pith. sign in

arxiv: 2604.08275 · v1 · submitted 2026-04-09 · 💻 cs.CL

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

Prisca Piccirilli , Alexander Fraser , Sabine Schulte im Walde This is my paper

Pith reviewed 2026-05-10 16:45 UTC · model grok-4.3

classification 💻 cs.CL
keywords metaphorliteral languageverb-object constructionsdistributional analysisNLP featurescorpus linguisticsconstruction-specific patterns
0
0 comments X

The pith

Metaphorical and literal verb-object constructions show no single consistent pattern across linguistic features.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper examines nearly 300 English verb-object pairs, such as metaphorical 'float an idea' versus literal 'suggest an idea,' across roughly two million corpus sentences. It draws on five NLP tools to pull more than two thousand features that track affective tone, word choice, sentence structure, and discourse connections. Overall comparisons across pairs indicate that literal contexts tend toward higher frequency, cohesion, and regularity while metaphorical ones show stronger emotional load, imageability, diversity, and constructional uniqueness. Yet when each pair is examined separately, the effects prove inconsistent, with most pairs diverging from the group trend. This leads to the conclusion that distinctions between metaphor and literal language are tied to specific constructions rather than following any uniform distributional signature.

Core claim

The analysis of 297 verb-object pairs in approximately 2 million sentences reveals that literal contexts generally exhibit higher lexical frequency, cohesion, and structural regularity, whereas metaphorical contexts display greater affective load, imageability, lexical diversity, and constructional specificity. Within-pair examinations show substantial heterogeneity, with most pairs displaying non-uniform effects. These findings indicate that no single, consistent distributional pattern distinguishes metaphorical from literal usage; differences are largely construction-specific.

What carries the argument

Contrastive analysis of 2,293 cognitive and linguistic features extracted by five NLP tools across metaphorical and literal contexts of 297 verb-object pairs.

If this is right

  • Metaphor research should prioritize construction-specific analyses over searches for universal distinguishing rules.
  • Large-scale corpus work with multi-tool feature extraction can expose fine-grained, pair-dependent usage contrasts in language.
  • Computational systems for metaphor detection or generation would gain accuracy by incorporating information tied to individual verb-object pairs.
  • Within-pair comparisons alongside cross-pair trends offer a more precise way to map how literal and figurative language actually behave.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • General-purpose metaphor identification tools that rely on broad distributional signals may underperform unless they also model specific verb-object combinations.
  • The observed heterogeneity raises the possibility that similar construction-specific patterns appear in other languages or other near-synonymous expression types.
  • Corpus-derived contrasts could be tested against psycholinguistic measures to check whether large-scale data align with human processing differences at the level of individual constructions.

Load-bearing premise

The 2,293 features extracted by the five NLP tools fully and without systematic bias capture all relevant affective, lexical, syntactic, and discourse properties that could distinguish metaphorical from literal uses.

What would settle it

Discovery of even one feature or small consistent set of features that reliably separates metaphorical from literal contexts across most of the 297 pairs would falsify the claim of no single pattern.

Figures

Figures reproduced from arXiv: 2604.08275 by Alexander Fraser, Prisca Piccirilli, Sabine Schulte im Walde.

Figure 1
Figure 1. Figure 1: Pipeline of our approach: based on an existing set of metaphorival vs. literal [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Metaphorical vs. literal VO preferences against total frequencies on a logarithmic scale, with points color-coded by dominance (green for literal-dominant, orange for metaphorical-dominant). The ratio ranges from 0 (entirely literal usage) to 1 (entirely metaphorical usage). Labels indicate the dominant VO followed by its corresponding counterpart in parenthesis, For example, met. VO breathe life (lit. VO … view at source ↗
Figure 3
Figure 3. Figure 3: Metaphorical vs. literal verb-only preferences (independently of their direct objects) against [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Tool-level aggregation of distributional dominance for 93 clearly-distinctive [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
read the original abstract

Metaphor pervades everyday language, allowing speakers to express abstract concepts via concrete domains. While prior work has studied metaphors cognitively and psycholinguistically, large-scale comparisons with literal language remain limited, especially for near-synonymous expressions. We analyze 297 English verb-object pairs (e.g., float idea vs. suggest idea) in ~2M corpus sentences, examining their contextual usage. Using five NLP tools, we extract 2,293 cognitive and linguistic features capturing affective, lexical, syntactic, and discourse-level properties. We address: (i) whether features differ between metaphorical and literal contexts (cross-pair analysis), and (ii) whether individual VO pairs diverge internally (within-pair analysis). Cross-pair results show literal contexts have higher lexical frequency, cohesion, and structural regularity, while metaphorical contexts show greater affective load, imageability, lexical diversity, and constructional specificity. Within-pair analyses reveal substantial heterogeneity, with most pairs showing non-uniform effects. These results suggest no single, consistent distributional pattern that distinguishes metaphorical from literal usage. Instead, differences are largely construction-specific. Overall, large-scale data combined with diverse features provides a fine-grained understanding of metaphor-literal contrasts in VO usage.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The paper conducts a large-scale corpus study of 297 English verb-object pairs (e.g., 'float idea' vs. 'suggest idea') drawn from ~2M sentences. Five off-the-shelf NLP tools extract 2,293 features spanning affective, lexical, syntactic, and discourse properties. Cross-pair aggregate analyses report that literal contexts exhibit higher lexical frequency, cohesion, and structural regularity, while metaphorical contexts show elevated affective load, imageability, lexical diversity, and constructional specificity. Within-pair analyses find substantial heterogeneity, with most pairs exhibiting non-uniform feature effects. The authors conclude that no single, consistent distributional pattern distinguishes metaphorical from literal usage; differences are largely construction-specific.

Significance. If the heterogeneity result survives controls for sample-size imbalance, the work supplies a valuable large-scale empirical demonstration that metaphor-literal contrasts in VO constructions lack a uniform signature and are instead highly pair-dependent. The scale (~2M sentences) and feature breadth (2,293 dimensions from multiple tools) constitute clear strengths, offering a finer-grained picture than prior small-scale or single-feature studies and supplying testable predictions for cognitive linguistics and computational metaphor detection. The absence of a single overarching pattern challenges assumptions in the literature that distributional cues are broadly reliable across constructions.

major comments (1)
  1. Within-pair analyses (Results section): The central claim that 'most pairs showing non-uniform effects' and thus 'differences are largely construction-specific' rests on per-pair statistical tests. Because sentence counts per pair are almost certainly highly skewed (Zipfian) in a 2M-sentence corpus, low-N pairs will have low power to detect differences, producing more non-significant results and inflating the count of 'non-uniform' pairs. The manuscript does not report per-pair N values, conduct power analyses, apply frequency-weighted aggregates, or correct for multiple testing across 2,293 features. This statistical design issue is load-bearing for the heterogeneity conclusion and could be addressed by re-running the within-pair tests with explicit N reporting, effect-size confidence intervals, or subsampling to equalize power.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their careful and constructive review. The concern regarding statistical power and reporting in the within-pair analyses is well-taken, and we address it directly below with plans for revision.

read point-by-point responses

  1. Referee: Within-pair analyses (Results section): The central claim that 'most pairs showing non-uniform effects' and thus 'differences are largely construction-specific' rests on per-pair statistical tests. Because sentence counts per pair are almost certainly highly skewed (Zipfian) in a 2M-sentence corpus, low-N pairs will have low power to detect differences, producing more non-significant results and inflating the count of 'non-uniform' pairs. The manuscript does not report per-pair N values, conduct power analyses, apply frequency-weighted aggregates, or correct for multiple testing across 2,293 features. This statistical design issue is load-bearing for the heterogeneity conclusion and could be addressed by re-running the within-pair tests with explicit N reporting, effect-size confidence intervals, or subsampling to equalize power.

    Authors: We agree that sentence counts per pair are likely to be highly skewed in a corpus of this size, which can reduce power for low-N pairs and potentially inflate the proportion of pairs classified as non-uniform. The original manuscript did not report per-pair N distributions, conduct power analyses, use frequency-weighted aggregates, or apply multiple-testing corrections across the 2,293 features. In revision we will add: (i) a summary of the per-pair sentence counts (including min, max, median, and a histogram), (ii) effect sizes with confidence intervals for all within-pair comparisons, (iii) a discussion of how low-N pairs affect the heterogeneity result, and (iv) a robustness check via subsampling from high-N pairs to equalize power. For multiple testing we will clarify that the analysis was exploratory and focused on directional consistency rather than isolated p-values; we will also report FDR-adjusted results for the main within-pair tests. These additions will make the heterogeneity claim more robust and transparent. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical corpus analysis

full rationale

The paper conducts a large-scale observational study on ~2M corpus sentences for 297 pre-identified VO pairs. It applies five external off-the-shelf NLP tools to extract 2,293 features and performs statistical comparisons (cross-pair and within-pair). No equations, first-principles derivations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the abstract or described methodology. The central claim of construction-specific heterogeneity follows directly from the data distributions rather than reducing to any input by construction. Potential statistical-power artifacts from uneven sentence counts per pair are a methodological concern but do not constitute circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The paper is observational and relies on standard assumptions about English corpus data and NLP feature validity rather than new free parameters or invented entities.

axioms (2)
  • domain assumption The 297 selected verb-object pairs are near-synonymous and representative for comparing metaphorical versus literal usage.
    The entire contrastive design rests on this selection being appropriate and balanced.
  • domain assumption The five NLP tools produce reliable measurements of the intended cognitive and linguistic properties.
    All 2293 features are treated as valid proxies for affective load, imageability, cohesion, etc.

pith-pipeline@v0.9.0 · 5519 in / 1413 out tokens · 55467 ms · 2026-05-10T16:45:35.285346+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages · 1 internal anchor

  1. [1]

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

    Introduction Metaphor is a "necessary" language feature of everyday thought and communication that allows speakerstoconceptualizeabstractideasintermsof more concrete domains (Ortony, 1975; Lakoff and Johnson, 1980; van den Broek, 1981; Schäffner, 2004, i.a.). While metaphor has been extensively studiedincognitivelinguisticsandpsycholinguistics (Gibbs, 198...

    work page internal anchor Pith review Pith/arXiv arXiv 1975

  2. [2]

    Datasets and Linguistic Features VOPairsWe downloaded theVOs from Piccirilli et al. (2024). They collected a set of 47VOs from previous work (Mohammad et al., 2016; Shutova, 2010; Piccirilli and Schulte im Walde, 2021; Stowe et al., 2022), which they semi-automatically ex- tended by collecting the most frequently observed shared direct objects of the corr...

    work page 2024

  3. [3]

    Comparison of Metaphorical and Literal Verb-Objects: Approach For comparing metaphorical vs. literalVOs, this work addresses two research questions: RQ1Arethereconsistentdifferencesincognitive and linguistic feature patternsbetween literal and metaphorical language (=cross-pair)? RQ2Are therespecificVOpairsshowing a par- ticularly strong contrast in cogni...

    work page

  4. [4]

    Analyses and Discussion Thissectionprovidesouranalysesanddiscussions regarding our two research questionsThese two main studies in Sections 4.2 and 4.3 are preceded by a frequency analysis (Section 4.1). 4.1. Metaphorical vs. Literal Preferences FrequencyTable 1 presents an overview of the distribution of literal and metaphorical sentences across our 297V...

    work page 1980

  5. [5]

    Fears, suspicions, resentments and hatred have‌poisoned ‌relationshipsacross that divide in ways that threaten us all

    with a very high frequency. This finding high- lights how crucial it is to consider the verb and its object as a unit, as verbs behave very differently in terms of their frequency and metaphoricity when considered with or without their objects. 4.2. Cross-Pair Analysis: Feature-based Comparison of Metaphorical and Literal VOs In this section, we address o...

    work page 1980

  6. [6]

    Conclusion This study provides a large-scale, feature-based comparison of metaphorical and literal verb–object constructions in natural English sentences. Lever- aging nearly two million corpus instances and over 2,200 linguistic features derived from five comple- mentaryNLPtools, we investigated both cross-pair and within-pair patterns of divergence. At ...

    work page

  7. [7]

    Limitations Several limitations should be acknowledged. First, our analysis is restricted to the English language, and the patterns observed in this work may not generalize to other languages with different mor- phological, syntactic, or metaphorical conventions, especially given the fact that theNLPtools we used are built on and for the English language....

    work page

  8. [8]

    Prisca Piccirilli is also supported by the Studiens- tiftung des deutschen Volkes

    Acknowledgments ThisresearchwassupportedbytheDFGResearch Grants SCHU 2580/4-1 (MUDCAT: Multimodal Di- mensions and Computational Applications of Ab- stractness) and SCHU 2580/7-1 and FR 2829/8-1 (MeTRapher: Learning to Translate Metaphors). Prisca Piccirilli is also supported by the Studiens- tiftung des deutschen Volkes. We are grateful to Annerose Eiche...

    work page

  9. [9]

    Only the Tip of the Ice- berg

    Bibliographical References Dawn G. Blasko. 1999. "Only the Tip of the Ice- berg": Who Understands what about Metaphor? Journal of Pragmatics, 31:1675–1683. Lera Boroditsky. 2000. Metaphoric Structuring: Un- derstanding Time Through Spatial Metaphors. Cognition, 75(1):1–28. Brian Bowdle and Dedre Gentner. 2005. The Career of Metaphor.Psychological Review, ...

    work page 1999

  10. [10]

    GeorgeLakoffandMarkJohnson.1980.Metaphors We Live By

    Assessing the Validity of Lexical Diversity Using Direct Judgements.Language Assess- ment Quarterly, 18(2):154–170. GeorgeLakoffandMarkJohnson.1980.Metaphors We Live By. University of Chicago Press, Chicago. Saif Mohammad, Ekaterina Shutova, andPeter Tur- ney. 2016. Metaphor as a Medium for Emotion: An Empirical Study. InProceedings of the Fifth Joint Con...

    work page 1980

  11. [11]

    InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics, pages 5375–5388, Dublin, Ireland

    IMPLI: Investigating NLI Models’ Perfor- mance on Figurative Language. InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics, pages 5375–5388, Dublin, Ireland. Association for Computational Linguistics. Mark Turner. 1996.The Literary Mind: The Origins of Thought and Language. Oxford University Press, Oxford. Raymondvand...

    work page 1996

  12. [12]

    TheSyntacticFlexibilityofGermanandEn- glish Idioms: Evidence from Acceptability Rating Experiments.Journal of Linguistics, 60:1–38. A. Supplementary Materials A.1. Met vs. Lit Verb-Object Pairs Table 3: Metaphorical vs. Literal Verb–Object Pairs with Corpus Frequencies (N=297) MetaphoricalVOLiteralVOMet Count Lit Count absorb knowledge assimilate knowledg...

    work page