Implicit toxicity is hate or abuse that carries no slur or profanity. "We shouldn't lower our standards to hire more women" insults women through who is attached to which claim, not through any offensive word. Whether a sentence like this is toxic often depends on its grammar: which group is the subject or object of which evaluative verb, and where a negation applies. That suggests detectors should use syntactic structure, meaning a dependency parse, a tree linking each word to the word it modifies. Work on semantic role labeling (systems that recover "who did what to whom") gives a warning, though. When an automatic parser makes mistakes, those mistakes flow downstream and limit the final system. Zhang et al. measured a 7.9-point F1 gap between automatic and hand-corrected trees, and found that bad syntax can be worse than no syntax at all (Syntax-Enhanced Self-Attention-Based Semantic Role Labeling). Abusive text is also exactly where parsers should do worst, because it is full of slang, sarcasm, coded spellings, and chat fragments.
This note proposes the measurement nobody has made and a narrowly targeted fix. First, we build a small set of hand-corrected parses over implicit hate corpora. With it we can compare a detector given perfect syntax against the same detector given automatic syntax and given none. That comparison tells us whether structure helps implicit toxicity detection at all once parser error is removed, and how much of that help automatic parsers throw away. Second, if the gap is real, we close it with a detector that uses the parse in only one place: the path between a targeted group and the words that evaluate it. During training, its trees are deliberately corrupted in the specific ways parsers actually fail on abusive text, so the model learns when to stop trusting them.
Core Claim
Our hypothesis has three parts. Implicit toxicity decisions depend disproportionately on a few dependency arcs, namely those connecting a target mention to its evaluative predicate and to any negation or modal scoping over it. Off-the-shelf parsers fail on these arcs more often than on average arcs in abusive registers. A detector that reads only these target-anchored paths, and is trained under corruption matched to real parser errors, recovers most of the gold-tree benefit without inheriting the "bad syntax is worse than no syntax" penalty. Each part is falsifiable separately. The first experiment can kill the whole agenda, and we treat that as an acceptable outcome.
Where This Sits
The implicit toxicity literature motivates the structural angle without using syntax. AmpleHate finds targets with a named-entity tagger and the sentence-level token, then amplifies attention between target and context. It outperformed contrastive baselines by an average of 6.87 points across seven implicit hate datasets, but names tagger errors as an unaddressed failure point. That failure is the same propagation problem in a different module. We replace its flat target–context attention with typed dependency paths, and we make the propagation explicit and measurable. Social Bias Frames shows that models identify the targeted group well (BLEU in the 70s–80s) but explain the implied claim poorly (BLEU 44–52). This is consistent with the target being easy and its relation to the predicate being hard. Latent Hatred and Tox-BART constrain the design. External knowledge graphs gave little benefit, and Tox-BART found that random tuples worked as well as retrieved ones. A structural signal therefore has to pass the same kind of shuffle test, which we build in. Features or Spurious Artifacts? found that masking identity words beats deleting them, because masking keeps the sentence's grammar intact. This is indirect evidence that structure carries signal, though that paper never parses.
The parsing literature supplies the methods and the tension we test. Zhang et al. supply the attention mechanism that injects dependency relations and paths into the Transformer's attention scores. He et al. supply two things: pruning candidates to within $k$ dependency hops of a predicate, and a faulty-tree generator. With that generator they found models kept over 90% F1 under trees that were 90% accurate. This contradicts the "worse than no syntax" finding, and our corruption curve is designed to settle the disagreement in this domain. Leveraging Denoised AMR for GEC randomly masks graph nodes and edges so that a model ignores corrupted semantic graphs, gaining 1.57 F0.5. We borrow the idea but make the corruption non-uniform, matched to measured parser confusions. Mimic and Conquer distills tree knowledge into a sequential model so that no parser runs at test time. It is the most serious competing approach and appears below as a baseline. CONDA is the closest bridge between toxicity and structure, but it uses flat, lexicon-assigned token slots rather than trees.
Research Program
The first component is a diagnostic resource. We hand-annotate Universal Dependencies trees for about 1,500 posts: roughly 600 implicit posts from Latent Hatred, 600 from SBIC, and 300 CONDA chat utterances. Annotators also mark the target mention, its evaluative predicate, and any negation, modal, or comparative operator that scopes over that link. This is enough to report parser accuracy on toxic registers, both overall and on target–predicate arcs specifically. It also lets us estimate a per-relation confusion matrix of parser errors for each register. No such measurement exists in the file. We test a biaffine parser on a BERT encoder plus two widely used toolkit parsers.
The second component is the detector. A RoBERTa-large encoder is shared across all conditions. The parse is used only to find target candidates and their nearby structure. Target candidates are identity-term matches, group-referring pronouns, and nominals that annotators frequently tag as targets. For each target, we keep the predicates and operators within $k$ hops, following He et al. We then add a learned bias to the attention scores between the target and those tokens, indexed by the label sequence of the dependency path, following Zhang et al. All other attention is left untouched. This restriction is what we expect to stop parse errors elsewhere in a noisy sentence from hurting the model. We also feed the parser's own arc probabilities to a per-path gate so that low-confidence paths can be down-weighted.
The third component is register-matched corruption. During training, each tree is resampled using the confusion matrix estimated from the gold subset. A subject arc that parsers on Latent Hatred mislabel as a modifier 20% of the time is corrupted at that rate. We compare this against uniform edge dropout of equal total rate, which is the Cao & Zhao recipe.
Baselines
The naive reference point is RoBERTa-large fine-tuned with no syntax under the same budget. Its published counterpart is BERT plus augmentation from Latent Hatred, which reports 70.4 F1 on binary implicit hate and 58.6 F1 on the six-way taxonomy. On SBIC, the reference is about 80.5 F1 on offensiveness classification from Social Bias Frames. On CONDA, Joint BERT reaches 0.59–0.76 F1 on implicit toxicity utterances (CONDA).
The strongest existing method is AmpleHate. It addresses the same target–context relation, using a tagger instead of a parse. The file reports only its 6.87-point average margin over contrastive baselines, and no reported result was found on Lacuna for AmpleHate's absolute F1 on Latent Hatred or SBIC. We will therefore reimplement it on our encoder and splits. Beating it means a statistically reliable gain, over five seeds, in macro F1 on Latent Hatred and SBIC. The gain should be larger on the structure-sensitive contrast set described below.
Tree distillation as in Mimic and Conquer is the strongest structure-using competitor that has no runtime parser. We train it with the same parsers as teachers. For veiled toxicity recall, the reference is fortification by Han & Tsvetkov, which raised recall from 1.2% to 51.1% at a cost of about 12% in non-offensive accuracy. Our method is complementary to fortification and can be stacked with it. We report both alone and combined.
Every system gets the same encoder, the same training data, 20 hyperparameter trials, and five seeds, and is selected on the same development split. The ablations that isolate the key ingredient are as follows. Applying syntax-aware attention to all arcs rather than target-anchored paths tests the restriction. Uniform corruption versus register-matched corruption tests the noise model. Gold trees versus automatic trees versus no trees on the annotated subset measures the ceiling. A shuffled-tree control, using trees from a different sentence of the same length, tests whether the structure is actually used, mirroring Tox-BART's random-tuple test.
Why not just…?
Why not just fine-tune a large pretrained encoder, which already encodes syntax internally? This is the main threat. Zhang et al. found explicit high-quality syntax still added about 3 F1 on top of BERT in role labeling, but toxicity is not role labeling. The gold-versus-none comparison on our annotated subset answers the question directly. If gold trees give no gain over no syntax, the encoder already has what it needs and we stop.
Why not just distill syntax into the encoder and skip runtime parsing? This may work as well, and for moderation at chat volume it would be cheaper. Distillation, however, learns from teacher parsers that make exactly the errors we measure, so it may absorb them. We run it as a full baseline. If it matches the parse-based detector on the contrast set, we will recommend it and report that runtime trees are unnecessary. Our contribution would then reduce to the ceiling measurement and the choice of teacher.
Why not just improve target finding, for example with a better tagger or an LLM that names the target? Finding the target is not where models struggle; Social Bias Frames already shows target identification at BLEU in the 70s–80s. The failure lies in relating the target to its predicate under negation or comparison. An ablation that gives AmpleHate gold target spans tests this. If gold targets alone close most of the gap, structure beyond the target is not the bottleneck.
Why not just prompt a large language model? Tox-BART found GPT-3.5 explanations vague and less specific about target groups than a tuned model. We still include a zero-shot and few-shot prompted model on the contrast set as a reference.
Proposed Experiments
The decisive first check runs on the 1,500 gold-annotated posts. We train with gold, automatic, and no trees and compare the results. Gold must beat no syntax on the structure-sensitive subset by a margin exceeding seed variance, or the agenda ends with a published negative result. Second, we sweep corruption from 0% to 40% arc error using the register-matched generator. This produces the accuracy-versus-tree-quality curve that decides between He et al.'s robustness finding and Zhang et al.'s "worse than no syntax" finding for this domain. We also locate the point where automatic parsers on Latent Hatred and CONDA actually sit on that curve.
Third, we build a contrast set of about 600 minimal pairs from Latent Hatred and SBIC posts. Each pair changes toxicity only by moving an attachment or a negation scope, while keeping the identity words fixed. This is where target-anchored structure should beat both the plain encoder and AmpleHate, and where the shuffled-tree control should collapse to the plain encoder. Fourth, we check whether structure reduces keyword bias. On the NYT Adversarial Set, vanilla BERT scores about 77.8% and explanation regularization reaches 90.1% (Contextualizing Hate Speech Classifiers). Identity terms there mostly sit as arguments of neutral predicates, so a detector that reads target–predicate paths should produce fewer false positives without the out-of-distribution F1 loss that masking incurs in Ramponi & Tonelli. Finally, we run full-corpus results on Latent Hatred, SBIC, and CONDA against all baselines. We report macro F1, implicit-class F1, veiled recall, and non-offensive accuracy together, so that recall gains bought with false positives are visible.
References
- Syntax-Enhanced Self-Attention-Based Semantic Role Labeling
- AmpleHate: Amplifying the Attention for Versatile Implicit Hate Detection
- Social Bias Frames: Reasoning about Social and Power Implications of Language
- Latent Hatred: A Benchmark for Understanding Implicit Hate Speech
- Tox-BART: Leveraging Toxicity Attributes for Explanation Generation of Implicit Hate Speech
- Features or Spurious Artifacts? Data-centric Baselines for Fair and Robust Hate Speech Detection
- Syntax for Semantic Role Labeling, To Be, Or Not To Be
- Leveraging Denoised Abstract Meaning Representation for Grammatical Error Correction
- Mimic and Conquer: Heterogeneous Tree Structure Distillation for Syntactic NLP
- CONDA: a CONtextual Dual-Annotated dataset for in-game toxicity understanding and detection
- Fortifying Toxic Speech Detectors Against Veiled Toxicity
- Contextualizing Hate Speech Classifiers with Post-hoc Explanation