Research Desk · Paris

Languages,
Metaphor & AI

Editor’s Guide

This is not quite an academic CV.

People often wonder what PhD students actually research. Philosopher Matt Might offers a useful image: if the sum of human knowledge were a perfect circle, doctoral research would mean spending years tapping away at one infinitesimal point on its edge. Eventually, perhaps, the boundary gives way—just enough to push it outward by a fraction.

What led me to thread the needle was idioms. Idioms gather the shared intelligence of language communities. In a sense, they work like encrypted language: their literal wording and intended meaning do not always point in the same direction.

Whether we translate, write, learn a foreign language, or speak one natively, we all know the frustration of searching for a word that truly fits. Finding an apt word is difficult enough; finding an apt idiom can feel like searching for a needle in a haystack.

When we do not already know an idiom, its split between literal and figurative meaning makes it especially difficult to find. This is why a semantically organised dictionary of idioms is needed: one that helps people search by what they mean, not only by the words they happen to remember.

A tiny explorer taps the edge of a large charcoal circle with a needle.

Research projects

01

PhD Project

A model for a bilingual dictionary of French idioms in economic discourse

Designed for writers, translators, and language learners.Explore the dictionary ↗

The questions

  1. How can people find an idiom when they know what they want to express, but not the words they need?
  2. Existing dictionaries often provide few authentic examples and little information about how frequently idioms occur across different types of discourse. How can a dictionary show users not only what an idiom means, but how it is actually used?
  3. Since no specialised dictionary of this kind currently exists, what workflow is needed to build one?

How it works

  1. The prototype organises idioms according to meaning, allowing users to search by semantic theme, constituent words, and commonly co-occurring words.
  2. I compiled and analysed a 2.2-million-word corpus of authentic economic discourse, drawing on podcasts, newspapers, magazines, and OECD reports. The corpus provides evidence of how frequently individual idioms occur and supplies authentic examples of their use.
  3. After developing the prototype, I designed a reusable workflow for producing specialised idiom dictionaries in fields such as politics and science. It covers subcorpus selection, idiom identification and annotation, semantic analysis, translation principles, and the lexicographic functions required to turn research into a usable dictionary.

Why it matters

  1. Even in the age of large language models, idioms remain difficult to identify and translate reliably.
  2. LLMs cannot reliably establish whether a usage example is authentic or provide corpus-based information about an idiom’s frequency across different types of discourse.
  3. A structured, corpus-based dictionary can therefore support writers, translators, and language learners while also serving as a verified linguistic resource for improving AI-generated responses.
02

From Words to Worlds

A study of metaphor, emotion, and large language models

Funded by the Cosmos Institute.Explore the project ↗

The question

How does the metaphorical expression of emotion differ between human languages and language generated by large language models?

How it works

The study compares emotional idioms in English, French, and Chinese with metaphors generated by four LLMs. It examines their source domains, conceptual mappings, and recurring figurative patterns.

Why it matters

Human idioms tend to ground emotion in the body, culture, and lived experience. LLM-generated metaphors often rely instead on external scenery, electromechanical imagery, or language that is stylistically vivid but conceptually weak. If AI-generated content becomes a dominant form of textual production, these recurring patterns could gradually reshape metaphorical expression and reduce linguistic diversity.
Read the preliminary conclusion ↓

Preliminary conclusion · Excerpt from an unpublished manuscript

This preliminary study examines how emotional meaning is metaphorically encoded in human languages and how these patterns compare with metaphorical language generated by large language models. By comparing English, French, and Chinese idioms with outputs from four LLMs, we identified a systematic divergence in their underlying conceptualisations.Human idioms remain deeply rooted in embodied experience, particularly in physiological sensations, facial expressions, and behaviour. By contrast, LLM-generated metaphors display distinct figurative strategies at the level of linguistic output.First, emotional states are frequently represented through external, non-bodily source domains, including natural landscapes and environmental conditions. Second, LLM outputs show recurring preferences for physical and electromechanical imagery—such as tension, vibration, electricity, and mechanical resistance. Although some of these images are attested in English, they remain marginal and are far less systematically clustered in human idiomatic usage.Third, LLM-generated language often relies on scenic or atmospheric descriptions that evoke an emotional mood without establishing a motivated conceptual correspondence between source and target domains. From a cognitive-linguistic perspective, this final strategy may be understood as a failure of metaphor: the expression is stylistically vivid yet conceptually empty, lacking the systematic mappings characteristic of genuine metaphor.These findings do not suggest that LLMs are incapable of producing emotionally appropriate language. Rather, they indicate that the figurative organisation of LLM-generated metaphors differs systematically from that of human languages. While LLM outputs may appear plausible and rhetorically rich, they often depart from the embodied and culturally stabilised mappings that underpin human idiomatic competence, producing language that is stylistically convincing but conceptually unmotivated.Beyond their descriptive implications, these findings raise broader concerns about linguistic diversity. If AI-generated content becomes a dominant source of textual production, recurring LLM-specific figurative patterns may gradually reshape metaphorical usage, favouring generalised and decontextualised imagery over culturally specific and embodied expression.

Publications & working papers

Publication · Langages, no. 236 · 2024

Does AI Possess Idiomatic Competence?

A study of AI translation of French fixed expressions into English and Chinese.

Read on HAL ↗Publication

La traduction des expressions figées dans un contexte lexicographique bilingue

Étude de cas de trois dictionnaires généraux français-chinois.

Read on ResearchGate ↗