FROM EXTRACTION TO HUMAN DECISION
A local model proposes possible metaphors. A second model checks each phrase. A ranker brings the most promising expressions forward. People make the final call.
The software orders attention; it does not decide meaning. Ranking helps a finite review effort reach better candidates sooner. Reviewers still decide what belongs.
HOW IT WORKS
One example: in “Cancer arrived as an unwelcome lodger,” cancer is what is being described and a difficult house guest is the borrowed image. The search looks for this illness-to-image relation—not metaphorical words in general.
“Find every metaphorical expression in this passage, including everyday ones. Return the exact phrase and the category of the borrowed image.”
Why this may help: illness is often implied rather than named. Broad prompting is intended to reduce missed expressions. The older narrow route also filtered passages before prompting, so the comparison does not isolate prompt wording.
“Is this exact expression used metaphorically in this passage?”
Why this may help: a different model family independently checks each phrase proposed by the first model. Only verified expressions move forward.
“Is it about illness?”
“Is it about lived experience?”
“How Menu-like is it?”
Why this helps: the answers remain visible instead of becoming one opaque score. They set the order; image categories reveal recurring patterns. People still decide what belongs.
A small feasibility check: across nine previously identified passages from five Danish interviews, the broad route surfaced all nine before verification. Second-model verification retained seven while removing about two thirds of all suggestions. This does not estimate recall, precision or Menu suitability. All private-text models run on project hardware.
FOR REVIEWERS
Reviewers in Denmark and the Netherlands get a small .json file by e-mail. The review page opens it in the browser, in their language; nothing is uploaded and the labels are exported as a small file to send back.
THE REVIEW WORKFLOW
Stages one and two conceal the rank on purpose. Only after the independent decisions are complete does the project team see the system's ordering.
Review a shuffled sample from across the list. The hidden rank prevents the software's judgement from influencing the reviewers.
Open blinded filtering 02 PPI panelDecide independently which expressions belong on a menu for people living with illness. Rank and score remain hidden.
Open panel voting 03 Project teamSee every candidate in rank order, explore the source domains, and check where retained and known-good expressions surfaced.
View the project outputWHY RANKING MATTERS
Published Metaphor Menu entries were planted as a positive control. Most landed in the top tenth of the full ranking.
See the recovery check Method 2 routesA simple comparison-word search misses most relevant expressions. The two-model workflow searches more broadly, then ranks what people review first.
Compare the approachesA SAFE PUBLIC VIEW
The visible examples come from English entries in #ReframeCovid: a crowdsourced, multilingual collection of alternatives to war language and other ways of framing COVID-19. It lets us show the workflow with text we may redistribute; it is not a patient cohort and does not validate a cancer Menu. The extraction, checks, categories and ranking were added by this project—not by the original dataset publishers.
Open the original #ReframeCovid collectionThis 415-item view contains 368 #ReframeCovid expressions and 47 phrases extracted from published Metaphor Menu check items. Reddit benchmark text is not reproduced.