Metaphor-Literal Repository Overview

Updated 12 October 2025

Metaphor-Literal Repository is an organized resource that stores, maps, and retrieves relationships between metaphorical and literal expressions across various linguistic structures.
It employs formal ontologies, neural sequence-to-sequence models, and embedding transformations to automate detection, paraphrase generation, and semantic mapping.
The repository supports diverse applications in NLP, digital humanities, and semantic analysis while addressing challenges like context modeling, standardization, and multimodal integration.

A metaphor-literal repository is an organized resource for storing, mapping, retrieving, and analyzing relationships between metaphorical and literal language expressions. These repositories serve as crucial infrastructure in computational semantics, NLP, digital humanities, and cognitive linguistics by systematizing the connections between figurative and literal forms at various levels of linguistic structure, representation, and modality. Recent research leverages ontologies, large annotated corpora, embedding-based mappings, and aligned datasets to formalize both the conceptual mechanisms and practical mappings underlying metaphor-literal relationships.

1. Fundamental Concepts and Formal Ontologies

The design of a metaphor-literal repository presupposes precise formalization of metaphorical mappings. Amnestic Forgery (Gangemi et al., 2018) establishes a prominent example by extending MetaNet and the Framester linked-data schema to capture both intensional (conceptual) and extensional (instance-level) aspects of metaphor. This OWL ontology reifies frames (i.e., structured conceptual domains such as CRIME and DISEASE) and encodes explicit role mappings between source and target frames, following Conceptual Metaphor Theory. Blending operations are formally represented, enabling representation of creative outcomes such as “stony woman,” where role-value assignments are merged or substituted across frames, often with referential ambiguities or quasi-truth effects.

Formal repositories arising from such ontologies allow for both semantic reasoning (e.g., SPARQL retrievals to construct new metaphors) and practical applications (e.g., automated generation, referential disambiguation). The class hierarchy and frame alignment—often graphically and mathematically specified (see Eqn. D(cds,ie) in (Gangemi et al., 2018))—provide a backbone for indexing both conventional metaphorical expressions and their literal analogues.

2. Automatic Generation and Mapping Methodologies

Several approaches for constructing metaphor-literal repositories emphasize automatic generation and mapping between literal and metaphorical forms.

The Metaphoric Paraphrase Generation paradigm (Stowe et al., 2020) employs both lexical replacement (using WordNet troponyms and vector-space similarity) and neural sequence-to-sequence “metaphor masking” (transformer-based architectures trained to substitute masked literal verbs with metaphoric counterparts).
Methods such as the MERMAID framework (Chakrabarty et al., 2021) build large-scale metaphor-literal parallel corpora by combining metaphor detection (BERT-based classifiers), masked language modeling for literalization, commonsense inference to ensure symbolic parity (with COMET), and discriminative decoding to favor metaphoricity.
Controlled generation frameworks (e.g., CM-Lex and CM-BART (Stowe et al., 2021)) learn FrameNet-derived embedding transformations between conceptual domains, enabling either direct lexical shift via vector arithmetic or sequence-to-sequence generation conditioned on explicit source-target frames.
Paraphrase transfer models (Ottolina et al., 2022) generalize masking and unmasking to verbs, adjectives, and nouns, quantifying “transfer rates” (e.g., 56% for verbs, 24% for nouns) as the fraction of inputs that can be mapped to a metaphoric form by a masked LLM.

Repositories built with these techniques typically store literal-metaphor pairs or sets, annotated with alignments, mapping type (e.g., lexical, syntactic), and confidence/quality metrics from either human or automatic evaluation.

3. Dataset Design, Coverage, and Resource Unification

Modern metaphor-literal repositories aggregate resources at varying levels of granularity, genre, and linguistic coverage.

MetaphorShare (Boisson et al., 27 Nov 2024) is a dynamic, collaborative repository aggregating open metaphor datasets from multiple languages, standardized via explicit in-text tagging (<m>, <l>, <t>, <u>) and tabular formats (CSV with required columns).
Extensive datasets such as CMDAG (Shao et al., 20 Feb 2024) (Chinese metaphor corpus with 28K sentences labeled for tenor, vehicle, and ground), annotated medieval Hebrew poetry corpora (Toker et al., 27 Feb 2024), and visual metaphor resources like MetaCLUE (Akula et al., 2022) enrich the repository’s scope beyond English and beyond the verbal modality.
Annotation schemes often provide structured tuples (e.g., (TENOR, VEHICLE, GROUND)), detailed metadata (genre, source, annotator profile), and sometimes discourse-level features (abstractness, emotionality; (Piccirilli et al., 2022)).
Functionalities for upload, search, labeling, and download are paired with rigorous validation steps (automatic and manual checks, license verification), metadata curation, and search backends built on Elasticsearch and SQL.

This unified approach addresses prior issues with scattered, poorly-documented, or non-interoperable metaphor resources.

4. Computational Models: Detection, Mapping, and Retrieval

Repository construction and use depend on reliable computational models:

Metaphor detection frameworks such as MelBERT (Choi et al., 2021) and the Expectation-Realization model (Uduehi et al., 2023) use deep contextualized encodings, late interaction, and dual-branch architectures (one for the “expected” literal meaning, one for the “realized” usage in context) to learn literal–metaphoric contrasts. They exhibit strong cross-lingual and cross-domain generalization (e.g., F1 ∼ 78.5% on VUA-18 for MelBERT).
Analysis of temporal embeddings (Ottolina et al., 2021) shows that diachronic variation in word meaning may affect both literal baseline and metaphor detection accuracy, suggesting that time-sensitive repositories can support studies of semantic drift and metaphorical innovation.
Metaphor-literal mapping knowledge is partially emergent in LLMs; however, studies (Ye et al., 5 Oct 2025) reveal that current LLMs’ repository-backed mappings can induce 15–25% conceptually irrelevant outputs and are highly sensitive to syntactic cues, underlining the limitations and the need for dynamic/context-based repository integration.

5. Applications and Evaluation Protocols

Metaphor-literal repositories support a spectrum of downstream applications and research purposes:

NLP: metaphor-aware generation (e.g., creative writing, dialogue systems (Li et al., 2022, Jhamtani et al., 2021)), detection (classification, sequence labeling), paraphrase augmentation (improving classifier F1 by +3% (Ottolina et al., 2022)).
Digital humanities: large-scale metaphor annotation in cultural heritage texts (e.g., Hebrew Piyyut corpus (Toker et al., 27 Feb 2024)), literary studies (CMDAG (Shao et al., 20 Feb 2024)), and cross-modal metaphor research (visual metaphors in MetaCLUE (Akula et al., 2022)).
Semantic and discourse analysis: repositories can encode not only word-level, but also phrase-level, discourse-level, and intention-annotated metaphor information (see the taxonomy of intentions (Michelli et al., 4 Jul 2024)).
Evaluation protocols typically include both automatic metrics (BLEU, BERTScore, SBERT-based similarity, F1/Accuracy for detection tasks) and large-scale human evaluation (fluency, metaphoricity, faithfulness). Empirical studies demonstrate the importance of creative and context-sensitive outputs (e.g., metaphor masking outperforms lexical replacement in metaphoricity (Stowe et al., 2020)).

6. Limitations, Challenges, and Future Directions

Several persistent challenges are highlighted:

LLMs over-relying on static metaphor-literal repositories without appropriate context modeling, resulting in considerable irrelevant or conceptually mismatched outputs (Ye et al., 5 Oct 2025).
The need for repositories to encode diachronic, genre, and discourse variation in metaphor usage (Ottolina et al., 2021, Piccirilli et al., 2022).
The integration of structured intent (taxonomies of communicative purposes (Michelli et al., 4 Jul 2024)), chain-of-thought rationale (CMDAG (Shao et al., 20 Feb 2024)), and multimodal data (images, prosody).
Standardization efforts remain essential for interoperability; repositories such as MetaphorShare (Boisson et al., 27 Nov 2024) address this by imposing unified formats, extensible tagging schemes, and robust metadata protocols.
Evaluations on figurative language remain challenging; repositories must offer strong annotation protocols (e.g., BIO schemes, intention labels), balanced coverage of literal/metaphoric expressions, and tools for easy search/filtering (elasticsearch-backed interfaces).

Future directions involve dynamic repository refinement (incorporating user feedback, context-localized mappings), tighter integration with LLM architectures for adaptive metaphor understanding, and expansion into low-resource languages and cross-domain datasets.

The development of robust, dynamic, and linguistically informed metaphor-literal repositories underpins both empirical research and application engineering for the analysis, generation, and interpretation of figurative language across natural and artificial systems.