A hate speech detector trained on one platform tends to misfire when moved to another. Gab, Reddit, Twitter and YouTube differ in slang, moderation, and how heated ordinary conversation gets, so the same raw score means different things on each. Speaker verification, the task of deciding whether two voice recordings come from the same person, faced the same problem decades ago with microphones and channels. Its standard fix is score normalization. A raw similarity is not judged on its own. It is compared with the similarities the system produces for a crowd of known non-matches, called a cohort, and the score is restated as "how many standard deviations above the crowd". That lets a single decision threshold work across recording conditions. The adaptive version picks, for each trial, only the cohort members closest to it.
We propose to carry that adaptive, per-trial normalization over to cross-platform hate detection. The detector stays frozen. The cohort is built from unlabeled posts on the new platform, and each post is scored by how much closer it sits to hateful prototypes than its benign platform neighbours do. The claim is specific and testable. Because the normalization statistics depend on the individual post, the method can change the ranking of posts, not just shift a threshold, so it can raise AUROC (area under the ROC curve). Re-tuning a global threshold or Platt scaling cannot do that. We also expect it to stack on top of the strongest disentanglement methods rather than compete with them.
Core Claim
Existing cross-platform detectors all work inside the representation. PEACE re-weights a RoBERTa detector's attention using frozen sentiment and aggression models. CATCH splits representations into platform-dependent target features and platform-invariant "causal" features, using a variational autoencoder and a reconstruction decoder. HATE-WATCH replaces CATCH's target labels with pseudo-labels, contrastive clustering and entropy reweighting. None of the three adapts the decision stage to the target platform: the output threshold is the same everywhere.
CATCH's own ablation is the motivating evidence that this matters. Removing its target-specific loss costs 10 to 18 percent, so platform-dependent information carries real predictive signal and cannot simply be discarded. Score normalization offers a different trade. It keeps the full representation, including platform-specific cues, and resolves platform bias at the scoring boundary, relative to that platform's own benign traffic.
The idea of relative competitiveness against a cohort is mature in speech. Huh et al. use cohort score distributions to absorb noise and quality variation. Jin et al. adapt thresholds to incoming audio using only a test-side cohort. Castán et al. normalize both sides of each trial, using unlabeled cohorts, to handle new conditions. Adaptive symmetric normalization (AS-norm) closes much of the gap between self-supervised and supervised speaker models in SDPN, reaching 1.29% EER on VoxCeleb1-O. These papers supply the method. They do not say whether it transfers to text, where the cohort must be mined from an unlabeled stream that may itself contain the "targets", meaning hateful posts.
The closest analogues outside speech constrain rather than pre-empt us. Score normalization for demographic fairness applies cohort calibration to face recognition, and language-dependent normalization partitions cohorts by spoken language. Both show that domain-matched cohorts help, but both stay within biometrics.
Research Program
We recast detection as verification. Let $f$ be a frozen source-trained encoder (RoBERTa-base, or the encoder of CATCH or HATE-WATCH). Let $p_1,\dots,p_K$ be hateful prototypes, obtained by clustering source-platform hateful posts in that encoder's space. Let $C$ be a cohort of presumed-benign posts from the target platform. The raw score of a post $x$ is $s(x)=\max_k \cos(f(x),p_k)$, and the prototype attaining the maximum is $p^*$. The normalized score is
$$\tilde s(x)=\tfrac12\Big[\tfrac{s(x)-\mu_Z(p^*)}{\sigma_Z(p^*)}+\tfrac{s(x)-\mu_T(x)}{\sigma_T(x)}\Big].$$
The two halves use different statistics. $\mu_Z(p^*)$ and $\sigma_Z(p^*)$ are the mean and standard deviation of $\cos(p^*,c)$ over the $N$ cohort posts closest to the prototype. They capture how close benign platform traffic typically gets to hate, which is platform-level drift such as slang. $\mu_T(x)$ and $\sigma_T(x)$ are the same statistics of $\cos(f(x),c)$ over the $N$ cohort posts closest to $x$. They capture whether this particular post is generically similar to everything nearby. That second term targets hubness, the tendency of some points to sit close to many others. Fei et al. connect z-score normalization to hubness in few-shot embedding spaces.
Because $\mu_T$ and $\sigma_T$ vary per post, $\tilde s$ is not a monotone function of $s$. That is what allows a ranking change.
The hard part, absent in speech, is cohort contamination. On Gab, hateful posts are common, so "nearest unlabeled neighbours" may be hateful. That would suppress exactly the posts we want to flag. We study three cohort constructions. The first is a static source-platform benign cohort, which is the mismatched baseline. The second is target-platform unlabeled posts filtered to the lowest-scoring fraction under the raw detector, a bootstrap. The third adds learnable impostor embeddings in the style of LIE-TAS-norm. That method learns cohort vectors with sub-centers under a log-likelihood-ratio cost. Its authors explicitly leave fully out-of-domain cohorts as open work, and our leave-one-platform-out setting tests exactly that. Here the vectors are initialized from the filtered target cohort and refined only with source labels.
The cohort size $N$ and the filtering fraction are tuned on a held-out source platform, never on the target. This addresses the brittleness in choosing top-$N$ noted in the speech literature.
Baselines
The naive reference is a RoBERTa-base detector trained on the source platforms and applied to the target with the threshold chosen on source validation data. Every other method must beat it, and we report its gap to an oracle target-tuned threshold to show how much of the cross-platform loss is pure threshold drift.
The ablation that isolates our key ingredient, per-post adaptivity, is global platform recalibration. It z-scores or quantile-matches the detector's logits using the mean and spread over the whole unlabeled target pool, which is the strongest label-free version of threshold re-tuning. It can fix macro-F1 at a fixed threshold but provably cannot change AUROC. A second ablation is AS-norm with a static source cohort, which isolates the value of using target-platform posts. A third is the T-side-only versus symmetric score, which isolates the hubness term.
The strongest existing method is HATE-WATCH, which reports average macro-F1 of 0.66 across GAB, Reddit, X and YouTube, against 0.63 for HateBERT and 0.62 for HateXplain (HATE-WATCH). CATCH reports 0.82 macro-F1 on GAB, against 0.81 for HATE-WATCH. No reported result was found on Lacuna for PEACE's numbers. We will retrain HATE-WATCH and CATCH in our pipeline on identical leave-one-platform-out splits, with the same backbone and the same hyperparameter search budget, rather than compare to reported figures.
Beating them means one of two outcomes. Either normalization on a plain RoBERTa detector matches HATE-WATCH's average macro-F1 without any target-side training, or normalization applied on top of HATE-WATCH yields a gain that is statistically reliable across five seeds and survives a paired bootstrap over posts.
Why not just…?
Why not just re-tune the threshold or apply Platt scaling on the target platform? Both are monotone transforms of the score, so they leave AUROC unchanged. They also need target labels unless done by quantile matching. Our global-recalibration ablation is precisely this alternative made label-free. If it captures all of the macro-F1 gain and our method adds nothing in AUROC, the per-post idea fails, and we will say so.
Why not just pool many platforms' data during training? Antypas and Camacho-Collados show pooling 13 datasets lifts TimeLMs-21 to 70.7% macro-F1 from a single-dataset best of 61.0%. That is a real threat. Pooling and normalization are orthogonal, though, and pooled training still meets unseen platforms at deployment. We therefore train a pooled detector on all but one platform and apply normalization to it. The question is whether the gain persists on the held-out platform.
Why not just adapt the model on unlabeled target data, through self-training or covariance alignment? Alignment of feature statistics in the style of CORAL improved cross-course dropout AUC by about 8% in MOOC transfer. That method needs gradient updates on target data and aligns covariance globally, which can blur class boundaries under very different base rates such as Gab versus YouTube. We run self-training and CORAL on the same unlabeled pool. If either beats normalization, the honest conclusion is that normalization is mainly a cheap, update-free alternative, and we report the combination as well.
Why not just prompt a large language model? Both CATCH and HATE-WATCH already report that specialized models outperform GPT-4 cross-platform. HATE-WATCH estimates a 15 to 20% misidentification rate for GPT-4's target labels. We include it only as a reference row.
Proposed Experiments
We use the HATE-WATCH platform suite: GAB (11k posts with targets), Reddit (40k), Twitter/X (25k) and YouTube (1k). We add Wikipedia and FRENK from PEACE as additional held-out platforms. All transfers are leave-one-platform-out. Metrics are macro-F1 at the source-chosen threshold, AUROC, the gap between source-threshold and oracle-threshold F1, and false-positive rate on benign posts.
The first decisive check is AUROC. Any significant AUROC gain over the frozen detector, which global recalibration cannot produce, confirms that per-post cohorts carry information beyond threshold drift.
The second is a contamination stress test. We inject known hateful posts into the target cohort at controlled rates spanning the platforms' base rates and measure how fast recall degrades for each of the three cohort constructions. If bootstrap filtering or learnable impostors do not hold up at Gab-like contamination, the method is limited to low-prevalence platforms.
The third tests the failure mode Yu et al. document, where ad-hominem counterspeech uses insults half the time. We score their 3,654 Reddit replies to hateful comments with a detector trained elsewhere. The prediction is that adaptive cohorts, which contain the platform's benign heated replies, cut false positives on counterspeech more than global recalibration does. If they do not, the hubness term is not doing the work we attribute to it.
References
- PEACE: Cross-Platform Hate Speech Detection - A Causality-Guided Framework
- Causality Guided Disentanglement for Cross-Platform Hate Speech Detection
- Cross-Platform Hate Speech Detection with Weakly Supervised Causal Disentanglement
- The VoxCeleb Speaker Recognition Challenge: A Retrospective
- Investigation of Cross-Show Speaker Diarization
- Improving Robustness of Speaker Recognition to New Conditions Using Unlabeled Data
- Pushing the Frontiers of Self-Distillation Prototypes Network with Dimension Regularization and Score Normalization
- Score Normalization for Demographic Fairness in Face Recognition
- Cross-Lingual Speaker Verification with Domain-Balanced Hard Prototype Mining and Language-Dependent Score Normalization
- Z-Score Normalization, Hubness, and Few-Shot Learning
- Trainable Adaptive Score Normalization for Automatic Speaker Verification
- Robust Hate Speech Detection in Social Media: A Cross-Dataset Empirical Evaluation
- Transfer Learning using Representation Learning in Massive Open Online Courses
- A Fine-Grained Taxonomy of Replies to Hate Speech