Measuring Dehumanizing Framing by Asking What Literally Belongs in the Target's Slot

Much hateful language avoids slurs. It describes a group with words that literally apply to something else: migrants who "swarm" a border, a community that "infests" a city, people who "breed." A human reader sees the problem right away. The sentence has put a group of people into a position that ordinarily holds insects, pests, or disease. We propose to measure this directly. For every noun phrase that could refer to a social group, in any grammatical role, we hide the phrase and ask a language model which words would literally fill that slot, given only the local grammatical frame around it. When the hidden phrase refers to people but the literal fillers are overwhelmingly non-human, we flag a humanness gap. We also name the source domain the fillers come from, such as vermin, disease, flood, or machinery. That gap, and the named domain, become an explainable signal for implicit toxicity detection.

The opportunity comes from three lines of work that have not been joined. Metaphor detectors model the clash between a word and its context well, but only inside narrow syntactic templates, usually verbs and their arguments, and only on clean text. Implicit-hate classifiers handle harmful text but find their targets with named-entity taggers or fixed identity lists. The one study that fed metaphor information into a hate classifier used hand annotations. Our contribution is an automatic, target-agnostic measure of dehumanizing framing that is not tied to any part of speech. We evaluate it where it should matter and where current systems are weakest: source domains and target groups never seen in training.

Core Claim

Much implicit dehumanizing hate is a selectional-preference violation, meaning a word appears in a slot whose usual fillers are a different kind of thing. Here the slot is one that normally takes non-human fillers, and it is occupied by a phrase referring to a human group. This violation can be measured for any grammatical position with one probing procedure: masking, then predicting literal fillers. The resulting signal should transfer across source domains and target groups better than representations learned end to end.

We do not expect large gains on in-distribution benchmarks. The Dutch study found that hand-annotated metaphor features added only 1.9 F1 to transformer classifiers, against up to 4.6 for SVMs (Improving Hate Speech Type and Target Detection with Hateful Metaphor Features). That result constrains our design, and we take it seriously. It suggests fine-tuned transformers already absorb much of this signal when training and test share the same metaphors. Our claim is instead about generalization to unseen metaphors and targets, and about explanations a moderator can check.

Relation to Prior Work

Four papers motivate the gap.

  • Dutch hateful-metaphor study (Improving Hate Speech Type and Target Detection with Hateful Metaphor Features). It showed that source-domain tags such as "Animals" or "Disease" help predict hate type and target in Dutch Facebook comments. The tags were written by human annotators and never extracted automatically. We automate that extraction.
  • Dehumanization framework (A Framework for the Computational Linguistic Analysis of Dehumanization). It scored vermin metaphor as the cosine similarity between a group label's vector and a fixed list of vermin words, aggregated over decades of newspaper text. It works on corpora, not sentences, and covers one domain. Our probe works on individual sentences and names whatever domain the fillers come from.
  • AmpleHate (AmpleHate: Amplifying the Attention for Versatile Implicit Hate Detection). It models how a target relates to its context. However, it finds targets with a named-entity tagger restricted to organizations, nationalities, religious or political groups, and places. When no entity is tagged, it falls back to the sentence-level classification embedding. It also models the target–context relation as generic attention, with no notion of what kind of thing the context implies.
  • Explanation regularization (Contextualizing Hate Speech Classifiers with Post-hoc Explanation). It pushes the classifier's attention from 25 identity words onto their context. It cut false alarms on neutral news text containing identity words, but the authors say it struggles when harm is carried metaphorically.

Two metaphor papers supply the mechanism we generalize.

  • Word Pair based Domain Mining (WPDM) (Bridging Word-Pair and Token-Level Metaphor Detection with Explainable Domain Mining). It compares the conceptual domain of a verb's arguments in the input against a literal dictionary example, and uses the resulting inconsistency score to weight a RoBERTa classifier. It reaches 83.8 F1 on MOH-X and 84.9 on LCC-Verb, where GPT-4 scores 69.7. It is limited to verbs and depends on WordNet and VerbNet coverage, which fails on slang.
  • Chinese metaphorical relation extraction (Chinese Metaphorical Relation Extraction: Dataset and Models). It extracts variable-length target and source spans, freeing detection from fixed word pairs. It links a target to a "source action" span but never names the unstated source domain, and it scores only 36.73 F1 on attribute relations.

Our probe keeps WPDM's idea of comparing against literal usage. It swaps the dictionary lookup for predicted fillers, so it works in any syntactic slot and with out-of-vocabulary words. Like the relation-extraction work, it frees detection from fixed templates, but it also names the implicit source domain.

One contradiction needs addressing. The Latent Hatred study (Latent Hatred: A Benchmark for Understanding Implicit Hate Speech) found that injecting ConceptNet and Wikidata facts did not help implicit hate detection. Metaphor systems, by contrast, gain from lexical hierarchies. Our reading is that entity facts are the wrong kind of knowledge, while knowledge about what kind of thing fits a slot is the right kind. We test this directly by comparing our probe against knowledge-graph injection.

Research Program

Candidate spans. We parse each sentence and collect every maximal noun phrase as a candidate referring expression, in any role. This includes nominalized adjectives such as "the unclean" and bare plurals. Roles include subject, object, copular complement, possessor, and the head modified by an adjective.

Group-referent scorer. A lightweight scorer on top of the encoder estimates whether each span refers to a social group. It is trained with weak labels from the target-group annotations in the Implicit Hate Corpus (IHC) and the Social Bias Inference Corpus (SBIC): spans that match an annotated target, lexically or by embedding similarity, count as positive. Because the scorer reads context, it can pick out coded or misspelled group references that an entity tagger misses.

Probe. For each span $t$, we build a reduced frame. It keeps the span's governing word and that word's other dependents, and replaces $t$ with a mask. For example, "[MASK] swarming across our border" becomes "[MASK] swarming across the border." We then take the language model's top-$k$ fillers for the mask. The humanness gap is

$$ g(t) = h(t) - \sum_{w} p(w \mid \mathrm{frame}(t))\, h(w), $$

where:

  • $h(\cdot)\in[0,1]$ is a small classifier over word embeddings estimating whether a word denotes a person;
  • $p(w \mid \mathrm{frame}(t))$ is the model's renormalized probability for filler $w$ in the reduced frame.

The humanness classifier is seeded from WordNet's person hierarchy but applied through embeddings, so it covers slang. We also assign each filler to the nearest prototype in a small inventory of source domains drawn from the dehumanization and critical-metaphor literature: vermin, disease, flood or natural force, predator, dirt or waste, object or machine, and other non-human. The resulting domain distribution becomes the explanation shown to a moderator.

Reduced frames and their risk. We strip the frame down because a model shown the full hateful sentence may already have learned the metaphor and predict "migrants" as the filler. The reduced frame asks what literally goes there. This is also the main risk: web-trained models may predict the figurative filler even in reduced frames. We therefore compare three probe variants:

  • reduced frame versus full sentence;
  • a masked encoder versus a generative model prompted to fill the slot;
  • a probe model lightly fine-tuned on literal-labeled tokens from VUA, TroFi, and MOH-X.

Using the probe. The gap and domain distribution for the highest-scoring group span enter the classifier the same way AmpleHate injects its relational vector: they are concatenated to the sentence embedding before the output layer. They can also be used alone as a signal that needs no training. For Dutch LiLaH we run the same pipeline with a Dutch encoder. Its gold source-domain annotations let us score the probe's domain predictions directly, which no English benchmark in our file allows.

New annotation. We build a small evaluation resource of about 1,500 implicit-hate examples drawn from IHC's inferiority-language class and from ToxiGen. Each is labeled with the target span, the source domain (if any), and the syntactic position of the clash.

Contrast set. Paired with it is a contrast set of minimal quadruples that vary two things: whether the referent is human, and whether the usage is benign or hateful. Examples: "rats swarmed the barn," "fans swarmed the stadium," "migrants swarmed the border," and "volunteers flooded the shelter with donations."

Baselines

Plain fine-tuned encoder. The naive reference point is a RoBERTa-base classifier fine-tuned on each benchmark. For context, BERT with augmentation reaches 70.4 binary F1 and 58.6 six-way F1 on IHC (Latent Hatred: A Benchmark for Understanding Implicit Hate Speech). A majority-class score is reported alongside it.

Strongest existing methods. These are SharedCon and AmpleHate. SharedCon clusters sentence embeddings and uses cluster centroids as anchors for contrastive training (SharedCon: Implicit Hate Speech Detection using Shared Semantics). It reports 78.5 macro-F1 on IHC and 84.3 on SBIC, and is the strongest single-dataset competitor. AmpleHate is strongest in breadth: it reports an average 6.87-point gain over contrastive baselines across seven implicit-hate datasets. The file gives no per-dataset number for AmpleHate: "no reported result found on Lacuna for AmpleHate on IHC individually." We therefore rerun both methods ourselves.

Fair comparison. All systems share the same RoBERTa-base backbone, identical splits, and the same hyperparameter search budget: twenty trials per method, selected on the validation split. Results are averaged over five seeds.

Metaphor-feature baselines. Three baselines represent the obvious ways of adding metaphor information:

  • the Dutch study's recipe, with token-level metaphor probabilities from a VUA-trained RoBERTa tagger standing in for the human metaphor annotations;
  • the dehumanization framework's vermin-vector cosine, computed per sentence;
  • ConceptNet and Wikidata injection as in Latent Hatred.

LLM baseline. A prompted LLM baseline asks an instruction-tuned model whether the text dehumanizes a group, and which group. There is no reported result found on Lacuna for LLM prompting on these implicit-hate benchmarks. The closest evidence is that prompted LLMs trail supervised systems on metaphor (Using Large Language Models to Perform MIPVU-Inspired Automatic Metaphor Detection).

Ablations. Four ablations isolate the key ingredient:

  • restricting probing to verb subjects and objects, which imitates WPDM's scope;
  • probing with full context instead of reduced frames;
  • keeping our group-span scorer but dropping the probe, which is AmpleHate with better targets;
  • replacing the humanness gap with generic metaphoricity.

What counts as beating a baseline. On IHC and SBIC we require the full system to stay within one F1 point of SharedCon. On the held-out-domain and held-out-group splits below, it must beat both SharedCon and AmpleHate by at least five F1 points, with seed-level confidence intervals that do not overlap. The five-point margin is our own pre-registered threshold, not a figure from prior work.

Why not just…?

Why not just fine-tune a strong encoder, perhaps on ToxiGen, and let it learn metaphor implicitly? This is the real threat. Fine-tuning on ToxiGen improves human-written benchmarks by 7 to 19 percent (ToxiGen: A Large-Scale Machine-Generated Dataset for Adversarial and Implicit Hate Speech Detection). Metaphor features added only 1.9 F1 to Dutch transformers. We therefore concede the in-distribution case in advance and make generalization the test instead. The evidence that implicit learning breaks off-distribution comes from metaphor work: RoBERTa fine-tuned on legacy data shows a 59 percent false-positive rate on contemporary headlines (NewsMet : A 'do it all' Dataset of Contemporary Metaphors in News Headlines). If ToxiGen-augmented RoBERTa matches our system on the held-out-domain split, the core claim is refuted.

Why not just prompt a large language model to judge dehumanization? Prompted models over-flag literal text and miss conventional metaphors (Using Large Language Models to Perform MIPVU-Inspired Automatic Metaphor Detection). GPT-4 reaches 69.7 F1 on LCC-Verb against 84.9 for a domain-inconsistency model (Bridging Word-Pair and Token-Level Metaphor Detection with Explainable Domain Mining). Neither result concerns toxicity directly, so we run the prompted baseline on every split. The contrast set is where we expect it to fail, by over-flagging benign crowd metaphors.

Why not just add an off-the-shelf metaphor detector's scores as features? Metaphoricity is the wrong target. Benign text about groups is full of metaphor ("flooded with support"), and news-style detectors already over-fire. Our generic-metaphoricity ablation tests exactly this substitution. If it matches the humanness gap, the humanness-specific part of the idea is unnecessary.

Why not just fix AmpleHate's target finder? A better span scorer might recover most of the gains without any probing. The "scorer without probe" ablation measures this directly. If it closes the gap on the held-out splits, the contribution shrinks to target finding, and we would report it that way.

Decisive Experiments

Held-out source domain. We train on IHC and ToxiGen with every example whose annotated or probe-detected source domain is vermin removed, then test on vermin-framed hate. We repeat this for disease and flood. This is the split where a signal measured by probing, rather than memorized, should pay off.

Held-out target group. We hold out groups among ToxiGen's thirteen and retrain, testing whether the group-span scorer and probe transfer to targets never seen in training.

Performance by syntactic position. On the new annotated set we break performance down by the clash's syntactic position. The verb-only ablation should lose most on copular, adjectival, and possessive framings. That is the direct test of the syntactic-scope claim.

Contrast set. We measure false positives on benign human-figurative and literal non-human sentences. The keyword stress test in the explanation-regularization paper, where accuracy on neutral news text rose from about 77.8 to 90.1 percent, is the precedent for this kind of check.

Domain validity. On LiLaH we report agreement between probe-predicted source domains and the gold human tags. This validates the explanations independently of classification accuracy.

Stop condition. If reduced-frame probes still predict human fillers for conventional hateful frames, the humanness gap collapses. We will report that as a finding about how far web-trained models have absorbed dehumanizing usage, since that result would itself matter.