cs.CLOct 2, 2026

To Jev or Not? Evaluating the Accuracy and Efficiency of Structured Decision Models for Hate-Speech Moderation

Authors: Demetris Paschalides, George Pallis, Marios D. Dikaiakos

Organizations: University of Cyprus, Computer Science Department, Nicosia, Cyprus

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

The scale of online content makes hate-speech moderation challenging, while Large Language Models (LLMs) enable harmful material to be produced and adapted more easily. Moderation therefore requires efficient classifiers that can accommodate different definitions of hate speech. Recent structured decision models accept natural-language criteria and select among specified answers, raising the question of whether they can meet these requirements without task-specific training. We present HATEDECIDE, an evaluation of six decision-model configurations on four hate-speech datasets against specialized moderation, zero-shot, commercial, and supervised baselines. We examine whether supplying a dataset's definition, or decomposing it into multiple questions, improves classification, and we measure their latency and cost. We find that commercial LLMs significantly outperform all decision models on only one dataset. Supplying definitions changes up to 28% of predictions without consistently improving classification, and decomposition significantly improves performance in only 20% of the comparisons. On a diagnostic set of test cases, the best hosted decision model comes within 1.6 macro-F1 points of the best commercial LLM at approximately 97% lower inference cost. These results identify opportunities for inexpensive moderation, while showing that explicit criteria and additional questions do not reliably improve classification.

Figures & tables

Figure 1: Overview of HateDecide . Out-of-the-box models classify texts directly (A), with or without the dataset’s definition, or through definition-specific decomposition (B), while supervised classifiers are trained on one dataset and tested within and across datasets. Each paired comparison is evaluated against gold annotations on identical texts.
DatasetText sourceTrainVal.Test (hate / non-hate)Additional labels and tasks
Dynamic [ 26 ]Texts written by annotators to challenge hate-speech classifiers.28,0703,7704,120 (2,268 / 1,852)Six conduct classes across 3,356 test texts.
MHS [ 20 ]Comments from YouTube, Reddit, and Twitter.26,1013,2453,203 (782 / 2,421)-
HateXplain [ 16 ]Posts from Twitter and Gab.15,3461,9201,924 (594 / 1,330)Three-class labels retain 548 offensive and 782 normal posts.
HateCheck [ 19 ]Texts testing abilities such as recognizing threats and counter-speech.--3,728 (2,563 / 1,165)Referenced group on 3,436 texts and general/directed hate on 2,563 texts.
Table 1: Datasets after label aggregation and deduplication. HateXplain combines offensive and normal as non-hate .
 DynamicMHSHateXplain (proxy)HateCheck
TargetVulnerable, marginalized, or discriminated- against groups; attacks based solely on being male, white, or heterosexual excludedA person or groupA social group or its membersA protected group, or an individual because of membership
IdentitiesOpen-ended, e.g., ethnicity, religion, gender, sexuality, immigration status, disability, classActual or perceived innate characteristicsUnspecified social groupsProtected characteristics, incl. age, disability, gender identity, pregnancy, race, religion, sex, sexual orientation
ConductDerogation, animosity, dehumanization, threats, and support for hateful entitiesPrejudice-driven hostile or malicious languageHatred, or intent to demean, humiliate, or insultAbuse
StanceIntentional, reporting or rejecting excludedNot statedContext decides, in-group use and quoted lyrics not hateCounterspeech, negated hate, reclaimed slurs non-hateful
QuestionsScope, intent, conduct typeTargeting, abuseTargeting, abuseTargeting, abuse (same target)
Hate ifIntent ∧\land (support ∨\lor (scope ∧\land hostile conduct))Targeting ∧\land abuseTargeting ∧\land abuseTargeting ∧\land abuse
Table 2: Dataset-specific definitions and derived decomposition rules. Full definitions appear in Appendix A . In the rules, ∧\land means AND and ∨\lor means OR. Scope, intent, targeting, and abuse denote affirmative component answers. Support denotes selection of the support category, hostile conduct denotes animosity, derogation, dehumanization, or threatening.
ModelAccessBackbone
Structured Decision Models
JevAPIUndisclosed
LayaLocalModernBERT-large
DeciderLocalQwen, 2B
SemIf/Qwen3.5-2BLocalQwen3.5-2B
SemIf/Qwen3.5-4BLocalQwen3.5-4B
Table 3: Evaluated models, access modes, and backbones.
 Dynamic ( N=4,120N{=}4{,}120 )MHS ( N=3,203N{=}3{,}203 )HateXplain ( N=1,924N{=}1{,}924 )HateCheck ( N=3,728N{=}3{,}728 )
ModelABΔ\DeltaABΔ\DeltaABΔ\DeltaABΔ\Delta
Jev.826.843++ 1.6.691.682−- 0.9.738.729−- 0.9.960.957−- 0.4
Laya.655.534−- 12.1.695.732++ 3.7.733.775++ 4.2.701.690−- 1.1
Decider.586.552−- 3.3.801.707−- 9.3.677.706++ 2.8.730.390−- 34.0
SemIf/Qwen3.5-2B.479.534++ 5.6.220.578++ 35.8.416.456++ 4.0.483.411−- 7.2
SemIf/Qwen3.5-4B.674.655−- 1.9.719.681−- 3.8.688.695++ 0.7.782.754−- 2.8
Table 4: Direct (A) and decomposed (B) binary macro-F1 on the full test sets. Both formulations classify the same texts and receive the same dataset-specific definition, together with the decomposition rule and component glossary, so each difference Δ\Delta (percentage points) is paired. Bold Δ\Delta indicates an adjusted interval that excludes zero.
Figure 2: Binary macro-F1 of supervised classifiers trained on the row dataset and tested on the column dataset without adaptation. Outlined cells mark within-dataset evaluation. HateCheck is evaluation-only.
Figure 3: Change in binary macro-F1 (percentage points) from no definition to the dataset-specific definition, on identical texts. Asterisks mark changes whose adjusted interval excludes zero (Section 3.4 ).
ModelConduct (Dyn.)Group (HC)Direction (HC)
Jev.538.989.682
Laya.192.929.694
Decider.313.959.610
SemIf/Qwen3.5-2B.196.939.269
SemIf/Qwen3.5-4B.406.962.594
Bespoke-Nimble, NF4.324.971.567
Table 5: Attribute macro-F1 with defined labels: Dynamic’s six conduct classes, and HateCheck’s referenced group and general-versus-directed hate.
Figure 4: Change in hate recall against change in hate precision from direct to decomposed classification of the same texts, for each model on Dynamic, MHS, and HateXplain. Colours mark gains and losses in macro-F1 (Table 4 ).
Figure 5: Macro-F1 against median complete-decision latency (a) and hosted cost per 1K texts (b) on identical HateCheck examples ( N=3,728N=3{,}728 ). Connected markers show direct (A) and decomposed (B) classification for five models. Both horizontal axes are logarithmic. Local monetary costs are omitted. Results for all evaluated models appear in Appendix H .
Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

ModelCheckpoint or identifierPrecisionInput limitDecision
Jevjev-1.13.0 (hosted API)--Choice question; option scores
Laya1c5edc17FP32, BF16 autocast1,024 tokens (2,048 for B)Option scores with the released temperature scaling
Deciderb37f7e1bBF16512 tokensOption-letter scores, scoring temperature 1.3
SemIf/Qwen3.5-2BQwen 15852e8c , SemIf ca3ba65fBF164,096 tokensNext-letter probabilities
SemIf/Qwen3.5-4BQwen 851bf6e8 , SemIf ca3ba65fBF164,096 tokensNext-letter probabilities
Bespoke-Nimbleadapter 594dfdcf , Qwen3.5-9B c2022362NF4, BF16 compute2,048 tokensLetter-code probabilities
Appendix
Table 6: Model revisions and inference settings. Hash prefixes identify the pinned checkpoints; hosted identifiers do not guarantee unchanged weights.
 DynamicMHSHateXplainHateCheck
ModelCh.Co.Br.Δ\Delta F1Ch.Co.Br.Δ\Delta F1Ch.Co.Br.Δ\Delta F1Ch.Co.Br.Δ\Delta F1
Jev8.74.14.6−- 0.510.65.25.4−- 2.77.96.11.8++ 3.92.81.41.4++ 0.1
Laya11.63.87.8−- 4.48.05.12.9++ 1.18.65.92.7++ 2.78.02.55.5−- 2.8
Decider9.56.13.4++ 3.34.51.82.7−- 1.012.83.79.1−- 4.715.611.44.1++ 5.8
SemIf/Qwen3.5-2B19.17.911.2++ 0.77.13.04.2−- 1.227.922.45.5++ 19.76.32.83.5++ 1.5
SemIf/Qwen3.5-4B20.97.113.8−- 7.216.311.64.7++ 4.120.416.14.3++ 10.311.64.67.1−- 0.8
Appendix
Table 7: Prediction-level effect of the dataset-specific definition relative to no definition, as percentages of texts: predictions that change (Ch.), change from incorrect to correct (Co.), and change from correct to incorrect (Br.). Δ\Delta F1 is the change in macro-F1, in percentage points. For HateXplain, the definition is the Davidson proxy.
 Binary Δ\Delta F1 (pp)HX-3 macro-F1
ModelDyn.MHSHXHCGen.Dav.
Jev−- 0.1−- 2.9++ 5.3++ 0.0.522.532
Laya−- 3.2++ 0.6++ 2.5−- 2.3.582.570
Decider++ 3.6−- 4.7−- 3.3++ 5.4.527.506
SemIf/Qwen3.5-2B−- 1.9++ 17.5++ 23.6++ 4.4.277.279
SemIf/Qwen3.5-4B−- 8.3−- 2.6++ 11.0−- 1.5.405.411
Appendix
Table 8: Effect of the generic definition. Left: change in binary macro-F1 from no definition to the generic definition, on the benchmark population (descriptive). Right: HateXplain three-class macro-F1 under the generic definition (Gen.) and the Davidson proxy (Dav.), on 1,000 posts.
 Median ms95th percentile msTexts/sUSD / 1,000 texts
ModelABABABAB
Jev335.6342.9536.9622.12.562.370.0330.054
Laya48.352.592.693.918.6817.49--
Decider133.6134.6292.5264.96.316.45--
SemIf/Qwen3.5-2B125.3250.5272.1540.46.563.28--
SemIf/Qwen3.5-4B183.7343.8325.4555.64.822.63--
Appendix
Table 9: Complete-decision timing and recorded hosted cost on HateCheck ( N=3,728N{=}3{,}728 ) for direct (A) and decomposed (B) classification. Dashes denote unassigned local monetary cost, not zero cost.
 F1Offensive as
ModelHateOff.Normalhatenormal
Laya.703.316.69246%32%
GPT-6 Sol.694.448.67142%14%
TF-IDF+LR.737.473.706--
DistilBERT.722.448.717--
Appendix
Table 10: Per-class F1 on HateXplain’s three-class task (1,000 posts; 309 hate, 285 offensive, 406 normal), and the share of offensive posts labeled hate or normal.
ModelDatasetABA*B −- A*
LayaDynamic.655.534.650−- 11.7 †
 MHS.695.732.696++ 3.6 †
 HateXplain.733.775.795−- 2.1
SemIf/Qwen3.5-4BDynamic.674.655.718−- 6.3 †
 MHS.719.681.709−- 2.8 †
 HateXplain.688.695.737−- 4.2
Appendix
Table 11: Direct (A), decomposed (B), and threshold-tuned direct classification (A*) macro-F1. B −- A* is in percentage points. † Adjusted interval excludes zero (18 comparisons). ∗ 800 texts per dataset.
More in this topic

Hate Speech Detection

Hate Speech Detection identifies hateful or abusive language, its targets and intent, to support safer and more accountable content moderation.

See all 99 papers

Explore similar work

CardsList
May 26, 2026

Attribute-Based Diagnosis of LLM Alignment with Hate Speech Annotations

Hate speech annotation is costly, subjective, and prone to annotator disagreement, making large-scale dataset construction challenging. We systematically analyze how well large language models (LLMs) align with human judgments across ten theoretically grounded subjective attributes, such as dehumanization, violence, and sentiment, evaluating both small and large variants of Llama 3.1 and Qwen 2.5. Our analysis reveals a consistent split across all models: behaviorally explicit dimensions (insult, humiliate, attack-defend) correlate strongly with human annotations, while evaluative dimensions (respect, sentiment, hate speech) are systematically inverted. Demographic persona conditioning reduces model confidence without improving alignment. Building on these insights, we propose combining attribute-level LLM predictions via a confidence-weighted Ridge regression to reconstruct continuous hate speech scores from the Measuring Hate Speech corpus, achieving R2R^2 of up to 0.71 and outperforming direct prompting baselines, demonstrating that structured attribute decomposition recovers a richer and more human-aligned signal than end-to-end label prediction alone.

Mohammad Amine Jradi, Faeze Ghorbanpour, Alexander Fraser

School of Computation, Information and Technology, TU Munich Munich Center for Machine Learning (MCML)

LLM EvaluationLLM AlignmentLLM-Assisted AnnotationHate Speech Detection
AskPDF
Sep 3, 2026

Evaluating Criterion-Conditioned Behaviour of Large Language Models in Content Moderation

Large language models (LLMs) demonstrate strong performance on standard content moderation benchmarks. However, these benchmarks often aggregate multiple moderation criteria into a single label, making it unclear whether models can disentangle them and reliably apply each criterion when making decisions. To study whether LLMs exhibit criterion-conditioned behaviour, we introduce Diagnostic Evaluation of COntent (DECO), a criterion-independent factorisation of content that enables controlled, criterion-level evaluation. We also introduce pairwise evaluation to compare model outputs across different criteria for the same input. Across four moderation datasets and four LLMs, we find that strong benchmark performance can hide substantial failures at the criterion level. Models struggle most when correct decisions depend not on overall harmfulness, but on the specific aspect of the content that the criterion requires them to assess. Our results highlight a key limitation of current content moderation benchmarks: strong performance on aggregated labels does not provide sufficient evidence that LLMs can reliably evaluate content with respect to individual moderation criteria. These findings call for the development of evaluation methods that explicitly measure criterion-conditioned behaviour.

Danting Zhang, Bei Peng, Robert Loftin

1Independent Researcher · University of Sheffield

LLM Safety BenchmarksLLM EvaluationContent Moderation
AskPDF
Mar 26, 2026

When Hate Meets Facts: LLMs-in-the-Loop for Check-worthiness Detection in Hate Speech

Hateful content online is often expressed using fact-like, not necessarily correct information, especially in coordinated online harassment campaigns and extremist propaganda. Failing to jointly address hate speech (HS) and misinformation can deepen prejudice, reinforce harmful stereotypes, and expose bystanders to psychological distress, while polluting public debate. Moreover, these messages require more effort from content moderators because they must assess both harmfulness and veracity, i.e., fact-check them. To address this challenge, we release WSF-ARG+, the first dataset which combines hate speech with check-worthiness information. We also introduce a novel LLM-in-the-loop framework to facilitate the annotation of check-worthy claims. We run our framework, testing it with 12 open-weight LLMs of different sizes and architectures. We validate it through extensive human evaluation, and show that our LLM-in-the-loop framework reduces human effort without compromising the annotation quality of the data. Finally, we show that HS messages with check-worthy claims show significantly higher harassment and hate, and that incorporating check-worthiness labels improves LLM-based HS detection up to 0.213 macro-F1 and to 0.154 macro-F1 on average for large models.

Nicolás Benjamín Ocampo, Tommaso Caselli, Davide Ceolin

Centrum Wiskunde & Informatica · Amsterdam, The Netherlands · University of Groningen +1

LLM-Assisted AnnotationHate Speech Detection
AskPDF