Platforms that moderate sexist content usually build one detector on English Reddit and Gab posts and then ship it to other countries, often adding a little local labeled data. We propose to treat this as what management researchers call a localization problem. A system built on home-market assumptions breaks in a host market, and the fix is to find out which assumptions fail before spending effort on repairs. Concretely, we propose a structured audit, borrowed from SWOT analysis, that runs before any local annotation. SWOT is a planning tool that crosses an organization's internal Strengths and Weaknesses with external Opportunities and Threats. In our setting, the internal side is how reliable the detector is on each category of sexism. The external side is which categories the target culture actually produces, including categories the source taxonomy never defined. The audit then decides how a fixed annotation budget is split across categories, and category definitions are used to retrieve candidate posts to label.
The bet is simple and falsifiable. Local data helps only when it lands on the detector's cultural blind spots. The most damaging blind spots are posts the model is confidently wrong about, and methods that pick examples by model uncertainty never select those. If the audit-guided budget beats random and uncertainty-based selection at equal cost on held-out target cultures, the localization lens has earned its place. If it does not, we will have shown that plain few-shot adaptation is all a platform needs.
Core Claim
Two findings in the literature look inconsistent. Generic data augmentation gave no significant gain for English sexism detection on the SemEval-2023 EDOS benchmark, according to Alexa at SemEval-2023 Task 10, which found that ensembling 18 transformers and shortening input length mattered instead. By contrast, targeted data was the strongest lever in two other settings. In HateModerate, which maps Facebook's community standards into 41 rules and builds tests for each, adding rule-specific data cut a hate detector's failure rate from 0.423 to 0.275. In Cross-Cultural Transfer Learning for Chinese Offensive Language Detection, adding just 50 Chinese examples beat both zero-shot transfer and training from scratch.
Our claim is that the deciding variable is where the added data falls relative to the model's category-level failures under a shift in culture. We also claim that a per-category audit, pairing internal reliability with external coverage of the target culture, finds those failures better than the model's own uncertainty does.
The management and SWOT literatures play different roles here. The management localization work motivates the design and supplies no method. China's FDI in BRICS countries documents expansions failing when centrally designed practices collide with host-country institutions. The Hilton business performance management case attributes successful localization to adapting appraisal to local norms, which included requiring documented facts when managers and staff disagreed. Both are qualitative case studies, and we take from them only the design principle: localize the assumptions, not the surface form.
SWOT supplies the structure of the audit. We follow the PEST-weighted power grid evaluation model in making a situational matrix computable instead of a brainstorm. That model scores regions from weighted political, economic, social and technical indicators. We depart from it by grounding every cell in measured quantities on text data, not expert-weighted indicators.
Research Program
The source system is a hierarchical multi-task detector trained on EDOS. EDOS is the SemEval-2023 benchmark of 20,000 Reddit and Gab posts labeled at three levels: sexist or not, four broad harm categories, and 11 fine-grained vectors such as condescending explanations or dehumanization. We follow the recipe of PingAnLifeInsurance at SemEval-2023 Task 10: domain pre-training on 2M unlabeled Reddit and Gab posts, then joint training of three classification heads on a DeBERTa-v3-large encoder.
The target settings are the non-Western sexism resources available to us. The first is the Indian Twitter misogyny taxonomy of Dehingia et al., which has five categories, including "justifying violence/dismissing inequality" and "asserting inferiority/rigid roles." The second is the Spanish half of EXIST, a bilingual tweet dataset with annotator demographics, used through Analysis of Annotator Demographics in Sexism Detection. The third is the gender portion of COLD, a Chinese offensive-language dataset. The fourth is Malayalam comments targeting women, from the resources behind CUET_Novice@DravidianLangTech 2025, where the labels support it.
The audit has three steps. First, we build a crosswalk between each target taxonomy and the EDOS hierarchy. Annotators map each target category to the nearest EDOS vector, or mark it as unmapped. Second, we label a small seed of 50 target examples and use it to estimate, for each target category $c$, the source model's reliability $\hat r_c$ (its recall on that category) and the category's prevalence $\hat p_c$ in the unlabeled target pool. Unmapped categories get a pessimistic prior on reliability, because the source model has no head for them. Third, the remaining budget $B$ is allocated in proportion to the expected mass of missed harm:
$$b_c = B \cdot \frac{\hat p_c\,(1-\hat r_c)}{\sum_{c'} \hat p_{c'}\,(1-\hat r_{c'})},$$
where $b_c$ is the number of examples labeled for category $c$.
In SWOT terms, a category with high prevalence and low reliability is a weakness exposed to a threat. A category that is well covered but rare needs little budget. Candidates for each category are retrieved from the unlabeled pool with the category's written definition, using an LLM or embedding retriever, independent of the detector's confidence. This follows HateModerate's rule-guided retrieval. It is the ingredient that can reach posts the detector confidently scores as benign, which matters because the EDOS overview reports that false negatives from subtle animosity dominate top systems' errors.
A secondary strand concerns whose local standard is being learned. EXIST shows about a 10-point gap in labeling rates between Spanish-speaking annotators (49% sexist) and English-speaking annotators (41%). We will therefore evaluate EXIST both against majority labels and against each language group's labels. This lets us check whether audit-selected data moves the model toward one group's standard. The step mirrors the Hilton case's insistence on documenting disagreements rather than averaging them away.
Baselines
The naive reference point is the EDOS-trained detector applied to target data with no localization, mapped through the crosswalk. This number shows how large the gap is. For scale, the cross-cultural transfer study reports that an English-trained XLM-R large model reached only 28.94 F1 zero-shot on COLD, against 63.48 when trained on Korean data. No reported result was found on Lacuna for EDOS-trained detectors on the Indian misogyny taxonomy or on fine-grained Spanish EXIST categories, so we measure these ourselves.
The strongest existing method on the source task is the multi-task system of PingAnLifeInsurance. It reports Macro-F1 of 0.8746 on binary Subtask A, 0.7094 on Subtask B and 0.5308 on Subtask C. The best Subtask B and C scores in the competition overall were 0.7326 and 0.5606, per the EDOS overview. We use this architecture as the backbone for every localization condition, so that differences come only from which target examples are labeled.
The strongest competitors for localization are two. The first is random sampling fine-tuning: the same seed, the same total budget, examples drawn uniformly from the target pool. The 50-example gains in the COLD study make this a serious threat. The second is uncertainty-based active learning, which labels the examples with the highest predictive entropy. All conditions share the same budgets (50, 200 and 1,000 labels beyond the seed), the same compute, and the same hyperparameter search of 20 trials tuned on a target dev split. Every condition is run with five seeds.
Ablations isolate the key ingredients. The first keeps the allocation formula but draws candidates by model uncertainty rather than definition-guided retrieval. The second keeps retrieval but allocates uniformly across categories. The third sets $\hat p_c$ uniform, which drops the external side. The fourth drops the unmapped-category prior.
Beating the baselines means a statistically reliable gain in target Macro-F1 over both random and uncertainty sampling at the 200-label budget, across paired seeds. The gain should be concentrated in recall on unmapped and low-reliability categories, and it must come without raising the false-positive rate on benign target posts.
Why not just…?
Why not just label random target data and fine-tune? This is the real threat. The COLD study shows that a handful of in-culture examples already moves performance sharply, and the gain follows a logarithmic curve, so random data may capture most of it. Our experiment settles this directly by plotting both learning curves. If the audit's advantage vanishes by 200 labels, we will report that the audit only helps in the smallest-budget regime, or that it does not help at all. We expect random sampling to under-sample rare but severe categories such as justification of violence, so the per-category recall comparison is where the question will be decided.
Why not just use uncertainty-based active learning? Uncertainty sampling finds examples near the decision boundary. The failures that matter under cultural shift are confident misses: posts expressing a local form of sexism that the source model scores as clearly benign. The EDOS overview reports that false negatives from subtle animosity dominate errors even within the source culture. The first ablation, which keeps the allocation but selects by uncertainty, tests exactly this. If uncertainty selection matches definition-guided retrieval, our mechanism story is wrong.
Why not just translate target posts into English and run the English detector? The cross-cultural transfer study ran this control and found that English models still missed offensive meaning in translated Chinese and Korean text, especially categories tied to local social tensions. We include translate-then-classify as a cheap baseline for completeness, but the evidence already predicts it will trail every condition that uses local labels.
Why not just prompt a large language model with the target taxonomy's definitions? This baseline requires no annotation budget. General moderation systems have failed badly outside their home domain: ExpGuard reports 0.3–0.6% F1 for commercial moderation tools on specialized-domain prompts, although its failure concerns domain jargon, not culture. We run a zero-shot LLM classifier given definitions as a baseline. If it matches the fine-tuned conditions on unmapped categories, the practical conclusion is that the audit should route those categories to the LLM rather than to annotators. That would still be an audit-driven decision, but a weaker claim than ours.
Proposed Experiments
The main experiment trains the EDOS source model and then localizes it to each target culture under every baseline and ablation, across three budgets and five seeds. We report target Macro-F1, per-category recall split into mapped and unmapped categories, and false-positive rate on benign posts.
Two further checks decide whether the idea stands. The first tests whether the audit's cheap estimates predict where localization gains appear: the product $\hat p_c(1-\hat r_c)$ computed from the 50-example seed should rank-correlate with the per-category F1 gain actually achieved by random fine-tuning. If it does not, the audit carries no signal, and the proposal fails regardless of headline scores. The second, on EXIST, compares models localized under majority labels with models localized under each language group's labels. It measures how much of the gain reflects one group's standard, which tells a platform whether "localized" means "matched to a local annotator pool" or something more stable.
A positive result would give platforms a cheap procedure, run before any annotation, for deciding where local labeling effort should go. It would also explain why targeted data works where generic augmentation did not. A negative result would show that random few-shot adaptation is sufficient, which is itself useful to know before building more elaborate localization machinery.
References
- Alexa at SemEval-2023 Task 10: Ensemble Modeling of DeBERTa and BERT Variations for Identifying Sexist Text
- HateModerate: Testing Hate Speech Detectors against Content Moderation Policies
- Cross-Cultural Transfer Learning for Chinese Offensive Language Detection
- China’s foreign direct investment in BRICS countries: status, problems and prospects
- A Case Study on the Business Performance Management of Hilton Hotels Corporation
- Research on Comprehensive Evaluation Model of Regional Power Grid Construction Cost Based on Combined Weight
- PingAnLifeInsurance at SemEval-2023 Task 10: Using Multi-Task Learning to Better Detect Online Sexism
- Violence against women on Twitter in India: Testing a taxonomy for online misogyny and measuring its prevalence during COVID-19
- Analysis of Annotator Demographics in Sexism Detection
- CUET_Novice@DravidianLangTech 2025: Abusive Comment Detection in Malayalam Text Targeting Women on Social Media Using Transformer-Based Models
- SemEval-2023 Task 10: Explainable Detection of Online Sexism
- ExpGuard: LLM Content Moderation in Specialized Domains