Source-linked AI summary
RAGing Against Misogyny: Improving Implicit Misogyny Explanations
Ivana Crescenzi, Arianna Muti, Maarten Sap, Tommaso Caselli, Debora Nozza
TL;DR
Implicit misogyny often depends on background knowledge, making its detection and explanation challenging for language models. This paper evaluates an English–Italian retrieval-augmented framework using encyclopedic, slang, and feminist sources. RAG improves detection on IMPLICIT-MIS but not SBIC+, while human evaluators prefer RAG explanations in both languages.
Problem
Implicit misogyny can require background knowledge to interpret, while LLMs may rely on shallow semantic associations and produce generic or unsupported explanations.
Source: §1 · p. 1Method
The paper evaluates an English–Italian RAG framework that retrieves contextual evidence from encyclopedic, slang, and feminist sources to detect and explain implicit misogyny.
Source: §1 · pp. 1–2, §5 · p. 3Results
Takeaways & Limitations
Feminist literature emerges as the most consistently useful source for detection and explanation.
Source: §9 · p. 9Human evaluation covers a subsample of 100 instances per language, so preference and faithfulness results are diagnostic rather than exhaustive.
Source: Limitations · p. 9
Abstract
from arXiv · showhide
Implicit misogyny often requires background knowledge to be understood. To address this challenge, we propose a Retrieval-Augmented Generation (RAG) framework for detecting and explaining implicit misogyny in English and Italian, leveraging encyclopedic, slang and feminist knowledge sources. Using the ImplicIT-Mis and SBIC+ datasets, we evaluate the contribution of each knowledge source to detection and explanation quality, investigating also whether RAG addresses known LLM failure modes in implicit misogyny. Results show that RAG improves recall on ImplicIT-Mis, but does not improve detection performance on SBIC+. However, human evaluation shows that RAG-based explanations are preferred over no-RAG explanations in both languages. Feminist literature emerges as the most consistently useful source for detection and explanation.
1 Introduction
Implicit misogyny can rely on coded, context-dependent meaning that requires background knowledge, leaving LLMs prone to shallow associations and unsupported explanations. The paper proposes an English–Italian RAG framework and finds dataset-specific detection gains alongside more-preferred explanations in both languages.
1 Introduction
- Implicit misogyny can use indirect, coded, context-dependent language that normalizes hostility toward women while remaining deniable.Interpreting it may require knowledge of sarcasm, cultural references, pragmatic cues, or insinuations about women’s roles and attributes.Source: §1 · p. 1
- LLMs may rely on shallow semantic associations and produce generic or unsupported explanations, while moderation systems under-moderate implicit hate speech.Structured explanations can make moderation decisions faster and more usable for moderators.Source: §1 · p. 1
- The framework retrieves contextual evidence to generate explanations for implicit misogyny in English and Italian online discourse.The study also examines whether posting context helps detection, which knowledge sources best support explanations, and whether RAG addresses known LLM failure modes.Source: §1 · pp. 1–2
- RAG improves detection on IMPLICIT-MIS but not SBIC+, while human evaluators prefer its more specific explanations over the baseline in both languages.Feminist literature achieves the highest recall on IMPLICIT-MIS.Source: §1 · p. 2
- The paper introduces a multi-source RAG framework, evaluates explanation quality through human annotation, and analyzes whether RAG mitigates known LLM weaknesses.The authors also release their pipeline and analysis.Source: §1 · p. 2
2 Related Work
Prior work has integrated external knowledge into harmful-language detection, but has mainly focused on classification. Research on implicit-hate explanations and reasoning motivates testing whether retrieval can ground both detection and explanations in relevant context.
2 Related Work
- Knowledge-based approaches have been applied to harmful-language detection, but prior retrieval work mainly targets classification rather than explanations.The open question is whether retrieval can explain why a text is uncivil.Source: §2 · p. 2
- Work on implicit-hate explanations reports that LLMs can rely on semantic associations instead of robust reasoning, while metaphorical hate also challenges LLMs.A separate line uses multi-persona debate for structured explanations of implicit sexism.Source: §2 · p. 2
3 External Knowledge Sources
The paper builds three knowledge sources to target different forms of implicit meaning: encyclopedic information, slang, and feminist literature. The sources are intended to cover factual references, offensive vocabulary, and theoretical or online ideological concepts.
3 External Knowledge Sources
- The three knowledge sources target named-entity and factual references, informal or Incel-specific vocabulary, and feminist concepts such as objectification and victim-blaming.Wikipedia, slang, and feminist literature address these respective needs.Source: §3 · p. 2
3.1 Wikipedia Knowledge Base
- Wikipedia retrieval filters Wikidata for human entities and topics in feminism, femicide, and violence against women before indexing relevant pages.The method first searches the domain-filtered index and falls back to live Wikipedia Search when no match is found.Source: §3.1 · pp. 2–3
3.2 Slang Knowledge Base
- The slang source selects Wiktionary definitions marked vulgar, derogatory, or offensive, using about 530 Italian entries and 22,000 English entries.English entries combine English slang, internet slang, and incel slang categories.Source: §3.2 · p. 3
3.3 Feminist Literature Knowledge Base
- The feminist-literature source combines Manne’s philosophical account of misogyny as social enforcement with an empirical analysis of online anti-feminist movements.Together, the books provide theoretical and online-discourse perspectives on implicit misogyny.Source: §3.3 · p. 3
4 Datasets
The experiments use an Italian corpus of implicitly misogynistic Facebook comments and an English dataset assembled from filtered misogynistic instances. The datasets differ in how implicit their examples are, a distinction relevant to evaluating external knowledge.
4 Datasets
- The experiments cover Italian IMPLICIT-MIS and English data drawn from the misogynistic subsections of SBIC and IMPLICIT HATE.Source: §4 · p. 3
- IMPLICIT-MIS contains 1,120 Italian Facebook comments from women-related news articles, all annotated as implicitly misogynistic.Its references to news events, public figures, and cultural context make it a testbed for contextual information in a non-English setting.Source: §4 · p. 3
- SBIC+ contains 2,344 English messages targeting women and is less implicit than IMPLICIT-MIS, with more slurs and explicit jokes.It was constructed from filtered instances of SBIC and the Implicit Hate Corpus.Source: §4 · p. 3
5 Methodology
The pipeline builds source-specific evidence for each input through entity-driven retrieval, compression, reranking, and context-driven filtering. Retrieval coverage differs by source and dataset, with feminist literature providing the most balanced coverage across the two datasets.
5 Methodology
- The pipeline filters entities, retrieves source-specific candidates, then compresses and reranks evidence before a second context-driven compression.The second compression is applied to Wikipedia and literature, while slang definitions are left untouched; thresholds were selected empirically through qualitative inspection.Source: §5 · pp. 3–4
- Entity extraction combines spaCy with name-recombination and capitalized-sequence rules, then filters entities using a predefined domain-keyword set.The keyword set was established before dataset inspection to avoid overfitting and support generalization.Source: §5 · p. 3
- Wikipedia candidates come from a local index and live API, slang candidates from matched Wiktionary terms, and literature retrieval selects similar two-sentence units above a similarity threshold.Wikipedia is reranked per entity; slang entries are reranked against the input, while literature retrieval and reranking coincide.Source: §5 · pp. 3–4
5.1 Retrieval Coverage
- 87.4% Wikipedia coverage on ImplicIT-Mis contrasts with 3.4% on SBIC+; slang coverage reverses the pattern at 9.7% versus 57.1%, while feminist literature covers 29.6% versus 40.8%.The coverage pattern reflects named entities in ImplicIT-Mis, derogatory vocabulary in SBIC+, and more balanced feminist-literature coverage.Source: §5.1 · p. 4, Table 1 · p. 4
6 Experiments
Experiments use a fixed instruction-tuned model and evaluate detection by recall, including prompts in both Italian and English for the Italian dataset. Human evaluation assesses explanation faithfulness and preference, with two expert annotators rating explanations independently.
6 Experiments
- Experiments use LLAMA-3.1-8B-INSTRUCT with temperature 0.3 and a 256-token generation limit.Other model parameters remain at their defaults.Source: §6 · p. 4
- Because all dataset instances are misogynistic, detection is evaluated using recall; Italian data are tested with prompts written in Italian and English.The text-only baseline uses the message alone, while the context setting adds the related news headline.Source: §6 · pp. 4–5
6.1 Qualitative Evaluation of Generated Explanations
- Faithfulness scores whether an explanation identifies the implicit assumption and is supported by the message and, when applicable, retrieved evidence, on a 0–2 scale.Two annotators score each explanation; the per-knowledge-base score is the sum of scores divided by the number of scores.Source: §6.1 · p. 5
- Two expert author-annotators rated explanations independently, with initial agreement varying by language and dimension and higher agreement after discussing and re-annotating disagreements.Initial Cohen’s κ was 0.505 for Italian and 0.454 for English on faithfulness, and 0.333 for English and 0.286 for Italian on preference.Source: §6.1 · p. 5
7 Detection Results
Adding retrieved knowledge improves detection over the headline baseline on ImplicIT-Mis, but no retrieval source beats the baseline on SBIC+. The observed dataset contrast cannot be attributed to language alone because language and dataset vary together.
7 Detection Results
- 0.923 with literature-based retrieval and 0.916 with slang-based retrieval are the best reported ImplicIT-Mis results, above the 0.891 headline baseline; no SBIC+ retrieval source beats its 0.966 baseline.Table 3 compares retrieval sources against the headline baseline for ImplicIT-Mis and the baseline for SBIC+.Source: §7 · pp. 5–6, Table 3 · p. 6
- Language and dataset vary jointly, and differences in implicitness, retrieval coverage, and knowledge needs may also contribute to the contrast in RAG effectiveness.The results therefore do not establish an intrinsic difference between Italian and English.Source: §7 · p. 6
8 Explanation Quality
RAG explanations were more faithful on IMPLICIT-MIS but not SBIC+, and annotators preferred them over the baseline in both languages. Feminist literature was the most consistently preferred source, although some explanations still missed pragmatic meaning.
Faithfulness
- IMPLICIT-MIS aggregated faithfulness was 0.781 vs. baseline 0.590; SBIC+ was 1.310 vs. 1.345, with no significant RAG-baseline difference on SBIC+.For IMPLICIT-MIS, the mean RAG score was 0.788 vs. 0.590 for baseline (∆= +0.198; p < 0.0001).Source: Table 4 · p. 6, §8 · p. 6
Preference
- RAG explanations received 121 vs. 51 preferences in Italian and 154 vs. 95 in English.Table 5 counts can include a text in both columns when annotators tie or disagree.Source: §8 · p. 6, Table 5 · p. 7
- In mutually exclusive comparisons, RAG was preferred 88 vs. 18 times in Italian and 86 vs. 27 in English; English had more ties, while Italian had more neither-preferred cases.The higher neither-preferred count for IMPLICIT-MIS indicates that explaining implicit misogyny in Italian remained challenging even with retrieval.Source: §8 · pp. 6–7, Table 6 · p. 7
Knowledge-source preferences
- Literature had the highest normalized preference: 0.408 on IMPLICIT-MIS and 0.538 on SBIC+, versus baseline scores of 0.215 and 0.460.Wikipedia scored below the English baseline at 0.361.Source: Table 7 · p. 7, §8 · p. 7
Retrieved context in explanations
- Wikipedia resolves cultural references, feminist literature supports ideological and ironic inference, and slang retrieval recovers community-specific derogatory meanings.Examples include a cultural reference to Elena Cecchettin, patriarchy as a structural cause, and the slang term “roastie.”Source: §8 · p. 7
Failure modes
- In selected diagnostic examples, RAG addressed 21 Italian and 9 English failure-mode issues, yet some explanations remained lexically related without pragmatic support.The analysis selected 35 Italian and 36 English sentences under its faithfulness or preference criteria.Source: §8 · p. 8
9 Conclusion
The evaluation finds dataset-specific detection gains and more-preferred RAG explanations in both languages. Feminist literature is the most effective contextual source, but some explanations still rely on shallow lexical associations.
Conclusion
- RAG improved detection on IMPLICIT-MIS, especially with feminist literature and slang knowledge, but did not improve detection on SBIC+.The authors’ manual analysis suggests SBIC+ requires less external knowledge than IMPLICIT-MIS.Source: §9 · p. 9
- RAG-based explanations were preferred over baseline explanations more often in both languages, with feminist literature the most effective contextual source.The paper also reports that some explanations rely on shallow lexical associations rather than pragmatic reasoning.Source: §9 · p. 9
Limitations
The study’s conclusions are bounded by dataset composition, subsampled human evaluation, uneven retrieval coverage, potentially irrelevant retrieved evidence, author annotations, and evaluation with a single main model.
Limitations
- The authors cannot guarantee that every SBIC+ instance is genuinely implicit; some English examples contain explicit jokes or slurs.Source: Limitations · p. 9
- Human evaluation covers 100 instances per language, so preference and faithfulness findings are diagnostic rather than exhaustive.Source: Limitations · p. 9
- Retrieval coverage varies substantially across sources and languages, limiting direct source-level comparisons.Source: Limitations · p. 9
- Retrieved evidence can be pragmatically irrelevant when retrieval follows lexical overlap rather than the implicit reasoning needed.Source: Limitations · p. 9
- The authors conducted the annotations, though they are experts in misogynistic language and rated anonymized explanations.Source: Limitations · p. 9
- Main experiments use one 8B model; generalization across model families and larger scales remains for future work.Source: Limitations · p. 9
Ethical Considerations
This section documents ethical practices and methodological details, including source-text handling, AI use and human review, prompt selection, retrieval and compression choices, and Italian examples of literature-informed interpretation.
Ethical Considerations
- Copyrighted texts and their derived index were stored securely for research-team use and are not released.The authors cite the applicable European text-and-data-mining exception for scientific research.Source: Ethical Considerations · p. 9
- Misogynistic examples are presented uncensored for transparency, reproducibility, and accurate analysis, despite the risk of exposing readers to offensive language.The authors state that the examples do not reflect their beliefs and are included only to present and analyze the data.Source: Ethical Considerations · p. 9
- The authors used AI tools for grammar and spelling checks and paraphrasing, then reviewed and edited the content and accept responsibility for it.Source: Ethical Considerations · p. 9
A Prompts
- Prompt variants were compared iteratively, and those yielding the most consistent, well-formatted JSON were retained.Templates distinguish input text, retrieved knowledge, and—where applicable—context accompanying Italian instances.Source: A Prompts · p. 11
- Baseline prompts ask models to infer implied assumptions and classify misogyny, while retrieval prompts add external knowledge such as Wikipedia, slang definitions, or gender-studies literature.The prompt set also includes Italian variants with contextual information and English-language variants for Italian instances.Source: §A.1 · p. 11, §A.2 · p. 11, §A.3 · p. 12
A.14 Italian Literature EN
- Italian literature examples interpret a comment as blaming rape victims who use alcohol or drugs and another as portraying feminists as hysterical and angry.Both examples are labeled as containing the implied assumption.Source: §A.14 · p. 14
B Retrieval Design Choices
- A qualitative review of 200 instances selected similarity thresholds of 0.3 for Wikipedia and slang and 0.6 for feminist-literature retrieval to balance relevance and coverage.Lower thresholds retained substantial irrelevant noise; higher thresholds removed useful evidence.Source: B Retrieval Design Choices · p. 14
- For final sentence-level compression, the authors chose 0.3 because 0.2 retained excessive noise and 0.4 removed potentially relevant information.These thresholds were compared in a second compression stage.Source: B Retrieval Design Choices · p. 14
- Recall on IMPLICIT-MIS (Wikipedia, IT, few-shot) rose from 0.881 without second-stage compression to 0.883 with it, while reducing noise passed to the model.The authors retained the stage because it slightly improved performance and reduced noise in retrieved evidence.Source: B Retrieval Design Choices · p. 14
C Results Across Model Sizes
Retrieval effects vary by model size and setting, while the human-evaluation subsections specify how explanations and annotators were assessed.
C Results Across Model Sizes
- Recall in the Italian few-shot baseline with Wikipedia and no second compression rose from 0.754 without RAG to 0.949 with RAG for the 1B model.Retrieval effects varied across model sizes and settings, so the authors say the pattern requires further testing.Source: Table 9 · p. 14, C Results Across Model Sizes · p. 14
D Human Evaluation Guidelines
- Human evaluators compared explanations on faithfulness to the input and, where applicable, retrieved context, and on overall human preference.Preference considered correctness, clarity, usefulness, and avoiding hallucinated or overgeneralized claims.Source: D Human Evaluation Guidelines · p. 14
D.1 1. Faithfulness / Groundedness
- Faithfulness was scored from 0 for unsupported or hallucinated explanations to 2 for fully grounded explanations, with 1 marking partial grounding.The examples distinguish a fully supported identification of sarcasm and a sexist stereotype from vague or invented explanations.Source: §D.1 · pp. 14–15
D.2 2. Human Preference
- Human preference labels were Prefer, Tie, or None, reflecting whether an explanation was clearly better, similarly good or flawed, or all were wrong.Annotators were asked to consider usefulness for understanding implicit misogyny and avoidance of hallucinated or overgeneralized claims.Source: §D.2 · p. 15
D.3 Inter-Annotator Agreement
- Inter-annotator agreement for faithfulness was reported across subsets and languages in Table 10.The supplied passage identifies the table's scope but provides no agreement values.Source: §D.3 · p. 15, Table 10 · p. 15
D.4 Annotators
- Two paper authors annotated independently, discussed disagreements, clarified guideline interpretations, and independently re-annotated disputed cases.One annotator was a postdoctoral researcher and online-misogyny expert; the other was a master's student in data science and business analytics.Source: §D.4 · p. 15