PaperlayerView exact revision on arXiv

Source-linked AI summary

From Triage to Impact: A Multi-Hop Reasoning Benchmark for Explainable Persian Hate Speech

Ghazal Kalhor, Sima Kamandloo, Mehrsa Shamsian, Behnam Bahrak

arXiv:2610.20370v1cs.CL

TL;DR

Persian hate-speech analysis remains constrained by limited fine-grained data and approaches that often use coarse or fragmented labels. This paper introduces a human-validated, seven-hop benchmark and evaluates LLMs across its tasks. Performance generally declines as reasoning becomes more fine-grained, while model-family advantages vary across tasks.

  • Problem

    Persian hate-speech analysis has limited fine-grained data, and existing approaches often use coarse or fragmented labels that do not capture how hate is expressed or framed.

    Source: §1 · p. 1

  • Method

    The paper constructs a human-validated Persian hate-speech benchmark with seven reasoning hops and evaluates state-of-the-art LLMs on it.

    Source: §1 · pp. 1–2

  • Results

    Performance generally declines on finer-grained reasoning tasks; proprietary models generally outperform open-weight models, while reasoning-oriented models show task-dependent benefits.

    Source: §1 · p. 2, §5 · p. 8

  • Takeaways & Limitations

    The findings highlight challenges for reliable, interpretable hate-speech analysis and offer guidance for culturally informed moderation of Persian-language social media.

    Source: §5 · p. 9

    The benchmark covers Persian content from X and Instagram, balances four triage categories rather than natural social-media prevalence, and is relatively small.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Hate speech has become increasingly prevalent on social media, raising concerns about its detection, interpretation, and moderation. Persian, a relatively low-resource language, lacks sufficient annotated data and computational resources for fine-grained hate speech analysis. In this study, we introduce a fine-grained, multi-hop benchmark for explainable Persian hate speech analysis, covering content triage, target identification, hate tactic and moral foundation analysis, textual evidence extraction, harmful narrative inference, and severity assessment. By evaluating state-of-the-art large language models (LLMs), we find that performance generally declines as tasks require more fine-grained semantic and inferential reasoning. Proprietary models generally outperform open-weight models, while reasoning-oriented models show no consistent advantage. Incorporating intermediate reasoning steps substantially improves severity prediction compared with one-shot assessment. Our analysis further reveals that Incitement to Violence or Harm is most strongly associated with higher severity, while political-affiliation targets are particularly associated with subversion-based moral framing. Overall, this work provides insights into LLM-based hate speech analysis and supports more accountable moderation of Persian-language social media. The benchmark is publicly available at https://github.com/kalhorghazal/From-Triage-to-Impact.

1 Introduction

The introduction motivates fine-grained Persian hate-speech analysis and presents a multi-hop benchmark that combines interpretable reasoning with human-validated annotations. Its reported findings show performance variation across tasks and model types, alongside associations between hate-speech features and severity or moral framing.

1 Introduction

  • Persian discourse carries real-world stakes across geopolitical contexts, including migration tensions, religious marginalization, and gender and political polarization.Source: §1 · p. 1
  • Persian hate-speech research is low-resource and often uses coarse or fragmented annotation, limiting analysis of how hate is expressed, framed, and potentially consequential.The introduction notes that implicit, multi-layered, morally framed hate cannot be reliably addressed using surface-level features alone.Source: §1 · p. 1
  • The proposed framework structures Persian hate-speech moderation as sequential, human-interpretable hops, with each step building on prior analysis.The authors present this as an alternative to post-hoc explanations, intended to support traceable and accountable decisions.Source: §1 · p. 1
  • The benchmark combines Moral Foundations Theory with seven hops spanning content triage, target and tactic identification, moral analysis, evidence, narrative, and severity.Initial annotations use Gemini 3.5 Flash and are validated against independent human annotations.Source: §1 · p. 2
  • The evaluation uses eight LLMs to examine performance across reasoning stages and model types, the effect of step-wise analysis on severity prediction, and tactic–severity associations.Source: §1 · p. 2
  • Performance generally declines on finer-grained reasoning tasks; proprietary models generally outperform open-weight models, while reasoning-oriented models vary in benefit across tasks.Intermediate analyses improve severity prediction over one-shot assessment, with variation across models; Incitement to Violence or Harm is most strongly associated with higher severity, and political-affiliation targets with subversion-based framing.Source: §1 · p. 2

2 Benchmark Construction

The benchmark combines existing Persian social-media data with newly collected X posts, then assigns multi-hop annotations and checks them against human judgments. Agreement and human–LLM alignment vary by task, with open-ended reasoning tasks presenting greater challenges.

2 Benchmark Construction

  • The dataset contains 1,200 texts—348 from prior work and 852 newly collected—with 300 instances in each of four initial content-triage categories.Source: §2 · p. 2
  • The 852 new posts were manually searched on X using Persian-language queries, then screened for triage categories and anonymized.The collection strategy targeted posts after the period covered by earlier datasets to reduce potential overlap and capture more recent discourse.Source: §2 · pp. 2–3
  • Gemini 3.5 Flash generated hop annotations using chain-of-thought prompts; two native Persian-speaking annotators independently labeled a stratified sample of 100 texts.Source: §2 · p. 3
  • Human agreement is higher for initial triage and severity classification, but lower for open-ended tasks, especially narrative inference and impact assessment.Agreement also varies across categorical and multi-label annotations.Source: §2 · p. 3
  • The first author resolves annotator disagreements to establish gold labels; Gemini alignment is high on most closed-ended tasks but lower on open-ended generation.The passage cautions that automatic metrics may miss semantic similarity, particularly in low-resource Persian.Source: §2 · p. 3

3 Experimental Setup

The experimental setup compares multilingual proprietary and open-weight models using independently executed, controlled prompts. The model selection supports comparisons by access type and reasoning capability.

3 Experimental Setup

  • The evaluation includes eight multilingual LLMs: four proprietary models and four open-weight models, all officially supporting Persian.Source: §3 · p. 3
  • The model set enables comparisons across proprietary versus open-weight access and reasoning versus instruction tuning.Source: §3 · p. 3
  • All models receive one-shot prompts, with chain-of-thought added to the multi-hop annotation prompt; temperature is 0 and top-p is 1.Each prompt is executed independently.Source: §3 · p. 3

4 Main Results

Model performance varies across the multi-hop Persian hate-speech tasks, with several fine-grained tasks proving more difficult and reasoning benefits differing by task. The analyses also identify associations between specific hate tactics and severity, and between identity bases and moral foundations.

4.1 Model Evaluation

  • Proprietary models lead triage and target and identity identification, while open-weight models show larger micro–macro gaps on less frequent categories.Reasoning variants show a recall advantage in triage, but their effects on target and identity identification vary by model and metric.Source: §4.1 · pp. 3–4
  • Tactic detection favors frequent labels, with mixed reasoning effects; moral-foundation identification is harder overall, and reasoning offers no consistent advantage.Proprietary models’ advantage on moral foundations is less uniform than on some other tasks; adapted F1 remains 71.2–89.8%.Source: §4.1 · pp. 4–5
  • Tactic rationales align best and moral rationales worst; BERTScore exceeds Jaccard, and proprietary models show their clearest advantage on moral rationales.Maximum similarity also exceeds average similarity, while open-weight models remain relatively competitive on target and tactic rationales.Source: §4.1 · p. 6
  • Proprietary models generally lead harmful-narrative similarity, though DeepSeek R1 and V3 are competitive; chrF++ trails semantic metrics, and reasoning gains vary by model and metric.DeepSeek R1 is close to GPT-5.2 on BERTScore F1 despite lower cosine similarity.Source: §4.1 · p. 6
  • Multi-hop analysis raises weighted κ for every model but lowers macro F1 for three; severity lags triage, while GPT-5.2 leads all potential-impact metrics.Macro F1 improves for five models and decreases for Grok, o3, and Qwen3 Think.Source: §4.1 · pp. 6–7

4.2 Correlation Analysis Across Reasoning Hops

  • Incitement to Violence or Harm has the strongest association with higher severity: 70% of its instances are Severe, compared with 51% for Calls for Discrimination or Exclusion.Severity associations are significant for five of six tactics; Conspiracy Theories is the exception.Source: §4.2 · p. 7, Figure 3 · p. 8
  • Political Affiliation has its strongest identity-base association with Subversion; sparse identity bases have more limited, variable patterns that warrant caution.Pairwise tests also find strong co-occurrences for Harm with Personal Identity and Organizational/Institutional Affiliation, and Betrayal with Political Affiliation.Source: §4.2 · p. 7, Figure 2 · p. 8
  • Target-group recall varies substantially: Migrants & Ethnic/Religious Minorities is well detected by several models, while Human Rights, Media & Academic Activists is consistently difficult.Recall exceeds 80% for five named models on the former category, but stays below 61% for every model on the latter and below 50% for all open-weight models.Source: §4.2 · pp. 7–8

5 Conclusion

The benchmark evaluation finds substantial variation across reasoning stages: performance generally declines on finer-grained tasks, and proprietary models generally outperform open-weight models. Intermediate analyses improve ordinal agreement across all models, while severity and moral-framing patterns vary with tactic and target.

5 Conclusion

  • Intermediate analyses improve weighted Cohen’s κ for all models, while macro F1 improves for five of eight models.The gains suggest more consistent improvement in ordinal agreement than in classification performance.Source: §5 · p. 8
  • Performance generally declines as tasks require finer-grained reasoning, and proprietary models generally outperform open-weight models.Benefits from reasoning-oriented models vary across tasks.Source: §5 · p. 8
  • Incitement to Violence or Harm has the strongest association with higher severity, while political-affiliation targets are particularly associated with subversion-based moral framing..Source: §5 · p. 8
  • Uneven performance across target groups highlights differences in LLM sensitivity to hate speech directed at distinct social categories.Figure 4 reports hate-speech detection recall across targeted groups for each LLM.Source: §5 · p. 9, Figure 4 · p. 9

Limitations

The study’s scope and dataset design limit how broadly its results represent Persian hate speech in real-world social media.

Limitations

  • The study covers Persian content from X and Instagram, so its assessment does not extend to other languages or platforms.The authors identify broader language and platform coverage as future work.Source: Limitations · p. 9
  • Balancing the four initial content-triage categories means the benchmark may not reflect their natural prevalence on social media.The authors propose evaluation on randomly sampled social media data to better represent real-world distributions.Source: Limitations · p. 9
  • The relatively small benchmark may limit coverage of less frequent target groups, identity bases, and hate tactics.Larger and more diverse benchmarks could broaden coverage and support more robust evaluation.Source: Limitations · p. 9

A Data Sources and Their Statistics

The appendix describes the datasets used to construct the benchmark and the two Persian-speaking annotators who independently validated automatically generated annotations.

A Data Sources and Their Statistics

B.1 Annotators

  • Two annotators, whose native language and annotation language were Persian, independently validated automatically generated annotations..Source: §B.1 · p. 10
  • The annotators were undergraduate Computer Engineering students with basic NLP coursework and received course-related research assistance as compensation..Source: §B.1 · p. 11

B.2 Annotation Task

Annotators independently applied a seven-step process to each text, moving from triage and target analysis through evidence extraction, narrative inference, and severity assessment.

B.2 Annotation Task

  • For each text, annotators independently labeled content triage, target groups and identity bases, hate tactics, and moral foundations..Source: §B.2 · p. 11
  • They then extracted textual evidence for the target, tactic, and moral foundation, described the harmful narrative, and assessed severity and potential impact..Source: §B.2 · p. 11

B.3 Key Instructions

  • Annotators worked without access to each other’s or Gemini’s annotations, selected all text-supported labels, and grounded narrative and impact responses in the text.For evidence extraction, they selected the shortest span directly supporting the annotation; external AI tools were prohibited.Source: §B.3 · p. 11

C Human Validation Results

The appendix identifies the human-validation tables and documents prompts for keyword detection and multi-hop annotation. The keyword prompts distinguish offensive language from dangerous speech, while the multi-hop template covers a broader analytical task.

C Human Validation Results

D.1 Keyword Detection Prompts

  • Keyword-detection prompts cover hate speech, offensive language, and dangerous speech, with offensive language distinguished from identity-targeted attacks and dangerous speech tied to incitement of violence or serious harm.The prompts ask models to extract relevant keywords from Persian text.Source: §D.1 · p. 12, Figure 5 · p. 12, Figure 6 · p. 12

D.3 Multi-Hop Prompt

  • The multi-hop analytical annotation prompt is presented in Persian and in English translation.Its stated task covers hate-speech detection, targeted-group identification, analysis of how hate speech operates, possible reasons for its persuasiveness, and potential impact.Source: §D.3 · p. 12

D.2 Ablation Study Prompt

The appendix presents the prompt template used for one-shot severity classification in the Persian hate-speech ablation study, along with its English translation.

D.2 Ablation Study Prompt

  • Figure 8 presents the prompt template used for one-shot severity classification of Persian hate speech in the ablation study.The supplied passage identifies the prompt’s purpose, not its contents or the ablation’s results.Source: §D.2 · p. 12, Figure 8 · p. 12
  • The Figure 8 prompt template is accompanied by an English translation.The caption identifies the translation but does not provide its wording.Source: Figure 8 · p. 12, §D.2 · p. 12

F Statistical Analysis of Moral Foundation vs. Identity Basis Associations

The appendix defines triage labels for Persian content and presents a multi-hop annotation prompt. It also describes a statistical table testing associations between moral foundations and identity bases.

F Statistical Analysis of Moral Foundation vs. Identity Basis Associations

  • The multi-hop annotation prompt covers hate-speech detection, target-group identification, how hate speech operates, reasons it may persuade, and potential impact.The prompt is presented in Persian and in English translation.Source: F Statistical Analysis of Moral… · pp. 13–14, Figure 9 · p. 13, Figure 10 · p. 14
  • The triage prompt separates offensive language, identity-based hate speech, dangerous speech that incites violence or serious harm, and content that is not harmful.The prompt instructs the model to select exactly one label for the text’s primary nature.Source: F Statistical Analysis of Moral… · p. 14
  • Table 9 reports significant associations between moral foundations and identity bases, tested with instance-level presence/absence indicators and χ2 tests of independence.P-values are Benjamini–Hochberg corrected, and Cramér’s V is reported as the effect size; the caption does not provide association values.Source: Table 9 · p. 15
Loading 2610.20370v1…