SPAR-Hate: Auditor-Guided Multi-Perspective Role Reasoning for Bilingual Hate Speech Parsing

Yifan Lyu; Dianqing Lin; Xinran Li; Jiaqi Qiao; Xiujuan Xu

cs.AIAug 22, 2026

SPAR-Hate: Auditor-Guided Multi-Perspective Role Reasoning for Bilingual Hate Speech Parsing

Authors: Yifan Lyu, Dianqing Lin, Xinran Li, Jiaqi Qiao, Xiujuan Xu

Organizations: Dalian University of Technology · Inner Mongolia University

Abstract

Hate speech research has moved from coarse-grained classification towards structured parsing, where systems jointly identify targets, supporting arguments, and target-level labels. Documents with multiple targets, conflicting local readings, or culturally coded language make these bindings difficult to recover. SPAR-Hate is an auditor-guided multi-perspective role-reasoning framework for bilingual hate speech parsing. It decomposes each document into local focus units, elicits evidence-grounded candidates from Victim, Moderator, and Cultural Bystander perspectives, resolves candidate conflicts under grounding and schema constraints, and reassembles sample-level predictions. Experiments on STATE-ToxiCN and a controlled TBO split show gains across local and API backbones, concentrated on strict joint target--argument--label metrics. Controlled comparisons favour separate candidate generation and arbitration for joint parsing. Within the tested pipeline, role-conditioned prompting yields higher exact-match joint scores than repeated generic prompting, while Chinese overlap-based results depend on the backbone. Structured teacher traces also support training a smaller student model.

Figures & tables

Figure 1: Examples of bilingual parsing errors. The English example shows omitted tuples and incorrect target–argument–label binding; the Chinese example shows a missed culturally localised homophonic slur.

Figure 2: Overview of SPAR-Hate. (a) The pipeline segments bilingual input into local focus units, generates role-conditioned candidates, arbitrates under evidence constraints, and reassembles sample-level outputs. (b) A Chinese focus unit illustrates role disagreement and Auditor selection before multi-target reassembly.

Dataset	Public Split	Final Split	Tracks
STATE-ToxiCN	official train/test	6424 / 1605	ZH-main, ZH-quadruple
TBO	public test only	3200 / 800	EN-main

Table 1: Datasets used in the experiments. TBO is repartitioned from its public test set with seed=20260415 .

API zero-shot model	EN	ZH
DeepSeek-v4-Pro	23.18	22.80
GPT-5.4	21.01	28.13

Table 2: Bold marks the highest score in each panel or metric column. Full-test average scores for the added API zero-shot baselines.

Local LLMs
Model	Target		Argument		T-A Pair		T-A-H Tri.		Avg.
	Hard	Soft	Hard	Soft	Hard	Soft	Hard	Soft
0-shot-Qwen3-14B	45.49	54.97	16.24	46.99	10.62	32.73	7.65	23.93	29.83
SPAR-Qwen3-14B (mainline)	48.73	55.75	21.03	56.85	12.36	34.14	8.71	24.97	32.82
SPAR-Qwen3.5-35B	45.84	53.12	20.09	57.95	11.23	35.87	9.26	27.24	32.58
API Models

Table 3: Main results on ZH-main and EN-main . Within each split, the best result in each column is boldfaced and the second-best is underlined.

Backbone	Integrated	SPAR
DeepSeek-v4-Pro	4.87 / 3.89	19.81 / 13.33
GPT-5.4	11.41 / 8.18	18.84 / 15.81

Table 4: Full-test fixed-Phase-1 integrated-prompt controls. The integrated condition retains the three lens descriptions and Chinese lexicon access, but removes separate candidates, clustering, and Auditor arbitration. Bold marks the higher value for each metric within a backbone.

DeepSeek-v4-Pro
Condition	T-A Pair		T-A-H Tri.
	Hard	Soft	Hard	Soft
Three roles	18.27	48.54	13.61	35.44
Repeated sampling	13.58	54.69	7.32	37.60
GPT-5.4
Three roles	17.42	45.48	13.32	34.38

Table 5: Repeated-sampling control on Chinese. Scores are percentages. T-A Pair and T-A-H Tri. Hard/Soft Macro-F1 are shown. Bold marks the higher score within each backbone. Complete metrics appear in Appendix H .

DeepSeek-v4-Pro
Condition	Targeted	NTA	TA
Three roles	32.03	21.10	13.91
Repeated sampling	19.56	7.33	3.14
GPT-5.4
Three roles	29.24	18.84	15.81
Repeated sampling	21.34	9.97	7.35

Table 6: Repeated-sampling control on English. Scores are percentages. Targeted, NTA, and TA are shown. Bold marks the higher score within each backbone. Complete metrics appear in Appendix H .

Setting	Tgt-H	Arg-H	TA-H	TAH-H	Avg.
Qwen3-14B (mainline)	48.73	21.03	12.36	8.71	32.82
w/o P1 segmenter	52.34	15.20	9.26	7.02	32.25
w/o P2 victim	50.40	15.95	10.53	7.63	31.16
w/o P2 moderator	49.90	16.51	10.86	7.75	31.77
w/o P2 bystander	50.18	16.10	10.50	8.20	32.85
w/o P3 arbitration	46.54	17.03	10.67	7.68	31.71

Table 7: Bold marks the highest score in each panel or metric column. Ablation summary on the bilingual main tracks. The main text keeps only the most diagnostic hard/exact metrics; complete bilingual ablation results appear in Appendix I .

Model	TAH-H	TAH-S	Avg.
0-shot-Qwen3-14B	7.65	23.93	29.83
SPAR-Qwen3-14B (mainline)	8.71	24.97	32.82
Distillation-Qwen3-4B	8.33	26.04	33.67

Table 8: Bold marks the highest score in each panel or metric column. Distillation summary across the three evaluation tracks.

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Item	Value
Backbone	Qwen3-14B
Hardware	RTX 3090 + RTX 4090D
Precision	bfloat16
Context length	8192
Tensor parallelism	2
Phase 1	local focus-unit segmentation; batch 4

Appendix

Table 9: Main local configuration for the reported Qwen3-14B mainline.

Item	Value
Role set	victim / moderator / cultural_bystander
Quality threshold	0.60
Cluster similarity	0.66
Similarity weights	target/argument/label/group = 0.45/0.45/0.05/0.05
Score bonuses	support role 0.12; candidate lexicon 0.05; cluster lexicon 0.08
Auditor lanes	consensus / defective / conflict / lexicon-hit / victim-microaggression

Appendix

Table 10: Key Phase 3 settings in the local 14B arbitration configurations.

Feature group	Signal
Grounding	argument in focus text; no invented evidence
Target anchoring	explicit target; canonical-target match
Structural validity	JSON validity; legal label; track-compatible fields
Group-label coherence	group –label consistency in ZH-quadruple
Soft priors	lexicon bonus; moderator floor; victim microaggression

Appendix

Table 11: Main feature groups used by the Phase 3 candidate scorer.

Step	Operation	Bound
Validate	Check JSON, label, argument grounding, and target anchor	quality $\geq 0.60$
Cluster	Compare target, argument, label, and optional group	similarity $\geq 0.66$
Route	Assign consensus, defective, or conflict lane	one local unit
Refine	Apply soft beacons and optional auditor call	one regeneration
Select	Rank with deterministic fallback	at most two clusters

Appendix

Table 12: Operational summary of local Phase 3 arbitration.

Measure	Retained-run summary
Raw-cluster p95	EN 3; ZH 4
Retained/final p95 and maximum	2
Conflict-lane share	55.9–61.7%
Auditor-use share	65.1–70.7%

Appendix

Table 13: Boundedness and routing diagnostics across the retained runs.

Language	Observed range	Default
EN	34.97–35.08	35.08
ZH	26.28–27.12	26.82

Appendix

Table 14: Four-hard-metric averages in the fixed-subset threshold replay. The results are not full-test significance estimates.

Object	R1	R2	Agree.	$\boldsymbol{\kappa}$
Final tuple	73.3	71.7	81.7	0.540
Bystander rationale	68.3	58.3	86.7	0.716

Appendix

Table 15: Acceptance rates, percentage agreement, and Cohen’s

\kappa

in the blind semantic audit.

Track	SPAR-DeepSeek	Direct references
EN	20/800 (2.50%)	0–9/800 (0.00–1.13%)
ZH-main	141/1605 (8.79%)	zero-shot: 37/1605; 66/1605

Appendix

Table 16: Counts and percentages of samples whose predicted tuple sets exactly match the gold sets.

Gold group	Samples	TAH-S	Quad-S
Sexism	302	36.66	34.49
Racism	223	40.56	35.57
Region	216	27.20	24.93
LGBTQ	93	34.75	32.13
Racism + Sexism	85	21.85	5.30

Appendix

Table 17: ZH-quadruple attacked-group diagnostic. Scores are Soft F1 and groups are defined from gold labels.

Item	Content
Text	First impression: Henan is poor and populous; Northeast China has one of the world’s lowest birth rates; public security in Xinjiang may not be very good.
Gold	$g_{1}$ : (Henan, “poor; populous”, 1) $g_{2}$ : (Northeast China, one of the world’s lowest birth rates, 1) $g_{3}$ : (Xinjiang, public security may not be very good, 1)
Phase 1	$g_{1}$ : Henan: poor; populous $g_{2}$ : Northeast China: one of the world’s lowest birth rates $g_{3}$ : Xinjiang: public security may not be very good
Phase 2	$g_{1}$ : only hateful candidates remain. $g_{2}$ : (Northeast China, one of the world’s lowest birth rates, 1/0). $g_{3}$ : (Xinjiang, public security may not be very good, 1/0).
Phase 3	$g_{1}$ : (Henan, “poor; populous”, 1) $g_{2}$ : (Northeast China, one of the world’s lowest birth rates, 1) $g_{3}$ : (Xinjiang, public security may not be very good, 1)
Baselines	$g_{1}$ : Dance is correct. SynChain outputs (Henan, poor, 0), missing “populous”. $g_{2}$ : Dance and SynChain both assign label 0. $g_{3}$ : Dance and SynChain both assign label 0.

Appendix

Table 18: Phase-wise processing of three local focus units.

g_{1}

g_{2}

, and

g_{3}

denote Phase 1 units. One unit may yield several sample-level tuples after Phase 4. The Chinese input is shown in English translation.

Item	Content
Text	some of y’all aren’t being held accountable for the dumb shit that y’all be doing….and it shows
Gold	(y’all, dumb shit, 1)
SPAR-Qwen3-14B	(y’all, dumb shit that y’all be doing, 1); (null, it shows, 0)

Appendix

Table 19: A remaining SPAR-Hate error: overly wide argument boundaries.

Field	Content
Text	The people spreading the claim that Henan people steal manhole covers are in fact Beijingers.
Gold tuple	(Beijingers, saying that Henan people steal manhole covers, 1)
Victim	(Beijingers, saying that Henan people steal manhole covers, 1)
Moderator	(Henan people, steal manhole covers, 1)
Cultural Bystander	(Beijingers, saying that Henan people steal manhole covers, 1)
Triage	lane=conflict ; initial_conflict=target_conflict

Appendix

Table 20: Chinese case study showing target-conflict resolution in Phase 3.

Field	Content
Text	that shit is not cute & dnt make u hard.
Phase 4 restored sample prediction	(bitches, Ghetto, 1) ; (bitches, ignorant, 1) ; (shit, shit, 1) ; (IMPLICIT_TARGET, alldat shit, 0)
Victim	(shit, shit, 1) (u, dnt make u hard, 1)
Moderator	(IMPLICIT_TARGET, that shit, 0) (IMPLICIT_TARGET, dnt make u hard, 0)
Cultural Bystander	(shit, shit, 0)
Triage	lane=conflict

Appendix

Table 21: English case study showing conflict-lane arbitration inside a multi-target sample.

Field group	Content
Identifiers	sample_id , group_id , source, split, language
Source text	original text and the current focus/group unit
Focus evidence	group_text , focus_text , canonical_target
Role evidence	tuples from the three roles, short rationales, quality scores, lexicon hit
Conflict diagnosis	triage lane, field conflicts, applied beacons, clusters
Teacher decision	teacher_final_tuple , teacher_rationale , audit confidence

Appendix

Table 22: Schema of the final constructed distillation rows.

Stage	Count	Note
Phase 3 group traces	7,752	all available local decision traces
Packaged train rows	3,739	base samples entering the teacher package
Teacher reasoning HQ	2,221	main warm-up variant
HQ Chinese rows	1,590	zh_group_hateful
HQ English rows	631	en_span_harmful
Teacher reasoning plus	3,194	reasoning variant with relaxed conflict threshold

Appendix

Table 23: Summary statistics for the teacher-trace construction pipeline.

Field	Content
Source text	Marriage is a necessity for women. Men can stay unmarried for life, but women cannot. In reality, countless men are even more eager. They will empty the family savings and act like simps just to marry and have children. Absurd.
Sample-level gold	(marriage, a necessity for women, 1); (countless men, empty the family savings and act like simps just to marry and have children, 1)
Current focus	In reality, countless men are even more eager. They will empty the family savings and act like simps just to marry and have children. Absurd.
Dataset variant	teacher_trace_reasoning_hq
Victim	tuple: (men, even more eager than women, and willing to empty the family savings and act like simps just to marry and have children, harmful); rationale: It describes men as simps and derides their behaviour, carrying clear gender stereotyping and disparagement.
Moderator	tuple: (men, willing to empty the family savings and act like simps just to marry and have children, harmful); rationale: The focus uses derogatory wording such as “simps” to negatively characterise men. The phrasing is attacking and demeaning, which violates community standards.

Appendix

Table 24: A condensed real training row from the constructed distillation data.

DeepSeek-v4-Pro
Condition	Target		Argument
	Hard	Soft	Hard	Soft
Three roles	61.22	58.76	27.79	67.97
Repeated	47.56	74.36	18.32	64.47
GPT-5.4
Three roles	52.62	61.94	23.12	56.62

Appendix

Table 26: Complete repeated-sampling results on Chinese. Scores are percentages. Repeated denotes repeated sampling. Bold marks the higher score within each backbone.

DeepSeek-v4-Pro
Condition	Target	Argument	Targeted	NTA	TA	Harm
Three roles	49.75	49.46	32.03	21.10	13.91	57.43
Repeated	47.26	35.10	19.56	7.33	3.14	47.81
GPT-5.4
Three roles	42.15	47.00	29.24	18.84	15.81	56.78
Repeated	40.13	38.44	21.34	9.97	7.35	53.13

Appendix

Table 27: Complete repeated-sampling results on English. Scores are percentages. Repeated denotes repeated sampling. Bold marks the higher score within each backbone.

Setting	Target		Argument		T-A Pair		T-A-H Tri.		Avg.
	Hard	Soft	Hard	Soft	Hard	Soft	Hard	Soft
Qwen3-14B (mainline)	48.73	55.75	21.03	56.85	12.36	34.14	8.71	24.97	32.82
w/o phase1 segmenter	52.34	64.46	15.20	48.92	9.26	34.43	7.02	26.39	32.25
w/o phase2 victim	50.40	60.67	15.95	46.48	10.53	32.84	7.63	24.78	31.16
w/o phase2 moderator	49.90	60.07	16.51	47.65	10.86	34.91	7.75	26.49	31.77
w/o phase2 bystander	50.18	61.11	16.10	50.90	10.50	36.93	8.20	28.90	32.85

Appendix

Table 28: Bold marks the highest score in each panel or metric column. Complete bilingual Qwen3-14B ablations.

Local LLMs
Model	Target		Argument		T-A Pair		T-A-H Tri.		Quad.		Avg.
	Hard	Soft	Hard	Soft	Hard	Soft	Hard	Soft	Hard	Soft
0-shot-Qwen3-14B	43.94	53.81	15.60	47.00	9.92	29.58	7.65	23.18	6.87	19.42	25.70
SPAR-Qwen3-14B (mainline)	48.54	58.58	19.61	53.94	11.00	34.14	8.51	26.61	6.18	22.37	28.95
SPAR-Qwen3.5-35B	50.81	60.84	19.57	52.05	13.64	38.50	10.59	29.94	8.68	24.48	30.91
API Models

Appendix

Table 29: Full Chinese four-field results. Best and second-best scores are marked within the local and API groups. Starred rows are historical STATE-ToxiCN values Bai et al. (2025) .

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 100 papers

ParsHate: A Benchmark Dataset for Hate and Target Detection in Persian
Sep 14, 2026Zahra Bokaei, Walid Magdy, Bonnie WebberHate Speech Detection
School of Informatics University of Edinburgh
From Triage to Impact: A Multi-Hop Reasoning Benchmark for Explainable Persian Hate Speech
Oct 14, 2026Ghazal Kalhor, Sima Kamandloo, Mehrsa Shamsian +1Hate Speech Detection LLM Evaluation
EPFL · Shahid Beheshti University · University of Social Welfare and Rehabilitation Sciences +1
An Explainable DistilBERT-BiLSTM-Attention Framework for Binary and Multi-Class Hate Speech Detection
Sep 23, 2026Rameesha Zia, Muhammad Shahid Iqbal MalikExplainable Text Classification Text Classification
School of Computing Sciences, Pak-Austria Fachhochschule, Institute of Applied Sciences and Technology, Haripur 22620, Pakistan

Sep 14, 2026J/K move · Enter open · S save

ParsHate: A Benchmark Dataset for Hate and Target Detection in Persian

Zahra Bokaei, Walid Magdy, Bonnie Webber

School of Informatics University of Edinburgh

Hate Speech Detection

Open paper AskPDF arXiv

We introduce ParsHate, a manually annotated dataset of 10,000 Persian tweets spanning 2013-2022, representing the first decade-long benchmark for hate speech detection in Persian. The dataset contains 31% hateful content and supports both hate detection and multi-label fine-grained target identification across seven structured target categories. ParsHate also distinguishes explicit and implicit hate, marks explicit and implicit targets, and provides span-level rationales. Data collection combines random and score-stratified temporal sampling to reduce keyword-driven bias while preserving natural label distributions. Applying SOTA models for Persian hate-speech detection on ParsHate shows moderate performance (79% F1), especially with samples from earlier years, and low performance with target identification (25.5% macro-F1). This emphasizes the diverse sampling of hate speech in ParsHate and its challenging nature that requires more advanced methods for better performance. Dataset is made publicly available.

Abstract

Hate Speech Detection

Explore similar work

ParsHate: A Benchmark Dataset for Hate and Target Detection in Persian

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

An Explainable DistilBERT-BiLSTM-Attention Framework for Binary and Multi-Class Hate Speech Detection