💬 NLP

Floating or Suggesting Ideas? A Large-Scale Contrastive Analysis of Metaphorical and Literal Verb-Object Constructions

This large-scale study analyzes nearly 300 English verb-object pairs to reveal that while metaphorical contexts generally exhibit higher affective load and lexical diversity compared to literal ones, there is no single consistent distributional pattern distinguishing them, as differences are largely specific to individual constructions.

Original authors: Prisca Piccirilli, Alexander Fraser, Sabine Schulte im Walde

Published 2026-04-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Prisca Piccirilli, Alexander Fraser, Sabine Schulte im Walde

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery: How does our brain tell the difference between saying something "literally" and saying it "metaphorically"?

For a long time, linguists and computer scientists have known that metaphors are everywhere. We say we "grasp" an idea (like holding a physical object) or "float" a suggestion (like a boat on water). But while we know what metaphors are, we haven't really had a massive, data-driven map of how they behave differently from literal language in the wild.

This paper is like a giant linguistic census. The researchers didn't just look at a few examples; they dug into 2 million sentences from the internet to compare pairs of phrases that mean roughly the same thing but use different words.

Here is the breakdown of their investigation, using some everyday analogies:

1. The Setup: The "Twin" Experiment

Imagine you have 297 pairs of "linguistic twins."

Twin A (Literal): "Suggest an idea."
Twin B (Metaphorical): "Float an idea."

Both twins are trying to do the exact same job: proposing a thought. But the researchers wanted to see if the environment around them was different. Did "floating" happen in a stormy, emotional sea? Did "suggesting" happen in a calm, orderly office?

To find out, they used five high-tech "microscopes" (NLP tools) to scan the sentences surrounding these twins. These tools measured things like:

Emotion: Is the sentence happy, scary, or angry?
Vocabulary: Are the words common and simple, or rare and fancy?
Structure: Is the sentence a rigid, predictable block, or a flowing, chaotic river?
Cohesion: Do the words stick together tightly, or do they drift apart?

2. The Big Picture: The "General Rules" (Cross-Pair Analysis)

First, the researchers looked at all the twins together to see if there were general rules. They found some clear patterns, like two different neighborhoods:

The Literal Neighborhood (The "Office"):
- Vibe: Stable, efficient, and repetitive.
- Characteristics: The words used here are high-frequency (everyone knows them), the sentences are very cohesive (words stick together like glue), and the structure is regular. It's like a well-organized spreadsheet.
- Analogy: It's like wearing a business suit. It's predictable, professional, and easy to process.
The Metaphorical Neighborhood (The "Art Gallery"):
- Vibe: Emotional, colorful, and diverse.
- Characteristics: These sentences are packed with emotion (fear, joy, disgust), sensory details (things you can see or touch), and vocabulary diversity (using many different words). They are less cohesive and more "jumpy."
- Analogy: It's like wearing a clown suit or a painter's smock. It's vivid, expressive, and grabs your attention, but it's less predictable.

The Takeaway: Generally, literal language is the "workhorse" (reliable, frequent, cohesive), while metaphorical language is the "artist" (emotional, sensory, diverse).

3. The Twist: The "Individual Personalities" (Within-Pair Analysis)

Here is where the plot thickens. The researchers then zoomed in on individual pairs to see if the "General Rules" applied to every single twin.

The Surprise: The rules didn't hold up for everyone!

The "True Believers": Some pairs were extreme. For example, "Breathe life" (metaphor) vs. "Instill life" (literal). These were so different that they looked like they were speaking different languages.
The "Chameleons": Most pairs, however, were mixed bags.
- Example: Take "Absorb a bill" (metaphor) vs. "Pay a bill" (literal).
- In some ways, "Absorb" looked more metaphorical (more emotional, more complex).
- But in other ways, "Pay" looked more metaphorical (more diverse vocabulary).
- Result: You can't just say "Metaphors are always emotional." For this specific pair, the metaphor wasn't the emotional one!

The Big Discovery: The difference isn't just about the verb (the action). It's about the whole construction (the verb + the object).

"Grasp an idea" is a metaphor.
"Grasp a rope" is literal.
The word "Grasp" changes its personality depending on who it's hanging out with.

4. The Conclusion: No Single "Metaphor Detector"

The researchers concluded that there is no single "magic switch" or universal pattern that separates all metaphors from all literal language.

Old Idea: "Metaphors are always emotional and complex."
New Reality: "It depends entirely on the specific phrase."

Some metaphors are as boring and structured as literal language. Some literal phrases are as wild and emotional as metaphors. The "metaphor-ness" is a property of the specific combination of words, not just the words themselves.

Why Does This Matter?

If you are building an AI (like a chatbot or a translator), you can't just teach it a simple rule like "If the sentence is emotional, it's a metaphor." You have to teach the AI to look at the specific context of the phrase.

In a nutshell:
Language is like a massive orchestra. We used to think the "Metaphor" section always played jazz (wild, emotional) and the "Literal" section always played classical (structured, predictable). This study shows that while there are general trends, every musician (every phrase) has their own unique style. To understand the music, you have to listen to the specific duet, not just the whole band.

1. Problem Statement

While metaphor is a fundamental component of human cognition and language, large-scale empirical comparisons between metaphorical and literal language remain limited. Previous research has largely focused on cognitive or psycholinguistic perspectives, often lacking systematic, data-driven contrasts where metaphorical and literal expressions compete to convey the exact same meaning.

The authors identify a gap in understanding:

Do metaphorical and literal expressions exhibit consistent, global distributional differences across a broad range of linguistic contexts?
Or are the differences highly specific to individual verb-object (VO) pairs, suggesting that "metaphor" is not a monolithic category with a single linguistic signature?

2. Methodology

Dataset Construction

Source: The study utilizes a curated set of 297 near-synonymous English Verb-Object (VO) pairs. Each pair consists of one metaphorical variant (e.g., float an idea) and one literal variant (e.g., suggest an idea) that share the same underlying semantic meaning.
Corpus Scale: The authors extracted approximately 2 million sentences from the ENCOW corpus (English Corpora from the Web), which covers diverse genres (blogs, news, academic texts). This ensures the analysis reflects naturalistic language usage rather than controlled experimental data.
Preprocessing: To prevent lexical bias (where the specific words themselves drive the features), the verb and object in every sentence were masked (replaced with MASK) before feature extraction. This forces the NLP tools to analyze the surrounding context rather than the target words.

Feature Extraction

The study employs five established NLP tools to extract 2,293 cognitive and linguistic features across four dimensions:

SEANCE (1,271 features): Sentiment, emotion, and social cognition (affective semantics).
TAALES (461 features): Lexical sophistication, word frequency, concreteness, and psycholinguistic properties.
TAASSC (355 features): Syntactic complexity and structural variety.
TAACO (168 features): Text cohesion (lexical overlap, connectives, semantic similarity).
TAALED (38 features): Lexical diversity metrics (e.g., Type-Token Ratio, MTLD).

Analytical Approach

The authors address two Research Questions (RQs) using two distinct analytical levels:

RQ1 (Cross-Pair Analysis): Do metaphorical and literal contexts differ consistently across all 297 pairs?
- Method: Wilcoxon signed-rank tests on mean feature differences between metaphorical and literal variants, normalized via z-scores. Features were filtered by statistical significance ( $p < 0.05$ ) and effect size ( $\geq 1$ standard deviation).
RQ2 (Within-Pair Analysis): Do specific VO pairs show strong internal divergence, and are there systematic patterns for "metaphor-dominant" vs. "literal-dominant" pairs?
- Method: Calculation of mean absolute differences per pair. Pairs with a divergence $\geq 1$ standard deviation across at least three tools were classified as "clearly distinctive."

3. Key Contributions

Large-Scale Contrastive Dataset: Creation of a dataset containing ~2 million sentences covering 297 metaphorical/literal VO pairs, enabling a controlled comparison of usage contexts.
Multidimensional Feature Analysis: The first study to simultaneously analyze over 2,200 features spanning affective, lexical, syntactic, and discourse-level properties to distinguish metaphor from literal language.
Granular Insight into Construction Specificity: Moving beyond "metaphor detection" to a fine-grained understanding of how metaphoricality manifests differently depending on the specific verb-object construction.

4. Key Results

A. Frequency and Distribution (Section 4.1)

Literal Dominance: Literal usages are generally more frequent (1.7M sentences) than metaphorical usages (333k sentences).
U-Shaped Pattern: High-frequency pairs tend to be either strongly literal or strongly metaphorical (lexicalized/conventionalized), while low-frequency pairs show balanced usage.
Verb-Object Unit: Analyzing verbs in isolation is misleading; the metaphorical nature is often driven by the specific VO pairing (e.g., reduce currency is balanced, but the verb reduce alone is overwhelmingly literal).

B. Cross-Pair Analysis (Global Tendencies) (Section 4.2)

When aggregating across all pairs, distinct global patterns emerge:

Literal Contexts: Characterized by higher lexical frequency, stronger local cohesion, structural regularity, and positive affect (institutional, goal-directed). They are easier to process (faster lexical decision times).
Metaphorical Contexts: Characterized by higher affective loading (negative affect, fear, disgust), sensorimotor grounding (embodied experience), greater lexical diversity, and constructional specificity. They rely on vivid, imageable, and semantically rich terms.

C. Within-Pair Analysis (Heterogeneity) (Section 4.3)

This is the most critical finding. The global tendencies do not translate to uniform behavior at the individual pair level.

High Heterogeneity: Of the 93 "clearly distinctive" pairs (those showing strong divergence), only a small subset showed consistent directional bias across all tools.
- 7 pairs were consistently literal-dominant across tools.
- 6 pairs were consistently metaphor-dominant across tools.
- 80 pairs showed mixed patterns, where different linguistic dimensions pulled in opposite directions (e.g., a metaphorical variant might be more cohesive but less affective than its literal counterpart).
Construction-Specificity: The divergence is driven by the specific construction, not the verb alone. For example, instill idea vs. breathe idea behaves differently than instill spirit vs. breathe spirit.

5. Significance and Implications

Theoretical Impact: The findings challenge the assumption that metaphorical language has a single, stable "linguistic signature." Instead, metaphor is usage-based and construction-specific. The differences between metaphor and literal language are often idiosyncratic to the specific lexical pairing rather than a universal property of "metaphor."
Methodological Impact: The study demonstrates the necessity of pair-level granularity in computational linguistics. Aggregating data without accounting for individual construction variance can obscure the true nature of figurative language.
Applications:
- NLP Tasks: Improves machine translation, sentiment analysis, and information retrieval by acknowledging that metaphorical and literal paraphrases have distinct contextual profiles.
- Future Modeling: Provides a rich feature set for training machine learning models to predict metaphorical usage, moving beyond simple detection to nuanced modeling of cognitive and linguistic properties.

Conclusion

The paper concludes that while there are broad, global tendencies distinguishing metaphorical from literal language (e.g., metaphors are more affective and diverse), these patterns are not consistent across all instances. The primary unit of analysis for metaphoricality is the verb-object construction, not the lexical verb or the category of "metaphor" in isolation. This necessitates a shift in computational approaches from treating metaphor as a monolithic class to modeling it as a spectrum of construction-specific variations.