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

1. Problem Statement

The study addresses a gap in psycholinguistic and computational research regarding literary metaphors. While everyday metaphors are well-studied, literary metaphors remain under-investigated, particularly concerning the temporal dimension.

• The Gap: Most experimental studies use literary metaphors created centuries ago (e.g., 19th century) but test them on contemporary readers. It is unknown whether the cognitive effort required to process these metaphors has increased due to the "temporal gap" between the author's era and the modern reader.
• The Context: The Italian language has undergone significant stylistic shifts over the last two centuries. 19th-century language (both literary and non-literary) was formal and elite-oriented, whereas 21st-century language has diverged: modern literature has shifted toward oral/informal styles, while the web (non-literary) has become a highly creative, hybrid linguistic space.
• Research Question: Do the processing demands of 19th-century literary metaphors vary when viewed through the semantic lens of the 19th century versus the 21st century? How does textual genre (Literary vs. Non-literary) modulate this evolution?

2. Methodology

The authors employed Diachronic Distributional Semantics to model semantic change and estimate processing costs without relying on human behavioral data.

A. Data Collection (Corpora)

The study utilized a total of 124 million tokens from Italian corpora, stratified by Epoch (19th vs. 21st Century) and Genre (Literary vs. Non-literary):

• 19th Century:
  • Literary: Project Gutenberg prose and poetry (30M tokens).
  • Non-literary: Technical manuals, newspapers, diaries, and immigrant newspapers (32M tokens).
• 21st Century:
  • Literary: Contemporary literature (2000–2020) (20M tokens).
  • Non-literary: Web-crawled data (blogs, news, educational sites) (42M tokens).

B. Metaphor Dataset

• Source: 515 Italian literary metaphors extracted from 19th-century texts (e.g., "Capelli di fiamma" / "Hair of flame").
• Structure: All metaphors followed the "Noun of Noun" (A of B) structure to ensure syntactic consistency.
• Components: Each metaphor consists of a Topic (the subject, e.g., "Hair") and a Vehicle (the metaphorical term, e.g., "Flame").

C. Computational Modeling: Temporal Aligned Embeddings

To compare semantic spaces across time, the authors used the Temporally aligned Word Embeddings with a Compass (TWEC) model (Di Carlo et al., 2019).

1. Compass Model: A word2vec (CBOW) model was trained on the entire combined corpus to create an atemporal "compass" (a shared coordinate system).
2. Temporal Alignment: Four specific context matrices were trained on the four distinct slices (19th Lit, 19th NonLit, 21st Lit, 21st NonLit). These were initialized using the Compass matrix to ensure all vectors exist in the same semantic space, allowing for direct comparison of cosine similarity across epochs.

D. Metrics

Four key measures were computed for every metaphor in every slice:

1. Cosine Similarity (CS): Between the Topic and Vehicle vectors. Used as a proxy for processing cost (Lower CS = Higher semantic distance = Higher processing demand).
2. Vector Coherence (VC): The cosine similarity of a single word's vector between the 19th and 21st centuries. Measures semantic stability over time.
3. Semantic Neighborhood Density (SND): The average cosine similarity between a word and its 500 nearest neighbors. Measures semantic density/flexibility.
4. Frequency (Freq): Logarithmic relative frequency of the words.

E. Statistical Analysis

• Linear Mixed-Effects Models (LMM): Used to test the effects of Epoch, Genre, and their interaction on CS.
• Model Selection: Incremental models were compared using AIC, BIC, and Log-Likelihood to determine if adding word-level predictors (SND, VC, Freq) improved the fit.

3. Key Results

A. Main Effect of Genre and Epoch

• No Main Effect of Time: Overall, the semantic similarity (CS) of metaphors did not significantly change simply due to the passage of time.
• Significant Interaction (Epoch × Genre): The evolution of metaphors diverged based on genre:
  • Literary: Metaphors became less similar (lower CS) from the 19th to the 21st century. This implies increased processing demands for modern readers encountering 19th-century literary metaphors in a modern literary context.
  • Non-literary: Metaphors became more similar (higher CS) from the 19th to the 21st century. This implies decreased processing demands in modern web-based language.
• Historical Context: In the 19th century, there was no significant difference between literary and non-literary genres (both were high-register, elite language). The divergence is a modern phenomenon.

B. Role of Single-Word Features

The best-fitting model included Vector Coherence (VC) and Semantic Neighborhood Density (SND) of the Topic and Vehicle.

• Topic Vector Coherence (VC): In 21st-century literary texts, lower VC (meaning the topic's meaning has shifted significantly over time) predicted lower CS (higher difficulty).
  • Example: The word "praise" (lode) shifted from a general meaning of "gratitude" in the 19th century to a specific academic meaning ("cum laude") in the 21st century, making it a less stable guide for interpreting metaphors.
• Vehicle Semantic Neighborhood Density (SND): Higher SND (denser semantic neighbors) generally predicted higher CS (easier processing).
  • This effect was strongest in 21st-century non-literary texts. Dense vehicles (e.g., "flour") are entrenched in the web's language, making associations easier. Sparse vehicles (e.g., "inebriation") remain difficult.

4. Key Contributions

1. Methodological Innovation: Successfully extended diachronic distributional semantics from single-word analysis to multi-word expressions (metaphors). This allows for the tracking of complex phrase evolution without human ratings.
2. Empirical Evidence of Stylistic Divergence: Provided computational evidence that the "gap" between literary and non-literary language in Italian has widened significantly in the 21st century, whereas they were stylistically homogeneous in the 19th century.
3. Processing Cost Dynamics: Demonstrated that the difficulty of processing a metaphor is not an inherent property of the metaphor itself but a dynamic interaction between:
   • The temporal stability of the terms (VC).
   • The semantic density of the terms (SND).
   • The genre of the text in which the metaphor is evaluated.

5. Significance and Implications

• Cognitive Linguistics: Supports the theory that literary metaphors require "poetic effects" (cognitive effort) because they generate weak implicatures. The study shows this effort has increased for modern readers in literary contexts because modern literature has simplified, making 19th-century metaphors feel more "alien" or distant.
• Language Evolution: Highlights that the Web has become a space of high creativity where novel metaphorical associations are frequent and thus "entrenched" (easier to process), whereas modern literature has lost some of its traditional "literariness," making older metaphors harder to integrate.
• Cross-Linguistic Contrast: The authors note that a similar study on English showed different results (English metaphors remained stable), suggesting that metaphor evolution is language-specific and tied to the broader stylistic trajectories of the specific language (Italian vs. English).
• Future Applications: Validates the use of temporal embeddings as a tool for studying the history of figurative language, offering a scalable alternative to expensive human behavioral experiments for diachronic studies.