor is not an easy task: for the purposes of this work, we followed previous studies and defined a metaphor as “literary” if it was retrieved from poetic or literary text (Jacobs and Kinder, 2017; Katz et al., 1988). From the selected texts, “A di B” (Eng. “A of B”) strings were extracted. Genitive metaphors were chosen because they frequently occur in literary text (Bambini et al., 2014), and they clearly display the two terms (topic and vehicle) of the metaphors. Moreover, this construction is suitable for automatic extraction, contrary to “A is B” metaphors, which in ecological literary texts usually include intervening terms (such as, for example, “A is adj B” as in “The night was a vast dark pupil”), which makes the automatic extraction more complex. Using the spaCy package (Honnibal and Montani, 2017), we performed a Part-Of-Speech (POS) tagging, and all the “NOUN of NOUN” strings were selected. Following (Hanks, 2006), a set of keywords from semantic classes that are frequently employed in metaphorical expressions, such as natural events and locations (river, storm, rain) and emotions (anger, joy), was then employed to filter the “NOUN of NOUN” strings. From the resulting list, metaphors were identified following the definition proposed by Fuoli et al. (2022), namely: “A string of words that describes one entity in terms of another unrelated entity by means of comparison”. Their annotation procedure focuses on the phraseological unit rather than on individual words, unlike earlier approaches such as the Metaphor Identification Procedure (Pragglejaz Group, 2007). Specifically, each string was examined within its original context, its contextual meaning was assessed, and it was determined whether the expression could also convey a more basic, literal meaning. If so, the string was annotated as metaphorical. The initial corpus comprised approximately 300,000 strings. After the semantic class filtering, 95,000 candidate strings were retained for manual inspection. These candidates were screened by the primary annotator until a final dataset of 400 metaphorical strings was obtained.

To validate the annotation procedure, two annotators independently labeled a random sample of 1,000 strings drawn from the filtered corpus. Of these, the primary annotator classified 180 strings (18.0%) as metaphorical. Inter-annotator agreement, assessed using Cohen's κ, was 0.78, indicating strong agreement according to established conventions in metaphor identification studies (Bolognesi, 2017; Pérez Sobrino and Ford, 2023). The 180 metaphorical strings identified within the validation sample were retained as part of the final set of 400 metaphors. The dataset was further enriched by adding an existing collection of “A of B”¹ literary metaphors, also drawn from 19th-century literary texts (Bambini et al., 2014). The final dataset included a total of N = 515 metaphors in the form of “A of B” (e.g., “capelli di fiamma”, Eng. “hair of flame”) and is fully available in the Literary Metaphors module of the Figurative Archive at https://neplab.shinyapps.io/FigurativeArchive/ (Bressler et al., 2026). As shown by a previous automatic analysis based on LLM-driven topic modeling reported in Bressler et al. (2026), the datasets cover a broad range of semantic domains, including human body, natural elements, materials, emotions, spiritual and abstract concepts, politics and war. These domains align with the majority of the common source domains, such as, for instance, body (e.g., “tongues of fog”), health and illness (e.g., “indigestion of praise”), animals (e.g., “dewlap of clouds”), plants (e.g., “eyes of gentian”), light and darkness (e.g., “light of poetry”), heat (e.g., “flame of faith”), tools (e.g., “umbrellas of vegetables”) and target domains, such as, for instance, emotions (e.g., “river of joy”), desire (e.g., “flaps of desire”), morality (e.g., “scent of innocence”), thoughts (e.g., “thoughts of smoke”), society (e.g., “bloom of men”), politics (e.g., “tangle of corruption”), religion (e.g., “seeds of religion”), communication (e.g., “battles of ideas”), as identified by Kövecses (2010). In terms of topic-vehicle order in the metaphorical expressions, some metaphors displayed a topic-vehicle (TV) order, with the first term being the topic and the second one being the vehicle, e.g., “capelli di fiamma” (Eng. “hair of flame”), and others displayed a vehicle-topic (VT) order, with the first term being the vehicle and the second one being the topic, e.g., “grumo di nuvole” (Eng. “clump of clouds”). Table 1 reports four examples of the metaphors included in the dataset, together with metadata regarding author, source, year of publication, topic, vehicle, and their order.

Table 1. Examples of literary metaphors included in the final dataset of the study.

T, Topic; V, Vehicle.

2.2 Training sets

To examine the diachronic evolution of metaphors, considering also the possible effect of textual genre, we collected a set of Italian corpora, for a total of 124 million tokens, composed of literary and nonliterary texts written in the 19th and the 21st century. Literary texts are defined not by fixed features but by dynamic, culturally situated prototypes encompassing content, form, function, and evaluative norms (de Geest and van Gorp, 1999; Steen, 1999). As a general criterion, following Meyer (1997), we operationalize literary texts as works belonging to established literary genres, i.e., poetry, prose, fiction, and drama, selecting for the 19th-century corpus, texts marked as such from Project Gutenberg and for the 21st-century corpus, privileging authors recognized through bestseller and award lists as culturally validated exemplars of the category. Table 2 reports a summary of the characteristics of the collected corpora.

Table 2. The composition of the literary and nonliterary sections of 19th-century and 21st-century corpora.

Size is in millions of tokens.

From the corpora, four training sets were built: 19th-century literary training set, 19th-century nonliterary training set, 21st-century literary training set, 21st-century nonliterary training set. The training sets included from 20 to 42 million tokens (19th century literary: 30 M; 19th century nonliterary: 32 M; 21st century literary 20 M; 21st century nonliterary: 42 M), in line with the mean size of large corpora employed by Di Carlo et al. (2019). The internal diversity of these datasets was designed to provide a proxy for the linguistic input accessible to the readers in each respective epoch. In the 19th-century training set, we included prose and poetry texts for the literary section and texts from technical manuals, newspapers, and diaries for the non-literary section. In the 21st-century training set, we included prose and poetry for the literary section and a collection of texts taken from the web, such as newspaper sites, blogs, and educational sites, for the non-literary section.

2.3 Training aligned spaces

Employing the Temporally aligned Word Embeddings with a Compass (TWEC) model (Di Carlo et al., 2019), we trained four sets of word embeddings on the four training sets.

The TWEC model exploits the dual representation of words derived from a word2vec model (Mikolov et al., 2013) based on a Continuous Bag of Words (CBOW) architecture, a feed-forward neural network trained to predict a target word given its context, relying on the theoretical assumption that most words do not change over time, and that words with a shifted meaning will appear in the context of words that did not change. This theoretical assumption is reflected in the creation of an atemporal VSM, called compass, against which the spatial coordinates of all the other temporal VSMs are then subsequently defined. In other words, a compass model is first trained on the entire corpus, providing semantic representations of words independently of time. After that, the compass's context matrix is used to initialize the training of a time- and genre-specific target matrix on each corpus, which allows us to extract the temporal word embeddings. Operationally, we trained the model on the whole training set (resulting from the combination of all training sets), and we extracted the resulting two atemporal matrices, an input-weight context matrix (C) and an output-weight target matrix (U). The target matrix U was used as “a compass”, i.e., it served as a reference to initialize all the other VSMs along the same coordinate system. We hence trained a set of context matrices C_i on each slice of the training set (namely, 19th-century literary training set, 19th-century nonliterary training set, 21st-century literary training set, 21st-century nonliterary training set). As a result, the context embeddings can differ according to the co-occurrence frequencies that are specific to a given temporal period. Each of the four resulting sets of word embeddings provides a representation of word meaning that differs by epoch (19th and 21st centuries) and genre (literary and nonliterary).

2.4 Measures of diachronic change

To semantically characterize the metaphors across time, we computed four measures of interest using the obtained sets of word embeddings, one at the metaphor level (cosine similarity between topic and vehicle – CS) and three at the word level, for each term of the metaphor (semantic neighborhood density – SND, vector coherence – VC, and frequency – Freq, see Table 3). Each measure was computed, for each metaphor, in all time and genre slices, to obtain a semantic characterization of the metaphor and its terms in each epoch and each textual genre. A graphical representation of the approach, including corpora collection, VSM training, and feature computation, is depicted in Figure 1.

Table 3. Measures of metaphor change, with corresponding formula and interpretation.

The value of n was set to 500, as in previous computational approaches to metaphors (Kintsch, 2000).

2.5 Statistical analysis

To understand the relationship between word- and metaphor-level measures, we computed a set of Pearson correlations between cosine similarity, semantic neighborhood density, vector coherence and frequency, corrected for multiple comparisons.

To test the hypothesis that cosine similarity varied through time and across different genres, we fitted a Linear Mixed-Effects Model (LMM – Pinheiro and Bates, 2000) using the lme4 package (Bates et al., 2015), considering cosine similarity as a continuous dependent variable, and genre and epoch as interacting categorical predictors. Then, to explore whether the evolution of metaphors is further shaped by different word-level variables, we fitted a series of LMMs using the lme4 package again, considering cosine similarity as a continuous dependent variable, genre and epoch as interacting categorical predictors, and word-level variables as continuous predictors. Specifically, to define the final statistical model, we incrementally added the word-level variables (i.e., first topic and vehicle semantic neighborhood density, frequency, and vector coherence separately, then pairs of topic and vehicle features, and finally all the variables), we tested if the addition of variables contributed to explaining the data variance by comparing the models' goodness-of-fit in terms of Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), and Log-likelihood (Bozdogan, 1987; Neath and Cavanaugh, 2012) and we assessed multicollinearity by checking VIF scores. All models included a random intercept to account for the variability of individual metaphors. Alternative random-effect structures including epoch and genre as random slopes were tested but discarded as the variance to estimate random effects of each factor was expected to be negligible, and the model yielded convergence issues. In all models, the categorical variables were effect-coded. Model assumptions were assessed through visual inspection of residual diagnostic plots, including histograms and Q–Q plots of residuals and random effects, which appeared approximately normal. All analyses were performed in R (R Core Team, 2025).

3 Results

3.1 Descriptive statistics

Table 4 reports descriptive statistics of the measures of interest as computed in the four corpora and sets of embeddings. The distribution of each variable is displayed in Figure 2. While measures like cosine similarity, semantic neighborhood density, and frequency maintain relatively normal distributions with consistent peaks across both centuries, vector coherence exhibits a significant departure from normality. Specifically, in the literary genre, both topics and vehicles demonstrate high vector coherence (approaching 1.0), suggesting that words in literary texts did not report drastic changes in meaning. Conversely, the words in the nonliterary genres show a great semantic shift.

Table 4. Descriptive statistics of word-level and metaphor-level measures reported as mean (SD).

CS, cosine similarity, computed between the topic and the vehicle of the metaphor; SND, semantic neighborhood density, computed as average cosine similarity between the word and its 500 closest neighbors; VC, vector coherence, computed as cosine similarity between the word at t₁ and the word at t₂; Freq, frequency, computed as the logarithm of the relative frequency of the word.

3.2 Correlation analysis

Results of the correlation analyses between single-word and metaphor-level measures are reported in Figure 3.

Cosine similarity, semantic neighborhood density, and frequency showed overall strong positive correlations across genre and epoch, as highlighted by the black triangles on the diagonal. Vector coherence, however, reported low correlations, confirming the different patterns of semantic change between literary and nonliterary genres.

Regarding word-level variables, semantic neighborhood density and frequency of both topics and vehicles were negatively correlated in all slices (see the blue squares), indicating that less frequent words tend to be more semantically dense across time and genre. A positive correlation emerged between frequency and vector coherence (see the light blue rectangles), especially in literary corpora, indicating that more frequent words have a more stable meaning. Moreover, semantic neighborhood density positively correlated with vector coherence (see the purple rectangle), suggesting that denser words have a more stable meaning.

Regarding the relations between word-level variables and cosine similarity (see green rectangles), we found that the latter was positively correlated with topic vector coherence in literary corpora, suggesting that when the meaning of the topics tends to change over time, metaphors are characterized by more semantically distant terms. Moreover, in nonliterary corpora, we found a positive correlation between both topic and vehicle semantic neighborhood density and cosine similarity, indicating that when metaphors are constituted by high-density terms, they tend to have a greater topic-vehicle similarity.

3.3 Linear mixed-effects models

The simple LMM including epoch and genre as interacting predictors revealed an effect of genre (β = −0.03, t = −6.85, p < 0.001), with metaphors reporting a lower semantic similarity in literary corpora compared to nonliterary ones, further qualified by its interaction with epoch (β = −0.07, t = −7.82, p < 0.001). No main effect of epoch was reported (p = 0.35). A diverging trend emerged: metaphors' terms become increasingly distant in literary texts going from the 19th century to the 21st century, while in nonliterary texts, metaphors' terms showed an increasing semantic similarity (Table 6, Figure 4).

Moving to the investigation of how the diachronic evolution of metaphors is further shaped by lexical-semantic features of topic and vehicle, the comparison of AIC, BIC, and Log-likelihood showed that the best fit to the data, while avoiding the introduction of multicollinear factors (VIF value < 5), is obtained by adding both topic and vehicle semantic neighborhood density as well as topic vector coherence to the simple model (see Table 5).

Table 5. Summary statistics of LMMs presenting AIC, BIC, and Log-likelihood.

AIC, Akaike's information criterion; BIC, Bayesian information criterion; Freq, word frequency; LogLik, log-likelihood; SND, semantic neighborhood density; VC, vector coherence. Values in bold indicate the models with lower Akaike Information Criterion (AIC).

The LMM showed a significant three-way interaction between topic vector coherence, genre, and epoch (β = 0.227, t = 3.17, p = 0.002), suggesting that the coherence of the topic influenced metaphor cosine similarity in contemporary literary texts but not in the other slices (Table 6, Figure 5A). Moreover, the model showed a significant interaction between vehicle semantic neighborhood density, genre, and epoch (β = −0.629, t = −3.44, p < 0.001), in addition to the main effect of vehicle semantic neighborhood density (β = 0.61, t = 8.01, p < 0.001). While the main effect indicated that higher semantic neighborhood density generally predicts higher cosine similarity, this relationship became stronger specifically in 21st century nonliterary texts (Table 6, Figure 5B).

Table 6. Outputs of the base LMM model and the best LMM model with single-word features on cosine similarity.

SND, semantic neighborhood density; VC, vector coherence. Values in bold indicate a statistically significant predictor (p < 0.05).

4 Discussion

In the present study, we investigated whether the distributional features, namely topic-vehicle semantic similarity, of Italian literary metaphors changed over time and as a function of textual genres and lexical-semantic features of the topic and vehicle. To do so, we collected 19th-century and 21st-century corpora (composed of literary and nonliterary texts), taken to reflect the linguistic input of today's and past readers. Then, we used the corpora to train different temporal vector space models, obtaining representations of word meanings at the two time points. Then, we examined how the semantic similarity between topics and vehicles of a set of 515 Italian metaphors, extracted from literary texts of the 19th century, varied. Crucially, while our models capture purely distributional, text-based data, we relied on established psycholinguistic literature to interpret semantic similarity as a computational proxy for potential processing demands.

Importantly, we did not limit the analysis to semantic similarity across epochs and genres, but we also included three single-word variables relevant in historical and metaphor research. Specifically, we considered i) a measure of the stability of word meaning over time (vector coherence; Cassani et al., 2021; Hamilton et al., 2016; Rodda et al., 2017); ii) frequency, as it affects both semantic shift (Englhardt et al., 2020; Feltgen et al., 2017; Sagi et al., 2009) and metaphor processing (Littlemore et al., 2018); iii) semantic neighborhood density (SND), which was shown to impact metaphor processing, with high density hindering metaphor comprehensibility (Al-Azary and Buchanan, 2017; Reid et al., 2023). Given the significant stylistic change and simplification of Italian over the past two centuries (Coletti, 2022), we expected that today's readers may experience classical literary metaphor differently from the readers of the 19th century. In particular, we assumed that today's readers do not share the cultural background of the authors of classical literary metaphors, and, as such, we expected that metaphors could entail greater processing costs for contemporary readers.

Overall, our findings revealed a more nuanced pattern than originally anticipated. While the semantic relationship between topics and vehicles remained remarkably stable over time, this apparent stability concealed important differences across textual genres as well as systematic contributions of lexical-semantic properties.

4.1 Diachronic evolution of literary metaphors

The first main finding of the study is that, overall, the semantic relationship between topics and vehicles of metaphors did not change from the 19th century to today, suggesting that the processing of metaphors has not become more costly as we moved away from the time these metaphors were originally written. However, a different representation of the metaphors across time emerges depending on the genre (literary or nonliterary). Indeed, the significant interaction between epoch and genre indicates that the semantic relationship between topic and vehicle followed a genre-dependent diverging trajectory over time. Starting from an equivalent semantic representation in 19th-century literary and nonliterary VSMs, metaphors' terms in the 21st century became increasingly distant (less similar) in literary texts, while they showed an increasing semantic similarity in nonliterary texts. This interaction suggests that the effort required to process these metaphors has actually decreased in modern nonliterary contexts; conversely, in the literary domain, these metaphors have become more semantically distant over time. From these results, we can argue that metaphors appeared as more striking expressions in contemporary literary texts, where the integration of the two semantic domains that compose the metaphors seemed to be more effortful, while the association of these domains seemed to be more straightforward in contemporary nonliterary texts.

This result can be linked to the nature of the corpora considered as literary and nonliterary texts, and some speculative hypotheses can be made in relation to broader patterns of stylistic changes that took place at a higher level. The modern nonliterary corpus consisted of web-crawled texts (Baroni et al., 2009), a linguistic variety that differs substantially from literature (Pistolesi, 2014). Scholars have noted that the language of the web is characterized by greater variability, brevity, fragmentation, fluidity, and loose use of language with an open, hybrid, and changeable nature (Santini, 2007), while modern Italian literature is marked by a loss of literariness and a shift toward orality, highlighted by lexicon and discursive signals specific to a standardized spoken language (Dardano, 2014). In the modern nonliterary corpus, the vectors of topics and vehicles tend to be closer, indicating more entrenched associations and easier connections between distant concepts. In the modern literary corpus, by contrast, the vectors tend to be more distant, suggesting that novel associations are less frequent and therefore more striking. We can conclude that in contemporary nonliterary language, novel metaphorical associations required reduced processing demands due to increased frequency of novel associations between words, while in literary language, they remain marked and, hence, cognitively demanding.

Beyond semantic shift and stylistic change, the evolution of metaphors may also reflect broader cultural dynamics. Certain metaphorical mappings could have gained traction through globalization, making certain conceptualizations more familiar across cultures where such mappings, for instance “TIME is a COMMODITY”, had not previously been established (Kövecses, 2015). Banaruee et al. (2024) attributed this phenomenon largely to Americanization, which they showed has influenced metaphor ratings across different languages.

It is noteworthy that, while in the 21st century the status of metaphors is quite distinct across genres, in line with the marked stylistic differences between literary and nonliterary texts reported for modern Italian (Aprile, 2014), this distinction was not present in the 19th-century texts. Indeed, the interaction effect suggests that in the 19th century, metaphors were processed similarly regardless of the textual genre. This can be linked to the shared stylistic features of 19th-century literary and nonliterary texts. Texts from both genres were cultural products designed for the educated portion of the population and therefore characterized by high-register language, with echoes of the classical tradition (Aprile, 2014). Although elements of stylistic innovation can be found starting from the 19th century, language essays, manuals, and newspaper texts, which compose our nonliterary training sets, continue the expressive heritage of earlier prose and poetry (Masini, 1994).

The possibility of interpreting these results from a language-specific stylistic perspective is also supported by the results that we obtained in a cross-linguistic proof-of-concept study, aiming at validating the methodology in another language. In Mangiaterra et al. (2024), we applied the same methodology to English literary metaphors but found a different pattern of results: while for Italian we found an interaction of epoch and genre, for English we found only an effect of genre. In particular, the two terms composing English metaphors were closer in literary text compared to nonliterary ones, with no significant change observed across the two epochs. Again, these results seemed in line with insights from the history of English, a language which reached a form similar to the contemporary one earlier than Italian (Wyld, 1936). The stable nature of the English language in the last two centuries also seems to extend to the stability of its stylistic features, such as metaphors. These cross-linguistic differences seem to suggest that the patterns emerging from this kind of analysis are language-specific and that metaphor evolution follows the broader stylistic trajectories of the language in which they are embedded.

4.2 Single-word features and their impact on metaphor evolution

The claim that metaphor evolution is embedded in the broader temporal changes of the language at stake is further strengthened when we consider the role of single-word variables. First, it is important to note that single-word variables showed known patterns of relationships, confirming the validity of our vector space models. In particular, vector coherence was associated with frequency, consistent with the findings that frequent words change less over time (Hamilton et al., 2016), and with semantic neighborhood density, consistent with the findings that words change more in the sparse portions of the semantic network (Ryskina et al., 2020). We also found that frequency was negatively correlated with semantic neighborhood density (consistent with Buchanan et al., 2001), suggesting that highly frequent words tend to occur in a variety of contexts and develop more loosely related relationships with their neighbors (less dense meaning), while low-frequency words tend to develop a higher level of specificity and tighter relationship with their neighbors (more dense meaning, see Rambelli and Bolognesi, 2024).

Moving to the analysis of the contribution of lexical-semantic features of topics and vehicles in shaping the diachronic evolution of metaphors, both the correlation analysis and the linear mixed models highlighted the centrality of the vehicle semantic neighborhood density and the topic vector coherence. Semantic neighborhood density of the vehicle had an overall effect, which appeared stronger in 21st century nonliterary text, as denser vehicles were associated with metaphors with semantically closer terms. As suggested by Al-Azary and Buchanan (2017), when a word has many close semantic neighbors, it lacks the flexibility to acquire new metaphorical associations. Instead, a vehicle with low semantic neighborhood density, with its looser semantic associations, may evoke the wide array of weak implicatures characteristic of literary metaphors, which resulted in greater appreciation (Reid et al., 2023), but also higher processing costs, as suggested by our results. In the 21st century nonliterary texts, high semantic neighborhood density pulls the terms together into the entrenched, easily processed associations reflected in our high similarity results, such as in the highly dense vehicle “flour” in the metaphor “sky of flour”, as compared to the sparser vehicle “inebriation” in the metaphor “inebriation of light”.

Moreover, we found that in 21st-century literary texts, where metaphors are associated with higher processing costs, these demands seem to be driven by the vector coherence of the topic. In deriving the meaning of a metaphor, the function of the topic is to help promote the salient features of the vehicle necessary to reach the intended interpretation. Topics whose meanings have changed greatly over time have probably acquired a sparser set of semantic relationships and are considered less concrete (Azarbonyad et al., 2017) and acquired later (Cassani et al., 2021). All these features contribute to making the shifted meaning of these topics less accessible and less prone to providing the straightforward contextual constraints necessary to guide the reader through the process of meaning derivation of complex literary metaphors. This can be connected to the role that Relevance Theory assigns to context (Wilson and Carston, 2007), which notably includes the metaphor's topic, in guiding the conceptual adjustment of the vehicle. A topic whose meaning has undergone substantial shift offers a weaker anchor from which that adjustment can proceed. An example of this process can be provided by the metaphor “vento di lode” (Eng. “wind of praise”). The topic “lode” (Eng. “praise”) shifted greatly between literary texts of the 19th-century (where the meaning was in the semantic domain of “kindness”, “gratitude”, “appreciation”) and 21st-century (where the meaning is in the semantic domain of “report card”, “degree”, “graduated”, given the prominence of the use in the expression “con lode” – Eng. “cum laude”), and therefore represented a less clear guide for the selection of the appropriate meaning of the metaphor, increasing its processing demands.

Overall, it emerged that the topic should have a stable meaning to guide the derivation of metaphor interpretation over time and reduce processing costs, and, at the same time, the vehicle should have a flexible meaning to allow the metaphor to emerge.

4.3 Limitations

Some limitations of the present study should be acknowledged. First, we note that cosine similarity between the two terms of a metaphor remains a distributional measure; while it successfully captures certain aspects of a reader's processing experience, including difficulty (McGregor et al., 2019), it does not represent a complete model of how metaphors are cognitively represented in the mind. What we presented is the diachronic evolution of semantic relationships, plausibly associated with a change in processing demands, rather than direct evidence of how literary metaphors are processed by readers. Future studies should combine diachronic distributional models with behavioral and neurophysiological measures of metaphor comprehension to directly validate the relationship between semantic similarity and processing effort across historical periods.

Second, our approach focused on the evolution of topic-vehicle relationships from a purely propositional perspective, being based on text-based distributional models of language. This left underexplored the imagistic and sensorimotor dimensions, which play a role in metaphor processing (Bambini et al., 2025; Canal et al., 2022; Khatin-Zadeh et al., 2026), especially for literary metaphors (Carston, 2010), and may have impacted the evolution of metaphors across time. However, this multimodal evolution cannot be explored with currently available diachronic models, both because text-based models exhibited weaknesses in dealing with imagistic and sensorimotor aspects of metaphors (Mangiaterra et al., 2025a,b) and because we currently lack the frameworks to train diachronic Vision-Language Models, making a visual equivalent to diachronic word embeddings unfeasible. Interestingly, in the Figurative Archive (Bressler et al., 2026), we found that semantic distance is negatively correlated with metaphor concreteness, with higher topic-vehicle semantic distance associated with lower metaphor concreteness. In light of this result, we can hypothesize that changes in topic-vehicle distance may also reflect changes in the perceptual grounding of the metaphor over time. For instance, metaphors whose topic and vehicle became closer over time may also be considered more concrete, although further targeted analysis would be needed to confirm this hypothesis.

Finally, the binary distinction between literary and nonliterary corpora necessarily led to the creation of genre proxies rather than strictly controlled experimental categories. Because these proxies conflate multiple stylistic dimensions, such as register and target audience, the inherent heterogeneity of the data introduces potential confounding variables. This is particularly relevant for the 21st-century nonliterary corpus, as its reliance on web-based texts introduces distinct shifts in medium, discourse type, and communicative function compared to the 19th-century materials. Consequently, the present findings should be interpreted as reflecting the evolution of literary metaphors within ecologically valid linguistic environments rather than the isolated contribution of historical time. Future studies could disentangle these factors by comparing more homogeneous textual genres across epochs (e.g., newspapers, essays, or narrative prose) or by explicitly modeling stylistic variables alongside temporal change.

5 Conclusions

This study examined the evolution of literary metaphors in terms of semantic similarity between topics and vehicles, expanding the diachronic application of word embeddings to the analysis of complex expressions such as metaphors. Our results showed that, while overall the semantic distance between topics and vehicles, taken as a proxy for processing demands, did not change between the 19th century and today, it varied in relation to different textual genres, following the broader patterns of stylistic evolution in the Italian language. Since literary language and nonliterary language were very similar in the past, metaphors did not differ in their topic-vehicle semantic similarity in the two genres. In today's language, however, metaphors in literary texts show greater topic-vehicle distance, reflecting a plainer and simpler stylistic register, while metaphors in the creative nonliterary language of the web show reduced distance, suggesting a greater ease of conceptual connection between domains.

Methodologically, this work confirms the possibility of applying temporal embeddings to examine the evolution of multi-word expressions (Jenkins et al., 2025), extending their scope to figurative expressions. Regarding metaphor processing, our results empirically highlight the importance of accounting for the textual context in which metaphors are embedded, which may determine different patterns of semantic elaboration (Steen, 1989). The point of optimal distance between topics and vehicles, which “should be sufficiently distant to emphasize differences but sufficiently close to maintain similarities” (Katz, 1989), seems to be the result of a complex balance, modulated by the nature of the words composing the metaphors but also sensitive to the broader stylistic features of the language. Ultimately, this study demonstrates that what makes a metaphor difficult to process is not an inherent property, but a dynamic process influenced by diachronic genre-based shifts and by the structure of single-word semantic networks.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. In the preparation of this work, the author(s) used GPT-5.2 and Grammarly in order to: Grammar and spelling check. After using these tools, the author(s) reviewed and edited the content as needed and take full responsibility for the publication's content.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Seana Coulson, University of California, San Diego, CA, United States

Reviewed by: Hassan Banaruee, University of Education Weingarten, Germany

Francesca Strik Lievers, University of Genoa, Italy

Footnotes

Thirty-seven of these metaphors displayed a definite article in the prepositional phrase, i.e., “A of the B” (e.g., “abbraccio del sonno”, Eng. “embrace of the sleep”).