PaperlayerView exact revision on arXiv

Source-linked AI summary

Meanings are like Onions: a Layered Approach to Metaphor Processing

Silvia Cappa, Anna Sofia Lippolis, Stefano Zoia

arXiv:2507.10354v2cs.CL

TL;DR

Computational metaphor analysis still struggles with context, pragmatics, and distinguishing relational mappings from associations. The paper proposes a layered framework combining contextual mapping, conceptual combination, and pragmatic intention; its theoretical model organizes both what metaphors mean and what they do.

  • Problem

    Computational metaphor analysis faces scarce data, contextual insensitivity, inconsistent evaluation, theoretical fragmentation, and difficulty distinguishing deep relational mappings from associations.

    Source: §1 · pp. 1–2

  • Method

    The paper proposes a layered framework joining contextual mapping and conceptual operations with pragmatic intention and communicative effects.

    Source: §1 · p. 2, §3.1.3 · p. 9

  • Results

    The theoretical framework organizes conceptual and pragmatic implicit knowledge to represent both what metaphors mean and what they do.

    Source: §1 · p. 2, §5 · p. 11

  • Takeaways & Limitations

    The proposal calls for shared research vocabulary, datasets, and formalization of pragmatic factors to support future computational experiments.

    Source: §5 · p. 12

    The operational framework is currently theoretical, and the lack of datasets hinders computational implementation of its pragmatic layer.

    Source: §5 · p. 11, §4 · p. 11

Abstract

from arXiv · show

Metaphorical meaning is not a flat mapping between concepts, but a complex cognitive phenomenon that integrates multiple levels of interpretation. In this paper, we propose a stratified model of metaphor processing that treats meaning as an onion: a multi-layered structure comprising (1) content analysis, (2) conceptual blending, and (3) pragmatic intentionality. This three-dimensional framework allows for a richer and more cognitively grounded approach to metaphor interpretation in computational systems. At the first level, metaphors are annotated through basic conceptual elements. At the second level, we model conceptual combinations, linking components to emergent meanings. Finally, at the third level, we introduce a pragmatic vocabulary to capture speaker intent, communicative function, and contextual effects, aligning metaphor understanding with pragmatic theories. By unifying these layers into a single formal framework, our model lays the groundwork for computational methods capable of representing metaphorical meaning beyond surface associations, toward deeper, more context-sensitive reasoning.

1. Introduction

The introduction frames computational metaphor analysis as challenged by limited data, context and pragmatic gaps, and difficulty distinguishing deep relations from associations. It proposes a layered framework joining conceptual interpretation with pragmatic intent.

  • Computational metaphor analysis faces scarce, mapping-limited data, contextual insensitivity, inconsistent evaluation, theoretical fragmentation, and LLM difficulty distinguishing relational mappings from associations.Source: §1 · pp. 1–2
  • The framework separates explicit surface meaning from implicit conceptual and pragmatic meaning, which are inferred through contextualization and implicature.Source: §1 · p. 2
  • The proposal layers contextual mapping, analogical conceptual operations, and pragmatic intention, joining CMT with SAT-inspired pragmatics to model what metaphors map and accomplish.Source: §1 · p. 2
  • The authors present the unified model as theoretical groundwork for richer datasets, standardized evaluation, and computational systems that better mirror how humans unlock implicit knowledge through metaphor.Source: §1 · p. 2

2. Related Work

The related-work section surveys cognitive theories, computational representations and datasets, then motivates integrating contextual and pragmatic dimensions into metaphor processing.

2.1. Metaphor theories

  • CMT maps source to target through systematic correspondences, while CBT blends selected elements from input domains; metaphor resistance can reflect weak explanatory power or preference for alternatives.Source: §2.1 · p. 3
  • Accounts of metaphor also emphasize shared frames, social interaction, and embodied experience, while noting limited attention to context and collaborative production.Source: §2.1 · p. 3

2.2. Metaphor representation

  • Existing ontologies encode blending spaces and mappings, but rule-driven approaches can lack scalability and tacit, context-dependent knowledge for open-ended or multimodal interpretation.Source: §2.2 · pp. 3–4

2.3. Metaphor datasets and benchmarks

  • The VUA corpus marks metaphorical tokens without source and target domains, while domain-annotated resources include Gordon et al. and the medical-domain BCMTD [25].Source: §2.3 · p. 4

2.4. Theory-driven computational processing of metaphor

  • Theory-guided detection incorporates CMT/CBT constraints, while interpretation methods often rely on single-word annotations or predefined targets; METCL operationalizes CBT.Source: §2.4 · p. 4

2.5. Metaphors as speech acts

  • Computational pragmatics only partially addresses why metaphors are used, in what contexts, and with what effects, despite speech-act theory framing utterances as actions with consequences.Source: §2.5 · p. 5
  • Illocutionary force is context-dependent and may blend intentions; treating intention as a feature of linguistic acts avoids presupposing mental states, while metaphorical speech acts can affect listeners.Source: §2.5 · pp. 5–6

3. How meaning is an onion: a proposal of three-dimensional meaning representation

The proposal frames metaphor as both a cognitive tool and a speech act, offering a layered vocabulary intended to ground future computational implementations.

3.1. The layered perspective

  • The framework treats metaphor as both a cognitive tool and a speech act, providing a vocabulary for future computational implementations.It presents metaphor as a three-dimensional meaning framework.Source: §3.1 · p. 6

3.1.1. Layer 1: The external context layer

  • Layer 1 records communicative-object context and annotator metadata, including domain, provenance, frame links, and interpreters’ cultural and demographic backgrounds.For the anti-smoking image, the example includes public-health domain, campaign provenance, and links to established metaphor resources.Source: §3.1.1 · pp. 6–7

3.1.2. Layer 2: The conceptual combination layer

  • Layer 2 represents source and target domains, their blending principle, and the emergent blend; the anti-smoking example connects cigarettes and bullets through lethality.The framework describes these components using the Blending Ontology.Source: §3.1.2 · p. 7
  • METCL provides a prototype-based combination mechanism whose generated blends were generally accepted by human judges and are proposed as a knowledge-graph completion tool.Its pipeline structures source and target, builds prototypes from ConceptNet features, and combines them into an abstract metaphor representation.Source: §3.1.2 · p. 8

3.1.3. Layer 3: The pragmatic layer

  • Layer 3 annotates evaluative attitude, illocutionary force and directive strength, along with psychological effects and efficacy relative to intention and context.Directive strength ranges from requests and suggestions to commands, prohibitions, or pleas.Source: §3.1.3 · p. 9

3.1.3. The pragmatic layer

  • For the anti-smoking image, the analysis infers a prohibitive directive discouraging smoking and possible concern, discomfort, or shock in viewers.The proposal focuses on annotating forces perceived as primary.Source: §3.1.3 · pp. 9–10

4. Implications for computational metaphor processing and future work

The section translates the layered proposal into research needs and computational applications. It calls for pragmatic-aware datasets and representations while outlining uses in metaphor generation, understanding, and socially relevant domains.

Datasets

  • Metaphor datasets should include conversational and multimodal data, pragmatic dimensions, perlocutionary effects, and meaning as a spectrum.The authors call for diverse annotators to help create fine-grained pragmatic gold standards.Source: §4 · p. 10

Knowledge representation

  • Metaphor representation should integrate literal, conceptual, and pragmatic levels, including discourse context, communicative intent, and emotional effects.The authors propose stratified ontologies that can track reinterpretation and resistance in conversation, informed by community norms and individual differences.Source: §4 · p. 10

Possible application of the framework in computational metaphor processing

  • LAG and METCL could complement one another by sharing RDF representations and combining their respective frame and prototypical-property resources.The proposed integration could use LAG to weight RDF triples with METCL property lists, while METCL could use LAG to identify missing frames or align features with ontological categories.Source: §4 · p. 11
  • Pragmatic-layer implementation is hindered by a lack of datasets, although LLM-based approaches such as LAG could extract pragmatic features for analysis.Source: §4 · p. 11
  • Layered metaphor generation could improve conceptual coherence and pragmatic fit, while metaphor-aware recommenders could map figurative queries to concrete product attributes.The proposed recommender use aims to match affective intent while keeping system reasoning transparent.Source: §4 · p. 11
  • Combining conceptual mappings with pragmatic implications could support metaphor interpretation, detection, paraphrasing, and explanation, especially in dialogue, narratives, and scientific texts.The described neurosymbolic approach parses surface form, maps conceptual domains, and applies pragmatic filters for speaker goals, emotional tone, and discourse situation.Source: §4 · p. 11

Metaphors for social good

  • Metaphor-aware tools could be used in education, science communication, doctor-patient interactions, and hate-speech contexts, where metaphors carry social, cultural, and emotional weight.Source: §4 · p. 11

5. Limitations and conclusion

The paper presents a theoretical three-layered framework joining conceptual and pragmatic implicit knowledge, intended as a basis for future empirical work. Its current limitation is the lack of computational implementation, alongside the need for shared theory, vocabulary, and formalization.

5. Limitations and conclusion

  • The proposed three-layered framework brings conceptual and pragmatic implicit knowledge together to represent both what metaphors mean and what they do.The authors position it as a direction for future research and a foundation for empirical work, datasets, evaluation metrics, and neurosymbolic methods.Source: §5 · p. 11
  • The framework is currently theoretical and lacks computational implementation.Source: §5 · p. 11
  • Further work should integrate CBT and pragmatics, develop shared dataset vocabulary and formalize pragmatic factors, then extend and test the framework across modalities and domains.Source: §5 · pp. 11–12

Declaration on generative AI

The authors used GPT-4 and Grammarly for grammar and spelling checks, then reviewed and edited the content and retained responsibility for it.

Declaration on generative AI

  • GPT-4 and Grammarly were used for grammar and spelling checks; the authors reviewed and edited the content and take full responsibility for it.Source: Declaration on generative AI · p. 12

Related papers

Loading 2507.10354v2…