Floating or Suggesting Ideas ? A Large-Scale Contrastive Analysis of Metaphorical and Literal Verb–Object Constructions
2026
Read Floating or Suggesting Ideas ? A Large-Scale Contrastive Analysis of Metaphorical and Literal Verb–Object Constructions (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-08275.txt.
Introduction
Piccirilli, Fraser, and Schulte im Walde analyze 297 verb-object pairs, like float idea versus suggest idea, in about 2 million sentences. Metaphor fills daily language, letting speakers express abstract ideas through concrete domains. While prior work studied metaphors cognitively, broad comparisons with literal language are few, especially for near-synonyms. With five NLP tools, the team extracts 2,293 features covering feelings, words, and syntax. They ask whether features differ between metaphorical and literal uses, and whether single verb-object pairs diverge internally.
Catalog record
| YEAR | 2026 |
|---|---|
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 52658 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.08275
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Floating or Suggesting Ideas ? A Large-Scale Contrastive Analysis of Metaphorical and Literal Verb–Object Constructions Abstract Metaphor is a pervasive feature of everyday language, enabling speakers to express abstract concepts in terms of more concrete domains. While prior work has extensively examined metaphors from cognitive and psycholinguistic perspectives, large-scale empirical comparisons between metaphorical and literal language remain limited, especially when they compete in conveying the same meaning. We address this gap through a systematic contrastive analysis of 297 near-synonymous English verb–object ( VO ) pairs (e.g., float idea vs. suggest idea ) in approximately 2 million extracted corpus sentences, allowing us to examine their contextual usage. Using five established NLP tools – SEANCE, TAALES, TAASSC, TAACO and TAALED – we derive 2,293 metaphor-related cognitive and linguistic features capturing affective, lexical, syntactic, and discourse-level properties. We address two research questions: (i) whether these features consistently differ between metaphorical and literal language ( cross-pair analysis ), and (ii) whether individual VO pairs show strong internal divergence between their metaphorical and literal variants ( within-pair analysis ). Cross-pair results reveal robust global tendencies: literal contexts are associated with higher lexical frequency, stronger cohesion, and greater structural regularity, whereas metaphorical contexts show increased affective loading, imageability, lexical diversity, and constructional specificity. In contrast, within-pair analyses reveal substantial heterogeneity and most pairs display non-uniform directional effects across cognitive or linguistic dimensions. Our findings suggest that there is no single, consistent distributional pattern that consistently distinguishes metaphorical from literal language. Instead, the differences we observe are largely construction-specific. Overall, by combining large-scale data with a broad range of cognitive and linguistic features, this study offers a fine-grained understanding of the contrast between metaphorical and literal VO usages. Keywords: metaphorical language, cognitive and linguistic empirical features, verb-object constructions Floating or Suggesting Ideas ? A Large-Scale Contrastive Analysis of Metaphorical and Literal Verb–Object Constructions Prisca Piccirilli 1 and Alexander Fraser 2,3 and Sabine Schulte im Walde 1 1 Institute for Natural Language Processing ( IMS ), University of Stuttgart, Germany 2 Technical University of Munich ( TUM ), Germany 3 Munich Center for Machine Learning ( MCML ), Germany {prisca.piccirilli, schulte}@ims.uni-stuttgart.de Abstract content 1. Introduction Figure 1 : Pipeline of our approach: based on an existing set of metaphorival vs. literal VO pairs, we extract sentences containing these VOs. We then run five NLP tools on all sentences (masking the VO, see Section 3 ) to extract more than 2,200 cognitive and linguistic features in order to analyze the contrast between metaphorical and literal language usage from (i) a cross-pair and (ii) a within-pair perspective. Metaphor is a "necessary" language feature of everyday thought and communication that allows speakers to conceptualize abstract ideas in terms of more concrete domains ( 24 ; 22 ; 35 ; 30 , i.a.) . While metaphor has been extensively studied in cognitive linguistics and psycholinguistics ( 10 ; 1 ; 12 ; 13 ; 32 , i.a.) , its empirical analysis in contrast to literal language use remains limited. For instance, float the idea and suggest the idea both convey the act of proposing an idea, but they differ in register, connotation, and metaphorical framing: float the idea employs a conceptual metaphor ( IDEAS ARE OBJECTS IN MOTION ), projecting an abstract act into a concrete physical domain ( 22 ) . In contrast, suggest the idea is literal and denotes the speech act directly. Pragmatically, the verb float conveys tentativeness or strategic intent, while suggest is more neutral and direct, and commonly used in formal and informal contexts ( 4 ; 31 ) . Understanding how metaphorical and literal expressions differ at scale is crucial not only for theoretical insights but also for practical applications such as machine translation, sentiment analysis, and information retrieval. In this paper, we fill this gap by conducting a large-scale, systematic analysis of metaphorical and literal language in contrast. We curate and analyze a dataset of approximately 2 million sentences containing semantically equivalent verb-object ( VO ) pairs, where one variant is metaphorical, e.g., float idea , and the other is literal, e.g., suggest idea . This paired design enables controlled comparison across usage contexts while holding the underlying meaning constant. Using a variety of NLP tools 1 1 1 https://www.linguisticanalysistools.org/ , we automatically extract a wide range of over 2,200 features, allowing us to explore cognitive, distributional and structural differences between metaphorical and literal usages. We conduct a large-scale and thorough descriptive analysis of these contrasts along two axes: we identify (i) which cognitive and linguistic features systematically distinguish metaphorical from literal language across VO pairs ( cross-pair ), and (ii) which VO pairs are “most strongly metaphorical” or “most strongly literal” across features but within-pair . Figure 1 provides a visual pipeline of our work study. In sum, our contributions are threefold: • We present a large-scale, contrastive dataset of metaphorical and literal verb-object expressions in naturalistic English text. • We conduct an empirical analysis of contextual feature-based differences between metaphorical and literal verb-object expressions. • We shed light on the internal metaphorical-literal differences at the verb-object level. By focusing on metaphorical and literal paraphrases with shared meaning, this work advances the computational study of figurative language beyond detection, offering a scalable and interpretable approach to metaphor-literal contrast. Our findings open new avenues for metaphor modeling and the injection of cognitive and linguistic knowledge, and it underlines the importance of metaphorical variation in language understanding tasks 2 2 2 All data and analysis files, including the full set of detailed outputs, are available in our repository at https://github.com/priscapiccirilli/Met-Lit-Contrast . . 2. Datasets and Linguistic Features VO Pairs We downloaded the VO s from 25 . They collected a set of 47 VO s from previous work ( 23 ; 31 ; 26 ; 33 ) , which they semi-automatically extended by collecting the most frequently observed shared direct objects of the corresponding verbs, resulting in a total of 297 metaphorical VO s and their corresponding literal paraphrases. VO Sentences 25 extracted 1,691 English sentences containing this set of VO pairs. We thus use this publicly available dataset, VOLIMET 3 3 3 https://github.com/priscapiccirilli/VOLIMET . However, since our goal is to analyze large amounts of natural language, we enlarge this dataset by automatically extracting all available English sentences containing the VO s from the ENCOW corpus ( 28 ; 29 , EN glish CO rpora from the W eb,) , where the object is the direct object of the verb. We extract the entire sentence regardless of its length or the distance between the verb and object, which allows us to capture contextual variations that are directly relevant to our features. ENCOW ’s coverage of multiple writing styles, registers, domains and genres (e.g., blogs, news articles, academic texts), ensures that our analyses reflect natural language in general. Our combined datasets therefore contain close to 2M English sentences covering the whole set of 297 metaphorical vs. literal VO s. Appendix 3 presents the full list of metaphorical and literal VO s and their sentence-level frequencies, reflecting how often each VO appears in context. Features Critically, we enriched our dataset with a range of cognitive and linguistic features: they serve as a basis for quantitative analyses of relationships