How I Extracted Metaphors from French Poetry (Without Handpicking Them)

A pre-LLM NLP approach using syntactic patterns, poetry corpora, and expert annotation

Most experimental studies on metaphor rely on a familiar trick: researchers construct sentences that are either metaphorical or literal, carefully matching them word by word.

This approach is convenient. It’s also deeply artificial.

During my PhD, I wanted to study metaphor processing using ecological language — metaphors as they actually appear in texts people read for meaning and beauty, not metaphors engineered in a lab. That led me to a deceptively simple question:

How can we select metaphors automatically, at scale, while keeping control over linguistic form?

This article explains the method I developed to do exactly that, using French poetry, syntactic patterns, and expert validation — without handpicking examples or writing “literal controls” by hand.

This work builds on earlier peer-reviewed research where I used eye-tracking to study how formal features of French poetry (such as visual layout and rhyme) shape real-time reading behavior, even when textual content remains unchanged.

Why ecological metaphors are hard to study

Metaphor research has a long tradition of norming metaphorical expressions on dimensions like familiarity, imageability, comprehensibility, or metaphoricity (e.g. Katz et al., 1988; Cardillo et al., 2010; Jacobs & Kinder, 2017).

The problem is that most datasets rely on:

constructed metaphors paired with constructed literal paraphrases

This introduces several issues:

• Literal “controls” often sound unnatural
• Register and style subtly change between conditions
• Syntactic structure is rarely identical
• The researcher’s intuition heavily shapes the material

In other words, we gain experimental control but lose ecological validity.

Poetry, on the other hand, is full of metaphors — but it’s messy. You can’t just scrape a poetry corpus and label everything metaphorical. You need a way to constrain the search without imposing meaning by hand.

The core idea: control syntax, not meaning

Instead of trying to define metaphors semantically, I flipped the problem.

I asked:

Can we select candidate metaphors automatically by constraining their syntactic form, and let meaning emerge from human judgments?

The intuition was simple:

• Certain syntactic constructions are highly productive in metaphorical language
• Those same constructions also appear in non-metaphorical uses
• If we control syntax, we can compare metaphorical vs non-metaphorical expressions without rewriting them

Step 1: Start with an ecological corpus

I worked with a corpus of French poetry, chosen deliberately:

• rich in figurative language
• stylistically diverse
• far from everyday “lab language”

The goal was not to model poetry per se, but to use it as a high-density source of metaphors.

The corpus was:

• POS-tagged
• syntactically parsed

This allowed systematic pattern extraction rather than manual reading.

Step 2: Extract candidate phrases using syntactic patterns

I focused on three simple but powerful constructions:

• Noun de Noun la prison du silence (“the prison of silence”)
• Noun “être” Noun le temps est un fleuve (“time is a river”)
• Adjective Noun une lumière cruelle (“a cruel light”)

These patterns are:

• frequent in poetry
• compact (short phrases)
• compatible with both literal and metaphorical interpretations

Using parsing + POS tagging, I automatically extracted thousands of candidate phrases matching these templates.

At this stage, nothing was labeled metaphorical yet.

Step 3: Random sampling to avoid cherry-picking

From the extracted candidates, I randomly sampled 1,000 phrases.

This step was crucial:

• no aesthetic preselection
• no “nice metaphors only”
• no intuition-driven filtering

Random sampling ensured that the dataset reflected the actual distribution of language in the corpus — including awkward, banal, or ambiguous phrases.

Step 4: Expert validation (meaning, not form)

To determine which phrases were metaphorical, I relied on expert judgments.

• 4 independent experts evaluated each phrase
• Each item was labeled as metaphorical or non-metaphorical
• Only items with high agreement (3/4 or 4/4) were retained

This resulted in:

• 60 metaphors
• 60 non-metaphors

All sharing the same syntactic patterns, all drawn from the same poetic sources.

At no point did we rewrite, paraphrase, or “correct” the phrases.

Why this matters methodologically

This approach has several advantages:

1. Ecological validity

All items come from real texts, written for expressive purposes — not experimental convenience.

2. Controlled form

By holding syntax constant, we reduce confounds related to sentence structure, length, or grammatical complexity.

3. Meaning emerges from data

Metaphoricity is not imposed by the researcher, but emerges from expert and participant judgments.

4. Scalability

The method is semi-automatic and reproducible. With different corpora, it can be adapted to:

• other genres
• other languages
• other figurative constructions

What comes next (beyond extraction)

These extracted metaphors were later used in large-scale online studies (around 1,500 participants) where people rated each phrase on:

• comprehension
• familiarity
• beauty
• metaphoricity

Because the material was ecologically grounded, the results revealed patterns that are often blurred in constructed stimuli — particularly the close relationship between aesthetic judgment and metaphoricity.

That part deserves its own article.

Why I still care about this approach today

In the age of large language models, it’s tempting to think that manual corpus work is obsolete.

I think the opposite is true.

LLMs are powerful because they absorb massive amounts of human language — but they don’t tell us why certain expressions feel metaphorical, beautiful, or striking. For that, we still need:

• carefully designed stimuli
• interpretable features
• human judgments
• and methods that respect how language is actually used

This metaphor extraction pipeline was my attempt to do exactly that.

If you’re interested in experimental NLP, psycholinguistics, or the intersection of language, meaning, and aesthetics, I’ll be sharing more work from this project — including large-scale behavioral results and modeling insights.