What Metaphors Leave Unsaid: Implied Inferences and Ideological Persuasion

When an editorial calls a policy a "storm" or a reform "a long road," the persuasive work is done by inferences the writer never states: storms are no one's fault and must be endured, roads have destinations and reward patience. This proposal treats those unstated inferences as the thing to measure. It tests whether they, and not the number or category of metaphors, explain why the same editorial changes the minds of liberal readers but leaves conservative readers unmoved, or the reverse. The link to reporting bias, the well-documented tendency of text to omit what is obvious to its readers, is direct. A metaphor's entailments are exactly the kind of content authors leave out, because spelling them out would weaken the rhetoric. Models trained on text therefore see the metaphor without its consequences.

The concrete claim is this. The reported finding that conservatives are "resistant" to metaphor is at least partly a measurement artifact of counting surface metaphors. If each metaphor is expanded into the specific, human-validated inferences it licenses about the political target, and those inferences are scored for which moral concerns they engage, the expanded representation should predict ideology-specific persuasion where metaphor counts do not. Building that expansion requires confronting reporting bias twice. The inferences are absent from the text. And language-model generators of those inferences are known to over-produce salient, dramatic content, so their output must be checked against what actual readers infer.

Grounding in prior work

The motivating evidence comes from Analyzing the Use of Metaphors in News Editorials for Political Framing. That paper annotated 12,006 metaphors in 300 New York Times editorials from Webis-News-Editorials18, with 14 source and target categories. It found medium effect sizes for liberals ($r \approx 0.3$–$0.35$) between metaphor counts and persuasive effect (Challenging, Reinforcing, No Effect). For conservatives it found no significant difference, although conservatives did respond to Journey sources and resisted Power & Control targets. Their analysis stops at counts and category labels. We keep their corpus and labels and change the representation of each metaphor.

There is also an apparent contradiction to resolve. When People are Floods detects seven dehumanizing source domains in 400,000 immigration tweets and finds that conservative ideology goes with 10.4% to 17.6% more metaphor use. The Pragmatics behind Politics multi-task trains RoBERTa with a metaphor-detection auxiliary task and finds the largest gains on conservative perspectives. Conservatives using and being signalled by metaphor while being "unmoved" by it is easier to reconcile if what matters is what the metaphor implies rather than how often metaphors appear. Both papers represent metaphor as a token label or domain tag. Neither looks at the inferences being transferred.

On the reporting-bias side, Do Neural Language Models Overcome Reporting Bias? shows that language models diverge from real-world frequencies more than raw n-grams do (KL 2.94 for n-grams versus 3.08 for RoBERTa and 3.77 for BERT/GPT-2) and favor dramatic events. This constrains our design: inferences generated by a model cannot be trusted as a proxy for reader inferences without checking. Two generation methods supply our technique. COMET showed that generators can externalize unwritten commonsense, with 77.5% of its ATOMIC generations judged plausible. Self-Talk elicits clarifications from the model itself. It also reports that 24.7% of these were irrelevant and 19.9% nonsensical even when downstream accuracy rose, which is why we human-validate every elicited inference rather than trusting downstream gains.

MiQA is the closest work joining metaphor and commonsense inference. It tests large models on 100 primary sensorimotor metaphors and finds that the large models score near human level. It has no political content, no ideology, and no persuasion outcome. Our object is contested policy metaphors, whose inferences may differ by reader. EMONA supplies the moral-foundation labeling scheme we use to score inferences. It also shows that 57% of morally loaded events were missed as "objective" by earlier opinion annotation, which is further evidence that moral force often sits in what is implied.

Research Program

The pipeline has three stages.

The first stage is inference elicitation. For each annotated metaphor in the editorial corpus, an instruction-tuned language model receives the sentence, the editorial's stance, and the metaphor span. It generates up to five statements of the form "if the target is like the source, then …", which COMET-style prompting externalizes as text.

The second stage is reader elicitation and validation. For a stratified subset of about 1,500 metaphors covering all 14 source categories, crowdworkers who self-identify as liberal or conservative perform two tasks. They write what the metaphor implies, without seeing model output. They also rate the model's inferences for plausibility. This produces three inference sets per metaphor: model-generated, liberal-generated, and conservative-generated. These sets let us measure the reporting-bias distortion directly, as the share of model inferences that neither reader group produces and the share of reader inferences the model misses. We also test whether model inferences skew toward the dramatic, using a "sensational versus mundane" rating that mirrors the reporting-bias literature.

The third stage is aggregation. Each validated inference is labeled with moral foundations using an EMONA-trained classifier, with the polarity of each foundation determined relative to the editorial's stance. Each editorial then becomes an "implied moral profile": counts of stance-supporting and stance-opposing inferences per foundation, weighted by plausibility. This profile, rather than metaphor counts, is the input to ordinal models predicting the Challenging/Reinforcing/No Effect label separately for liberal and conservative readers.

Baselines

The naive reference is the majority persuasive-effect label for each ideology. Because the classes are imbalanced, every model must beat it on macro-F1 and not just accuracy.

The first serious baseline restates the editorial paper's analysis as a predictor. Its features are metaphor counts plus counts in the 14 source and 14 target categories, fed to the same ordinal regression with the same cross-validation folds. It represents the existing explanation. Its reported evidence is $r \approx 0.3$–$0.35$ for liberals and no significant effect for conservatives (Analyzing the Use of Metaphors in News Editorials for Political Framing). No predictive accuracy for persuasive effect on this corpus is reported. Our first experiment establishes that number.

The strongest existing learned method is RoBERTa with a metaphor-detection auxiliary task as in The Pragmatics behind Politics, fine-tuned end to end on full editorials. Its reported results are on other tasks: .805 accuracy on political affiliation and .860 on political perspective with the emotion auxiliary. No reported result was found on Lacuna for this model on persuasive-effect prediction. We run it on identical folds with the same hyperparameter search budget as our models. We also add a variant that receives our inference text appended to the editorial, so that any advantage cannot come from the extra text alone.

The third baseline is zero-shot and few-shot prompting of the same model used for elicitation, given a liberal or conservative reader persona and asked for the persuasive effect directly. For context, frontier models reached only about 0.33 macro-F1 zero-shot on narrative framing in Narrative Media Framing in Political Discourse, though that is a different task.

Two ablations isolate the key ingredient. The literal-paraphrase ablation rewrites each metaphorical sentence literally while preserving denotation, using the entailment-filtered rewriting of ENTRUST. We then elicit inferences from the paraphrase. If this does as well, the signal is the editorial's content, not what the metaphor licenses. The source ablation compares model-elicited profiles against reader-elicited profiles on the human subset, which measures what reporting bias in the generator costs.

Beating the baselines means a statistically reliable improvement in conservative-reader macro-F1 over both the count/category model and fine-tuned RoBERTa, using paired bootstrap over folds, together with a significant likelihood-ratio gain from the inference features once the count features are included.

Why not just…?

Why not just fine-tune a classifier on the editorials and let it learn the implications? This is the real threat. A model fine-tuned on 300 documents could in principle absorb implied content. With 300 labeled editorials and three-way labels for each ideology, though, end-to-end models have little data from which to learn which unstated inferences matter. The RoBERTa baseline and its inference-augmented variant settle the question. If the plain fine-tuned model matches our profile model, we conclude that explicit expansion adds nothing and report that.

Why not just ask a frontier model how a conservative reader would react? Persona prompting skips the representation entirely. But the evidence on related tasks is poor. Frontier models fall back on stereotypes in narrative framing (Macro-F1 about 0.33 in Narrative Media Framing in Political Discourse) and accept reversed causal links about 50% of the time in polarized climate discourse (Assessing LLM Reasoning Through Implicit Causal Chain Discovery). We run persona prompting as a baseline. If it wins, the reporting-bias concern about model inferences is weaker than we assume.

Why not just use moral-foundation dictionaries or the existing domain labels? Lexicons see only stated words, and EMONA's 57% figure shows how much moral content they miss. The count/category baseline and a dictionary feature set run on the same folds will show directly whether the implied inferences add anything beyond what is written.

Why not just trust model-generated inferences without human readers? Self-Talk's rates of irrelevant and nonsensical clarifications, together with the bias of language models toward dramatic content, make this unsafe. The source ablation quantifies how much accuracy and fidelity is lost when reader inferences are replaced with model inferences.

Decisive checks

The study succeeds on its central claim if three things hold. First, implied-inference profiles significantly predict conservative persuasive effect while counts do not. Second, the literal-paraphrase ablation loses most of that gain. Third, reader-elicited profiles outperform model-elicited ones by a margin that tracks the measured gap between model and reader inferences. That third result would show that reporting bias in the generator directly limits how well persuasion can be modeled.

A secondary check addresses the editorial-versus-Twitter tension. We apply the elicitation step to the seven dehumanizing domains in the immigration tweets and ask whether implied-inference profiles separate conservative from extreme-liberal uses of creature metaphors better than domain tags do. This would explain the "Creature Exception" as a difference in which inferences are aimed at which target.

The sample of 300 editorials is small, so the analysis plan and feature definitions will be preregistered and effect sizes reported alongside significance tests. A null result is also informative. It would show that the conservative null in the editorial data reflects the labels or the readers rather than the measurement.