Holistic Figurative Language Understanding and Generation

This direction is split into other directions: syntactic scope limitation. Its last version stays readable and citable here.

Version 1 of 152 papers54 conceptsfrozen 2026-10-04

Human language is rarely literal. We "pour our hearts out," "kill processes," and describe "sweet voices." For decades, Natural Language Processing (NLP) treated these expressions as outliers or noise. However, a growing body of research suggests that figurative language—encompassing metaphors, idioms, hyperbole, and sarcasm—is not a poetic ornament but a fundamental cognitive mechanism for communicating complex ideas.

The shift in the field is moving away from treating these tropes as isolated linguistic quirks toward a unified, "holistic" understanding of non-literal meaning.

The Intuition of Semantic Friction

The traditional way to detect a metaphor is to look for "semantic friction"—the clash that occurs when words from two incompatible domains are forced together. In the sentence "the market crashed," a financial entity (the market) is paired with a physical action (crashing) typically reserved for solid objects.

Early statistical models, such as those developed by Ekaterina Shutova and Simone Teufel, leveraged this friction by using unsupervised distributional clustering. By observing that "stirring" usually applies to liquids, a model can infer that "stirring excitement" is metaphorical because "excitement" does not share the physical properties of a liquid. This approach was later refined by Yulia Tsvetkov and colleagues, who demonstrated that metaphors are fundamentally conceptual. By training models on high-level semantic features like "abstractness," they could detect metaphors across different languages (English, Spanish, Farsi) without needing language-specific training data.

Beyond Single Words to Discourse and Relations

A recurring critique in the literature is that early models were too narrow, often focusing exclusively on single-word verbal metaphors. Weicheng Ma and Soroush Vosoughi argued for a shift toward classifying the "metaphoricity" of entire sentences, noting that reducing the task to a single pre-identified word fails to capture the complexity of real-world language.

More recently, researchers have reframed the problem as relation extraction. Instead of just labeling a word as "metaphorical," the CMRE dataset introduced by Guihua Chen and Tiantian Wu forces models to explicitly link a "target" concept (e.g., life) to its "source" description (e.g., a journey). This captures the underlying cognitive mapping that makes the metaphor work. Similarly, the ConFiguRe dataset by Dawei Zhu and colleagues pushes analysis to the discourse level, recognizing that figurative meaning often unfolds across multiple sentences rather than a single verb-object pair.

The Multi-Task Synergy

One of the most significant recent insights is that different types of figurative language are "a match made in heaven" for AI models. Naveen Badathala and Pushpak Bhattacharyya demonstrated that training a single model to detect both hyperbole and metaphor simultaneously leads to massive performance gains.

This synergy suggests that figurative tropes share a common "figurative fingerprint." Gregorios A. Katsios and colleagues have even used these fingerprints for authorship attribution, proving that the specific way an individual uses creative language is a deeply personal and identifiable trait.

From Detection to Explanation and Generation

The field is now moving beyond simple "true/false" detection toward explainability and creative generation. The FLUTE dataset, developed by Tuhin Chakrabarty and Smaranda Muresan, requires models to provide natural language explanations for why a sentence is sarcastic or metaphorical. This ensures that models are not just relying on statistical shortcuts but are actually reasoning about the non-literal intent.

On the generative side, the challenge is to turn dry, literal text into creative expressions. The MERMAID framework by Chakrabarty uses "symbolism" and discriminative decoding to replace literal verbs with metaphorical ones. In the visual domain, the Abstract-to-Concrete Translator (ACT) addresses the "concreteness bias" of vision-language models by rephrasing abstract concepts (like "edgy") into concrete descriptions that models can actually "see."

Why This Matters

Understanding figurative language is the key to unlocking empathy and nuance in AI. As Gyeongeun Lee and Natalie Parde found, metaphors and hyperboles are critical signals for detecting empathy in online support communities. Without the ability to parse these non-literal layers, AI remains "metaphor blind"—struggling to distinguish a speaker's intended meaning from their literal words. By building models that can explain, interpret, and generate these tropes, we move closer to a machine intelligence that truly understands the human experience.

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_holistic_figurative_language_understandi_d06229c8,
  title = {Holistic Figurative Language Understanding and Generation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/holistic-figurative-language-understanding-and-generation/rdr_d06229c88f3d4cf38ad15f065d7ac4ee},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_holistic_figurative_language_understandi_d06229c8_v1,
  title = {Holistic Figurative Language Understanding and Generation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/holistic-figurative-language-understanding-and-generation/rdr_d06229c88f3d4cf38ad15f065d7ac4ee/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}