Recovering Culture-Specific Word Senses for Zero-Shot Hate Speech Transfer

A hate speech classifier trained only on English and applied to Italian or Spanish fails in a revealing way. It reads the Italian exclamation porca puttana or the Spanish intensifier puta as a slur, because those words translate to English words that are slurs. Monolingual Italian models classify 99% of tweets containing such taboo interjections correctly. English-trained models get 3% right (Exposing the limits of Zero-shot Cross-lingual Hate Speech Detection). The failure has a specific cause. Italian packs "prostitute" and "exclamation of frustration" into one word, and English splits them across different words. A shared multilingual embedding space carries the English association across anyway. Linguists have measured how common this mismatch is: across BabelNet, Open Multilingual WordNet and the CLICS database, over 99.9% of concept pairs that some language expresses with a single word are not expressed that way in every language (On Universal Colexifications).

We propose to fix this at the level of individual words. For each word the English classifier depends on heavily, a dense retriever finds the concept that word expresses in its sentence. It searches a multilingual concept index built from wordnets linked through a shared interlingual concept index, so concepts missing from the target language's own wordnet can still be found through another language's entries. The system then replaces the word with an unambiguous word for that concept before classification. It uses no target-language task labels. In two sentences: an English-trained classifier fails abroad because foreign words bundle meanings differently, so we look up which meaning is in play from a cross-language dictionary of concepts and rewrite the word into one the classifier cannot misread. The result is an automatic, cross-lingual, zero-shot version of a fix that has so far existed only as a supervised, 24-word Italian pipeline.

Core Claim

A large share of the zero-shot transfer gap on taboo interjections and pejorative epithets comes from mismatches in how languages package senses into words, and it can be closed without target labels. The method rewrites only the few words the classifier relies on, each into a single-sense anchor word for the concept retrieved in context. Two things are new relative to prior work. First, the sense inventory is the union of all CILI-linked wordnets rather than the target wordnet alone. CILI, the Collaborative Interlingual Index, is a list of language-independent concept identifiers that national wordnets map their synsets onto. Using the union matters because the exclamatory or pejorative sense a target wordnet lacks is often lexicalized as a separate entry in some other language. Second, the rewrite is chosen by retrieval in context, not from a hand-built list.

Relation to Prior Work

Three papers motivate the proposal by locating the gap. Exposing the limits of Zero-shot Cross-lingual Hate Speech Detection used occlusion-based token attributions to show that taboo interjections drive zero-shot errors on HatEval Spanish, AMI and HaSpeeDe. It also found that misogyny transfer degrades far more than anti-immigrant transfer, with a 36% macro-F1 drop for XLM-R. That work is purely diagnostic, and we reuse its attribution step as our trigger. Multilingual HateCheck shows that the same weakness appears in functional tests for non-hateful profanity and slur use across ten languages. On Universal Colexifications supplies the linguistic reason to expect the problem in every language pair, not only Romance ones.

One paper supplies the injection method. PejorativITy trains an Italian classifier to decide whether one of 24 polysemous words (balena, gallina, ...) is used pejoratively. It then substitutes a single-sense anchor, grassa or cetaceo. Gold senses raise misogyny F1 from 0.68 to 0.87, and predicted senses give +9 macro-F1. We keep the substitution idea. We drop its three limits: it is monolingual, it uses a closed lexicon, and it needs sense labels. Retrieval over linked wordnets replaces all three, using the interlingual querying that Intrinsically Interlingual: The Wn Python Library for Wordnets makes practical.

Two retrieval papers supply technique. Cross-lingual Contextualized Phrase Retrieval shows that encoding a phrase together with its sentence lets a dual encoder (two separate encoders whose outputs are compared by vector similarity) separate senses across languages. CCPR-LaBSE reaches 78.5 phrase retrieval accuracy against 48.8 for plain LaBSE. However, that work relies on 1.3 billion phrase pairs mined from parallel text and targets translation. We use the same context-conditioned encoding, but against wordnet glosses and examples, so no parallel corpus is needed.

Three lines of work constrain the design. Data-Efficient Hate Speech Detection via Cross-Lingual Nearest Neighbor Retrieval retrieves labeled examples from a 265k-instance multilingual pool using bge-m3. It needs labeled target seeds, and it degrades past about 2,000 retrieved examples because the source datasets define hate inconsistently. This is why we retrieve concepts rather than labeled sentences. Disentangled Contrastive Learning for Zero-Shot Multilingual Dense Retrieval separates language-specific from semantic features. It assumes the semantic part is the same across languages, which is exactly the assumption colexification data contradicts, so we treat it as a competing encoder rather than a solution. Sociocultural knowledge is needed for selection of shots and CReHate warn that culture-specific slurs and annotator disagreement exist even within one language. We therefore do not claim to resolve contested labels, only sense confusion.

Research Program

The system has three stages, all built on an English-trained XLM-T or XLM-R classifier that is left untouched.

The first stage selects trigger words. Occlusion attribution, following the zero-shot hate speech diagnostic paper, scores each target-language token by how much removing it changes the hate probability. Only tokens above a threshold $\tau$ are passed on, typically one to three per tweet. Restricting rewrites to these tokens keeps the rest of the sentence, including real slurs the classifier handles correctly, intact.

The second stage retrieves the concept. Each CILI synset $s$ is indexed by encoding its glosses, example sentences and lemmas from every available wordnet in Open Multilingual WordNet, using bge-m3. Each trigger token $t$ in sentence $x$ is encoded in context, with a CCPR-style span marker, as $e(t \mid x)$. Retrieval then picks

$$s^* = \arg\max_{s \in \mathcal{C}(t) \cup \mathcal{N}_k(t)} \cos\big(e(t \mid x),\, g(s)\big),$$

where $g(s)$ is the synset's index vector, $\mathcal{C}(t)$ is the set of synsets the target wordnet lists for $t$'s lemma, and $\mathcal{N}_k(t)$ is the $k$ nearest synsets from the whole multilingual index regardless of lemma. The second set is the key ingredient. It lets an exclamatory use of puttana land on a frustration or intensifier concept that Italian WordNet never attaches to that lemma, provided another language's wordnet lexicalizes it separately. If the best score falls below a calibrated floor, the token is left unchanged. Calibration uses only English and the unlabeled target text, by holding out lemmas with known senses from the index.

The third stage rewrites the word. The system substitutes the lemma of $s^*$ that has the fewest senses in the inventory. It prefers a target-language lemma and falls back to English when none exists. CONCAT, which appends the concept gloss instead of replacing the word, is tested as an alternative.

The program also includes an analysis study. For each language pair, we measure how often attributed tokens have a "mismatched packaging". This means the retrieved concept differs from the concept of the token's dictionary-equivalent English word. We then test whether this rate predicts the per-language transfer gap. A positive correlation would turn the colexification result into a usable predictor of where zero-shot moderation will fail.

Baselines

The naive reference point is plain zero-shot transfer: XLM-R and XLM-T fine-tuned on English HatEval and English AMI and applied directly to the targets. On the taboo-interjection subsets the cited reported figures are 3% accuracy on Italian and 58% on Spanish, against 99% and 82% for monolingual models (Exposing the limits of Zero-shot Cross-lingual Hate Speech Detection). We rerun it with our own seeds so that every system shares the same classifier checkpoint.

The strongest existing method that fits the zero-shot constraint is translate-test. A machine translation system renders each target tweet into English before the same classifier sees it. Good translation systems often render idioms by meaning, so this baseline attacks our exact failure case. No reported result found on Lacuna for translate-test on HatEval Spanish, AMI or MHC, so we will produce it with one open translation model, identical classifier weights and no tuning beyond decoding defaults.

The second strong competitor adapts the retrieval approach of Data-Efficient Hate Speech Detection via Cross-Lingual Nearest Neighbor Retrieval to the zero-shot setting. Unlabeled target tweets serve as queries in place of the labeled seeds. The retrieved labeled non-target examples, capped at their reported 2,000-instance optimum, are added to the English training data. In its original few-shot setting that method reports up to 10% macro-F1 improvement over monolingual training with fewer than 50 target labels. That number is not directly comparable to our setting, which is why we rerun it.

We also run a profanity-masking baseline that deletes every token on a public multilingual profanity list. We run a zero-shot instruction-tuned LLM prompted in the target language. Finally, we swap in the DCL encoder for the retrieval stage, which checks whether a disentangled encoder alone fixes sense retrieval. All systems share the same English training data and classifier, and each gets the same hyperparameter budget, tuned only on English development data.

Three ablations isolate the key ingredients. Restricting retrieval to $\mathcal{C}(t)$, the target wordnet only, tests whether cross-lingual concept recovery matters. Substituting the same number of random high-attribution tokens with neutral words tests whether gains come from sense resolution or simply from deleting trigger words. On PejorativITy, plugging in gold word-level pejorativity labels gives the upper bound. Beating the baselines means two things at once. On the taboo-interjection and pejorative-epithet subsets, the method must exceed both zero-shot transfer and translate-test with non-overlapping bootstrap intervals. On MHC's hateful functionalities, especially slur use (F7, as low as 43.3% for XTC in Spanish), it must lose no more than one point.

Why not just…?

Why not just translate the tweet into English first? This is the real threat. If translate-test matches us on the interjection subsets, the honest conclusion is that sentence translation already solves the sense problem. Our contribution would then shrink to the packaging-mismatch analysis. The file offers indirect reasons to expect partial failure: translating datasets misses dialect-specific slurs and references (Exploring Large Language Models for Hate Speech Detection in Rioplatense Spanish). Rewriting the whole sentence also risks normalizing real slurs, which the MHC slur and counter-speech functionalities will expose. The head-to-head on the same subsets settles the question. A combined system that applies our rewrites before translation is also tested.

Why not just mask profanity? Masking removes the false positives but also removes real abuse. Words like puta are hateful in some contexts and exclamatory in others, and that context dependence is the whole problem. The random-substitution ablation and the MHC "hate expressed via profanity" functionality test directly whether sense-aware rewriting beats blind deletion. If masking wins on both, the retrieval stage is unnecessary.

Why not just retrieve labeled examples, as in nearest-neighbor augmentation? Sentence-level neighbors are chosen by overall similarity. They cannot tell a zoological balena from a pejorative one. They also import conflicting label definitions, and the reported degradation past about 2,000 examples is the symptom (Data-Efficient Hate Speech Detection via Cross-Lingual Nearest Neighbor Retrieval). Our zero-shot adaptation of it is a baseline. If it closes the interjection gap, our concept-level design is not needed.

Why not just prompt a large multilingual LLM? LLMs may know Italian expletives. But CReHate reports a 6.79% accuracy gap for GPT-4 between cultures even within English (CReHate). Prompting also leaves no inspectable record of which sense the model assumed. We include the LLM baseline. If it dominates, our rewriting can still be applied as a pre-processing step, and we report whether it helps the LLM.

Proposed Experiments

Training uses English HatEval and English AMI only. Zero-shot targets are HatEval Spanish, AMI-2018/2020 Italian and HaSpeeDe Italian. We measure macro-F1, and accuracy on the taboo-interjection subsets annotated following the zero-shot diagnostic paper. MHC is run across all ten languages, scored per functionality. Sense retrieval is evaluated directly on PejorativITy's word-level labels, using pejorative-versus-neutral accuracy of the retrieved concept's anchor. That gives a check on the mechanism independent of the classifier.

The decisive checks are four. Interjection accuracy on AMI Italian must rise well above 3%. Recovering half the gap to 99% would be a strong result. Full-index retrieval must beat target-wordnet-only retrieval, which shows that concepts recovered from other languages carry the effect. Immigrant-targeted HaSpeeDe performance and MHC slur accuracy must not fall. The per-language rate of mismatched packaging must correlate with the per-language zero-shot gap across the MHC languages.

Coverage is the main risk. Wordnets lag behind slang, and some exclamatory senses may exist in no wordnet. We therefore report the fraction of trigger tokens that fall below the retrieval floor. That number marks the limit of what any structured concept index can do here.