Disambiguating Literal and Figurative Language in Context

This direction is split into other directions: literal-figurative disambiguation. Its last version stays readable and citable here.

Version 1 of 135 papers39 conceptsfrozen 2026-10-04

When a person says they are "breaking the ice," they are rarely wielding a mallet against a frozen pond. To a human, the distinction between a literal physical action and a social metaphor is intuitive, guided by the surrounding conversation. For natural language processing systems, however, these Potentially Idiomatic Expressions (PIEs) represent a semantic minefield. Because the surface form—the sequence of words—is identical in both literal and figurative cases, models cannot rely on simple dictionary lookups. Instead, they must perform token-level disambiguation based on the surrounding discourse.

The Signal in the Noise: Lexical Cohesion

The foundational intuition behind distinguishing literal from figurative use is lexical cohesion. In a literal context, the component words of an expression tend to have strong semantic ties to the rest of the sentence. If you are literally "spilling beans," the context likely contains words like "floor," "kitchen," or "can." If you are figuratively "spilling beans" (revealing a secret), the context involves "secrets," "telling," or "surprises."

Linlin Li and Caroline Sporleder pioneered unsupervised methods to exploit this. In Unsupervised Recognition of Literal and Non-Literal Use of Idiomatic Expressions, they demonstrated that literal usages are "knowledge-lean" signals: they link more naturally to the surrounding discourse's vocabulary. By modeling these relationships as a cohesion graph, researchers can identify literal usage simply by measuring how "at home" the individual words feel in their environment.

From Hand-Coded Features to Contextual Embeddings

As the field moved toward neural architectures, the focus shifted from explicit graph modeling to capturing "sentential context" through sequence labeling. Eunsol Choi and Luke Zettlemoyer showed in Neural Metaphor Detection in Context that complex, feature-heavy models are often outperformed by simple BiLSTMs using contextualized embeddings like ELMo.

The intuition here is that modern embeddings already encode the "flavor" of a word's usage. A word like "twisting" in the context of "incredulity" (figurative) occupies a different vector space than "twisting" in the context of a "leg" (literal). This distinction is vital for tasks like literal motion detection, where Irene Manotas and colleagues highlight the difficulty of teaching models to recognize when a verb describes a physical displacement versus a metaphorical state.

Routing Interpretations via Attention

Recent work has moved toward specialized architectures that explicitly "route" a phrase toward its most likely meaning. Ziheng Zeng and Suma Bhat developed PIER (Potentially Idiomatic Expression Representation generator), which uses an attention fusion mechanism.

Instead of forcing a single representation, these models evaluate the context and decide whether to treat a phrase as a single non-compositional unit (the idiom) or a collection of literal words. This is often supported by external knowledge. For instance, Zeng also introduced IEKG, a commonsense knowledge graph that provides "if-then" reasoning for idioms. If the model sees "lick Y's boots," the graph provides the figurative implication ("insincere"), helping the model decide if that fits the current discourse better than a literal physical action.

The Frontier: Euphemisms and Multilingual Transfer

The challenge of literal/figurative disambiguation extends beyond idioms into the realm of euphemisms—polite substitutions for harsh realities. Anna Feldman has led several shared tasks (FigLang 2022, FigLang 2024) showing that while Large Language Models like GPT-4 are powerful, they still struggle with the "unseen term" problem.

To combat this, researchers are experimenting with multi-modal and multi-task signals. Aykut Erdem and İlker Kesen found that providing models with visual imageries of literal descriptions helps them "see" the absurdity of a literal interpretation in a figurative context. Meanwhile, Debora Ciminari and Uliana Sentsova are pushing these capabilities into multilingual contexts, demonstrating that while English-trained models can transfer some "figurative logic" to other languages, specialized monolingual training remains essential for capturing the unique cultural nuances of how different societies "soften" or "color" their language.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_disambiguating_literal_and_figurative_la_927cbb57,
  title = {Disambiguating Literal and Figurative Language in Context},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/disambiguating-literal-and-figurative-language-in-context/rdr_927cbb574778402a8ea0804c34346da3},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_disambiguating_literal_and_figurative_la_927cbb57_v1,
  title = {Disambiguating Literal and Figurative Language in Context},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/disambiguating-literal-and-figurative-language-in-context/rdr_927cbb574778402a8ea0804c34346da3/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}