The Challenge of Processing Novel Metaphors in Language Models

22 concepts10 papers1 proposal

When we encounter the phrase "spending time," we rarely pause to picture a wallet or a transaction. The metaphor is so deeply embedded in everyday language that it has effectively fossilized into a standard word sense. However, if a poet writes about "unzipping a shadow," our brains immediately engage in active, cross-domain mapping to construct a brand-new conceptual bridge.

In natural language processing, this distinction creates a profound challenge known as the conventionality bias. While modern neural language models appear highly proficient at identifying metaphorical language, evidence suggests they are not performing abstract analogical reasoning. Instead, they rely heavily on frequency-based pattern memorization, excelling at highly conventionalized metaphors while failing catastrophically when confronted with novel, creative, or idiosyncratic expressions.

The Two Axes of Metaphorical Language

To understand why models struggle, we must decouple how frequently a word is used metaphorically from how creative that usage actually is. As Natalie Parde et al. demonstrate in Exploring the Terrain of Metaphor Novelty: A Regression-Based Approach for Automatically Scoring Metaphors, metaphoricity and novelty are independent dimensions. A phrase can be explicitly metaphorical yet entirely lack creative novelty due to linguistic conventionalization.

Early computational lexicons, such as the one introduced by Wakako Kuwahata et al. in Senses of Polysemous Nouns: Building a Computational Lexicon of Basic Japanese Nouns, recognized the necessity of explicitly defining these hierarchical sense extensions. Without distinguishing between established extensions and novel figurative usages, a system cannot determine whether an expression represents a standard linguistic rule or a creative departure.

The Mechanism of Conventionality Bias

Why do state-of-the-art neural models fail when metaphors become creative? The answer lies in their training objective: predicting words based on statistical co-occurrences.

When a model encounters a conventional metaphor, it can rely on learned word senses and high-frequency patterns. In An analysis of language models for metaphor recognition, Arthur Neidlein et al. evaluated language models on their ability to distinguish conventional versus novel expressions. The authors discovered a stark performance disparity: models excel at identifying the conventional instances but struggle to generalize to less common or newly coined figurative language. This suggests that high accuracy scores on standard benchmarks mask a reliance on simple word-sense disambiguation rather than a deep comprehension of abstract metaphorical principles.

This limitation becomes highly apparent in downstream reasoning tasks. Tuhin Chakrabarty et al. tested models on pragmatic reasoning in Figurative Language in Recognizing Textual Entailment. They found that while models easily resolve common idioms, novel comparisons frequently trigger reasoning failures. The models tend to fall back on memorized, frequency-based associations instead of executing true semantic logic. As summarized by Xiaoyu Tong et al. in Recent advances in neural metaphor processing: A linguistic, cognitive and social perspective, current architectures lack the underlying cognitive mechanisms required for active cross-domain mapping, leaving them tethered to the surface statistics of the training data.

Formalizing and Annotating Novelty

To fix this bias, researchers must first precisely define what makes a metaphor "novel." Traditional approaches often rely on subjective crowdsourced scores, which can be noisy and inconsistent.

To address this, Sebastian Reimann et al. proposed a more robust framework in When is a Metaphor Actually Novel? Annotating Metaphor Novelty in the Context of Automatic Metaphor Detection. By systematically comparing a word’s contextual usage against its established dictionary definitions, the authors isolated true semantic novelty. Their experiments confirmed that when novelty is strictly defined by dictionary deviation, current NLP models exhibit a severe performance gap, further complicated by subword tokenization and word frequency issues.

Another facet of novelty is intentionality. In Guidelines for the Annotation of Deliberate Linguistic Metaphor, Stefanie Dipper et al. focused on "deliberate" metaphors—expressions consciously chosen by a speaker to spark creative comparison. These are often signaled by "markedness," which manifests as unusual grammatical structures, typographical emphasis, or highly unexpected domain mixing. Capturing these markers allows researchers to isolate deliberate, creative language from rote, fossilized phrases.

Why the Novelty Gap Matters

Failing to distinguish between conventional and novel metaphors is not just a theoretical limitation; it restricts the utility of NLP models in specialized domains where creative language carries high diagnostic or stylistic weight.

In sociolinguistics, novel metaphors serve as crucial stylistic signals. Markus Egg et al. investigated this in A corpus of metaphors as register markers, demonstrating that non-conventional and extended metaphors are not distributed randomly. Instead, they function as deliberate register markers that signal participation in specific discourse styles, such as persuasive political speeches or religious sermons.

Even more critically, the structural nature of a metaphor can serve as a clinical indicator. In Using Automated Metaphor Identification to Aid in Detection and Prediction of First-Episode Schizophrenia, E. Dario Gutiérrez et al. applied machine learning to interview transcripts of at-risk individuals. They noted that the critical diagnostic signal for predicting a transition to psychosis lies not in the sheer quantity of metaphorical language, but in the specific presence of idiosyncratic or deviant metaphors. A model blinded by conventionality bias would fail to separate these highly predictive, deviant linguistic disturbances from benign, standard metaphors.

Toward Dynamic Representations

To move beyond static word-sense lookup, recent work attempts to model metaphors as evolving, dynamic processes within discourse. In Broadening the coverage of computational representations of metaphor through Dynamic Metaphor Theory, Xiaojuan Tan et al. integrated the YinYang Dynamics of Metaphoricity (YYDM) model into a frame-based computational approach.

Instead of treating a metaphor as a fixed label, the authors formalize explicit "metaphoricity cues"—such as repetition, novelization, and domain mixing—to track how a metaphor becomes active, evolves, and shifts meaning throughout a text. By capturing these dynamic interactions, such frameworks aim to bridge the gap between cognitive linguistics and machine learning, pushing models away from rote memorization and toward true, adaptive semantic reasoning.

Key Foundational Work

Senses of Polysemous Nouns: Building a Computational Lexicon of Basic Japanese Nouns (1996) — contributed 1 concept elements to this direction.
Using Automated Metaphor Identification to Aid in Detection and Prediction of First-Episode Schizophrenia (2017) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_conventionality_bias_f04ef7f5,
  title = {conventionality bias},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/conventionality-bias/rdr_f04ef7f56fc84ac7b3f4befe02a8638c},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_conventionality_bias_f04ef7f5_v1,
  title = {conventionality bias},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/conventionality-bias/rdr_f04ef7f56fc84ac7b3f4befe02a8638c/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake