Toxicity classifiers are almost always trained on moderated, pseudonymous platforms like Reddit and then pointed at places with very different social rules: semi-anonymous Q&A sites, fringe boards, dark-web forums. Psychologists call the behavioral change these places induce online disinhibition, meaning people drop their usual restraint when they feel unseen. The change cuts both ways. It produces more abuse, and it also produces more candid self-disclosure and mutual support. On the dark-web forum Dread, posts are measurably more toxic than in topic-matched Reddit communities, yet they carry significantly more emotional and appraisal support (Parallel Communities Across the Surface Web and the Dark Web). A detector moved naively into such a space faces two failures at once: it misses new forms of harm, and it flags blunt, profane, supportive talk as abuse.
Adversarial domain adaptation is the standard tool for moving a classifier to a new environment without target labels. A second network, the domain discriminator, tries to tell source from target examples, and the encoder is trained to fool it, so that the learned features stop encoding "which platform is this." We argue that this tool, applied in the obvious way, is built to make the second failure worse. The words that mark supportive posts on anonymous platforms (swearing, slurs used affectionately, raw descriptions of crises) look like the words that mark toxic posts on the source. Forcing the two distributions to overlap pulls supportive target posts toward the source's toxic region. Our proposal has two parts. The first is to condition the alignment on predicted support as well as predicted toxicity, so that the discriminator compares supportive posts with supportive posts. The second is to correct for the fact that the two platforms really do have different rates of both behaviors. We will test whether this lowers the rate at which supportive posts are flagged as toxic on the target platform without giving up toxicity detection.
Core Claim
Disinhibition shifts break the assumption behind standard adversarial alignment, which is that only the input distribution changes while the meaning of features stays fixed. Across anonymity regimes, both the base rates of toxic and supportive behavior and the social meaning of profanity change (A Comparison of Common Users across Instagram and Ask.fm to Better Understand Cyberbullying documents "affectionate profanity" and defensive "upstanding" that dictionary methods mistake for bullying). We make two predictions. First, marginal and toxicity-only conditional alignment will raise the false-positive rate on supportive target posts relative to a model with no adaptation, even when overall target F1 improves slightly. Second, conditioning the discriminator on a benign-disinhibition head and reweighting for estimated class-prior shift will remove most of that damage. The new measurable object is what we call the sanctuary false-positive rate: the fraction of human-annotated supportive target posts that the toxicity head flags.
Relation to Prior Work
Two papers come closest, and both treat "domain" as something other than the social regime. Improving Generalization of Hate Speech Detection Systems to Novel Target Groups via Domain Adaptation adapts hate classifiers across demographic target groups within HateXplain. It reports 5.8% to 10.7% relative F1 gains over zero-shot, finds curriculum pseudo-labeling as strong as adversarial training, and flags negative transfer as unresolved. This paper motivates our concern about negative transfer and supplies two of our baselines. Its domains share one platform and one moderation regime, whereas ours differ precisely in anonymity and moderation. Domain Classification-based Source-specific Term Penalization for Domain Adaptation in Hate-speech Detection uses an offline domain classifier and attribution penalties to stop models from leaning on source-specific words. It reaches 60.5 average macro-F1 against 58.0 for vanilla fine-tuning across Waseem, HatEval and Vidgen, at about 4.5x training cost. That approach suppresses words. Ours keeps the words and changes what they are aligned with, which matters when the same word is a shortcut on one platform and a marker of intimacy on another.
The method itself borrows from two sources. Conditional Adversarial Domain Adaptation conditions the discriminator on the outer product of features and predicted class probabilities, with entropy weighting. It addresses the "mode mismatch" problem, in which one class in the source gets aligned with a different class in the target. Its conditioning uses a single label space, and we add a second head whose job is to separate exactly the modes that toxicity labels conflate. Adversarial Topic-Aware Prompt-Tuning for Cross-Topic Automated Essay Scoring splits representations into an adversarially aligned shared part and an unaligned topic-specific part. We reuse that split for platform context. The invariance-discriminability literature, for example Chen et al. and Shi et al., constrains the design: it shows that aggressive alignment collapses classes into uninformative overlap, so our alignment weight is kept small and is ablated.
Parallel Communities Across the Surface Web and the Dark Web supplies the data regime and the empirical motivation. It is purely descriptive and never tests how a Reddit-trained classifier behaves on Dread. Its finding that GPT-4o-mini annotates support with over 96% accuracy also raises the most serious competing approach, which we address directly below.
Research Program
The model is a RoBERTa-base encoder with two classification heads. The first predicts toxicity and the second predicts emotional or appraisal support. On the source platform, toxicity labels come from existing annotated data. Support labels come from a 5-shot GPT-4o-mini labeling pass replicating the Parallel Communities protocol, checked against a human-annotated subset. A small platform-context adapter in the style of ATOP's topic-specific prompts is left out of alignment, and both heads may read it. The rest of the representation is aligned adversarially through a gradient reversal layer, a layer that passes features forward unchanged but flips the sign of the gradient on the way back, so the encoder learns to confuse the discriminator.
The discriminator receives the encoder features $f$ combined with the concatenated predictions $g = [p_{tox}, p_{sup}]$ through the randomized multilinear map from CDAN. Each example is weighted by
$$w(x) = \big(1 + e^{-H(g(x))}\big)\cdot \frac{\hat{\pi}_t(\hat{y}(x))}{\pi_s(\hat{y}(x))},$$
where $H$ is prediction entropy, $\hat{y}(x)$ is the predicted joint (toxic, supportive) cell, $\pi_s$ is the source proportion of that cell, and $\hat{\pi}_t$ is the estimated target proportion. We obtain $\hat{\pi}_t$ by inverting the source-validation confusion matrix applied to target predictions. The first factor is CDAN's entropy weighting. The second keeps the discriminator from pushing the model to equalize toxic and supportive rates that genuinely differ across platforms. For orientation, Detoxify gives 8.08% mean toxicity on Dread against 6.82% on Reddit, and the support gap is significant.
The analysis side builds the evaluation the field lacks. We will annotate held-out target posts for toxicity, for support, and for whether profanity is used hostilely, affectionately, or defensively. Sanctuary false-positive rate can then be measured directly, and so can the narrower subset of supportive posts that also contain profanity.
Baselines
The naive reference point is source-only fine-tuning: the same encoder trained on labeled source data and applied unchanged to the target. No reported result was found on Lacuna for Reddit-to-Dread or Instagram-to-Ask.fm toxicity transfer, so this run defines the floor on our own test sets. On the Waseem/HatEval/Vidgen benchmark, the corresponding vanilla figure is 58.0 average macro-F1 from the Dom-spec paper, and we will reproduce it as a calibration check.
The strongest existing text-domain method on a published benchmark is Dom-spec attribution penalization, with 60.5 average macro-F1 (DeepLIFT variant) and 60.6 when combined with identity lists (Domain Classification-based Source-specific Term Penalization…). We run it on our platform pairs with the same encoder and the same tuning grid size. On its own benchmark, beating it means exceeding 60.5. More importantly, on our pairs we will compare sanctuary false-positive rate, which we expect it to handle poorly because it suppresses the very terms that supportive anonymous posts use.
Curriculum labeling from Improving Generalization… is the strongest non-adversarial unsupervised competitor, since that paper found it matched or beat adversarial training. We will run it with class-wise thresholds. We also run plain gradient-reversal alignment and CDAN with entropy weighting conditioned on toxicity alone. These isolate whether conditioning on support, as opposed to conditioning in general, is what helps.
The strongest practical competitor is GPT-4o-mini labeling of unlabeled target data followed by distillation into the same RoBERTa model. It needs no alignment at all.
The proposal's key ingredient is isolated by ablations that remove, one at a time, the support conditioning, the prior-correction weights, and the unaligned context adapter.
All methods share the encoder, the unlabeled target pool, five seeds, and an equal hyperparameter budget of 20 configurations, with model selection on a small labeled target validation split shared across methods. To count as a win, our method must cut sanctuary false-positive rate by a statistically significant margin, measured by a bootstrap over seeds and posts, relative to toxicity-only CDAN. It must do this while matching or exceeding curriculum labeling and LLM distillation on target toxicity macro-F1, within one standard deviation.
Why not just…?
Why not just fine-tune a large pretrained model and skip adaptation? Both cited hate-speech adaptation papers report that vanilla fine-tuning loses out of domain: 58.0 against 60.5 macro-F1 in Dom-spec, and 5.8% to 10.7% relative F1 behind adaptation in the target-group study. Dom-spec's qualitative analysis shows a vanilla model flagging a mental-health post because it contains a gendered slur, which is our failure mode in miniature. The source-only run will show whether this holds across anonymity regimes as well.
Why not just use curriculum pseudo-labeling instead of a minimax game? This is a real threat, because it was the best average performer in Improving Generalization…. Our concern is structural. Pseudo-labels on the target come from the source-trained model, which already marks profane supportive posts as toxic, so self-training should reinforce that error. If curriculum labeling matches our sanctuary false-positive rate, the support-conditioning claim is false. We will then report that, and we will test support-aware thresholds inside curriculum labeling as the simpler fix.
Why not just have an LLM label the target platform and distill? GPT-4o-mini reached over 96% accuracy on support in Parallel Communities…, so it may also label Dread toxicity well. The proposal's value then shrinks to producing a cheap, private, local model, which matters for dark-web and intimate-partner-violence settings where sending text to an API is ethically fraught (A Survival Guide for Iranian Women…). We test this head-on. We also test our alignment on top of LLM pseudo-labels, since the two are compatible. If distillation alone wins on both metrics, the honest conclusion is that alignment is unnecessary once a capable labeler is available, and we will state that.
Why not just mask profanity and identity terms? In Latent Hatred, only 3.2% of implicit hate contains explicit expletives, so masking does nothing for implicit harm. Dark-web text also gives ordinary words specialized meanings (Shedding New Light on the Language of the Dark Web). Masking also deletes the cues that distinguish affectionate from hostile swearing. We include a masked source-only model as a cheap reference.
Proposed Experiments
The main testbed is the 29 topic-matched Reddit and Dread communities from Parallel Communities, using Reddit as the source and unlabeled Dread as the target. We will annotate 2,000 Dread and 1,000 Reddit posts, three annotators each. Sampling will be stratified so that supportive-and-profane posts are well represented.
The second pair is Instagram to Ask.fm from the common-users collection, with 1,500 annotated Ask.fm items. It adds a test the Reddit/Dread pair cannot provide: because the same people post on both platforms, we can check whether the unaligned context adapter encodes the platform rather than the person. If users' context vectors cluster by platform more than by user, the adapter is doing its intended job.
A third, fully labeled shift is EDOS Gab to Reddit, provided the source platform of each post can be recovered from its metadata. Its 1M-post unlabeled pools per platform serve as adaptation data, and the published 82.88% Sub-task A macro-F1 from NLP_CHRISTINE anchors in-domain performance. We also run on Waseem/HatEval/Vidgen and on the HateXplain target-group splits, to confirm the method does not regress where prior numbers exist.
Three checks decide the claim. The first: plain gradient-reversal alignment and toxicity-only CDAN must raise sanctuary false-positive rate above the source-only floor on Reddit-to-Dread. If they do not, the motivating failure is absent and the proposal reduces to a benchmark. The second: adding support conditioning must lower that rate, and the ablation without prior correction should show partial recovery, which would tell us how much of the damage comes from mismatched class rates versus mismatched modes. The third: across all pairs and seeds, we count negative transfer, meaning runs that score below source-only on target macro-F1, for each method. Prior-corrected support conditioning should have the fewest such runs. The LLM-distillation comparison sits alongside all three and determines whether the contribution is a better adaptation method or a smaller, local substitute for an API labeler.
References
- Parallel Communities Across the Surface Web and the Dark Web
- A Comparison of Common Users across Instagram and Ask.fm to Better Understand Cyberbullying
- Improving Generalization of Hate Speech Detection Systems to Novel Target Groups via Domain Adaptation
- Domain Classification-based Source-specific Term Penalization for Domain Adaptation in Hate-speech Detection
- Conditional Adversarial Domain Adaptation
- Adversarial Topic-Aware Prompt-Tuning for Cross-Topic Automated Essay Scoring
- Transferability vs. Discriminability: Batch Spectral Penalization for Adversarial Domain Adaptation
- Information-Theoretical Learning of Discriminative Clusters for Unsupervised Domain Adaptation
- A Survival Guide for Iranian Women Prescribed by Iranian Women: Participatory AI to Investigate Intimate Partner Physical Violence in Iran
- Latent Hatred: A Benchmark for Understanding Implicit Hate Speech
- Shedding New Light on the Language of the Dark Web
- NLP_CHRISTINE at SemEval-2023 Task 10: Utilizing Transformer Contextual Representations and Ensemble Learning for Sexism Detection on Social Media Texts