Paper deep dive
SPAR-Hate: Auditor-Guided Multi-Perspective Role Reasoning for Bilingual Hate Speech Parsing
Yifan Lyu, Dianqing Lin, Xinran Li, Jiaqi Qiao, Xiujuan Xu
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 95%
Last extracted: 8/26/2026, 3:52:12 AM
Summary
The paper introduces SPAR-Hate, an auditor-guided multi-perspective role-reasoning framework for bilingual hate speech parsing. It addresses challenges in identifying targets, arguments, and labels in documents with multiple targets or culturally coded language by decomposing text into local focus units and generating candidates from Victim, Moderator, and Cultural Bystander perspectives. These candidates are clustered and arbitrated under grounding constraints before reassembly. Experiments on STATE-ToxiCN (Chinese) and TBO (English) demonstrate improved performance on strict joint target-argument-label metrics compared to baselines.
Entities (10)
Relation Signals (9)
SPAR-Hate → evaluatedon → STATE-ToxiCN
confidence 98% · Experiments on STATE-ToxiCN and a controlled TBO split show gains
SPAR-Hate → evaluatedon → TBO
confidence 98% · Experiments on STATE-ToxiCN and a controlled TBO split show gains
SPAR-Hate → backedby → Qwen3-14B
confidence 95% · The main local experiments use Qwen3-14B
SPAR-Hate → backedby → GPT-5.4
confidence 95% · SPAR-GPT-5.4 records EN NTA/TA scores
SPAR-Hate → backedby → DeepSeek-v4-Pro
confidence 95% · SPAR-DeepSeek-v4-Pro records the highest ZH-main average
SPAR-Hate → performstask → Target-Argument-Label Parsing
confidence 95% · SPAR-Hate is an auditor-guided multi-perspective role-reasoning framework for bilingual hate speech parsing... jointly identify targets, supporting arguments, and target-level labels
SPAR-Hate → usesperspective → Cultural Bystander
confidence 95% · SPAR-Hate ... elicits evidence-grounded candidates from Victim, Moderator, and Cultural Bystander perspectives
SPAR-Hate → usesperspective → Moderator
confidence 95% · SPAR-Hate ... elicits evidence-grounded candidates from Victim, Moderator, and Cultural Bystander perspectives
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
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. Full-test integrated-prompt controls, component ablations, and bounded-arbitration diagnostics identify the contribution of separated perspective generation and arbitration. Structured teacher traces also support training a smaller student model.
Tags
Links
- Source: https://arxiv.org/abs/2608.22018v2
- Canonical: https://arxiv.org/abs/2608.22018v2
Trouble viewing inline? Open PDF directly →
Full Text
69,208 characters extracted from source content.
Expand or collapse full text
SPAR-Hate: Auditor-Guided Multi-Perspective Role Reasoning for Bilingual Hate Speech Parsing Yifan Lyu Affiliation: Dalian University of Technology Dianqing Lin Affiliation: Inner Mongolia UniversityCorrespondence:stevelyu811@gmail.com Xinran Li Affiliation: Dalian University of Technology Jiaqi Qiao Affiliation: Dalian University of Technology Xiujuan Xu Affiliation: Dalian University of Technology 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. Full-test integrated-prompt controls, component ablations, and bounded-arbitration diagnostics identify the contribution of separated perspective generation and arbitration. Structured teacher traces also support training a smaller student model. 1 Introduction Hate speech detection identifies hateful expressions whose interpretation depends on context, social judgement, and platform standards (Schmidt and Wiegand, 2017; Fortuna and Nunes, 2018). Fine-grained work now includes rationale annotation, target–argument parsing, and tuple extraction (Pavlopoulos et al., 2021; Mathew et al., 2021; Zampieri et al., 2023; Bai et al., 2025). Structured hate parsing predicts one or more target-level tuples, each linking an attacked target, its supporting argument, and a target-level label. Figure 1: Typical LLM errors in bilingual hate parsing, including omitted tuples, imprecise target–argument binding, and failed interpretation of culturally localised homophonic slurs. Figure 1 shows omitted tuples, faulty target–argument binding, and a missed homophonic slur. These failures require consistent tuple prediction and culturally grounded interpretation. Large language models remain skewed towards Western cultural representations (Naous et al., 2024), weakening performance on implicit, coded, and locally grounded hate (ElSherief et al., 2021; Nozza, 2021; Ocampo et al., 2023; Lin et al., 2026). Cultural knowledge can change target identification, evidence selection, and harm attribution. Social position also shapes interpretations of hostile speech (Spears, 2021), so the same utterance can support different grounded readings. SPAR-Hate comprises four stages: Segment, Perspective-Guided Role Generation, Arbitrate, and Reassemble. Phase 1 produces local focus units. Victim, Moderator, and Cultural Bystander generators produce evidence-grounded candidates for each unit. Phase 3 clusters and selects candidates under grounding and schema constraints, and Phase 4 restores sample-level predictions. The Chinese Cultural Bystander receives weak lexical context for coded language and local slang Bai et al. (2025); Xiao et al. (2024); Lu et al. (2023). Across local and API settings, SPAR-Hate improves the stricter joint structural metrics on Chinese STATE-ToxiCN and the controlled English TBO split. Component ablations test each phase, integrated-prompt controls test separated perspective generation and arbitration, and distillation transfers the structured traces to a smaller student model. Contributions. The task formulation aligns Chinese and English structured parsing at field level while retaining their original annotation contracts. The framework combines local focus-unit segmentation, three perspective-conditioned generators, constrained arbitration, and sample-level reassembly. Main experiments, full-test integrated-prompt controls, ablations, and diagnostics test strict structural recovery and structured teacher-trace transfer. 2 Related Work 2.1 From Hate Speech Classification to Structured Tuple Parsing Hate speech research began largely as text classification, but class labels alone do not support fine-grained semantic understanding. Early work focused on definitions, features, and classification models Schmidt and Wiegand (2017); Fortuna and Nunes (2018); Waseem and Hovy (2016); Davidson et al. (2017). Later multilingual shared tasks and richer annotations showed that multilingual hate analysis requires more than single-label prediction Basile et al. (2019); Ousidhoum et al. (2019). Toxic Spans, HateXplain, TBO, and STATE-ToxiCN then moved the field toward fine-grained localisation, rationale annotation, target-argument parsing, and Chinese quadruple parsing Pavlopoulos et al. (2021); Mathew et al. (2021); Zampieri et al. (2023); Bai et al. (2025). SPAR-Hate focuses on local target–argument–label bindings under separate Chinese and English annotation contracts and evaluates their sample-level reconstruction. 2.2 Implicit Hate, Cultural Context, and Cross-Lingual Fragility Implicit, subtle, and context-dependent hate remains difficult to detect. Coded and indirect expressions weaken model performance ElSherief et al. (2021); Ocampo et al. (2023); Hartvigsen et al. (2022), while HateCheck identifies related functional failures Röttger et al. (2021). Cross-lingual zero-shot models can also misread language-specific non-hateful taboo expressions as hate signals Nozza (2021). Explainable hate-speech detection uses rationales, social bias frames, and stepwise explanations Mathew et al. (2021); Sap et al. (2020); Yang et al. (2023). Structured parsing adds the requirement that target, argument, and label remain jointly aligned. SPAR-Hate supplies cultural knowledge as weak context and retains only text-grounded candidates. 2.3 Role-Conditioned Reasoning and Constrained Aggregation RoleLLM and multi-perspective role-playing show that role conditioning elicits distinct knowledge and reasoning biases Wang et al. (2024). Proposer–aggregator methods coordinate several model outputs Du et al. (2023). AutoGen and MetaGPT organise role allocation and structured workflow handoffs, with MetaGPT encoding these handoffs through standard operating procedures Wu et al. (2023); Hong et al. (2024). SPAR-Hate constrains role outputs to comparable target–argument–label candidates before aggregation. Its distillation traces retain focus-unit decomposition, role hypotheses, conflict diagnosis, and final arbitration, linking the method to chain-of-thought, self-consistency, and reasoning distillation Wei et al. (2023); Wang et al. (2023); Shridhar et al. (2023); Hsieh et al. (2023). 3 Methods: The SPAR-Hate Framework SPAR-Hate combines local focus-unit decomposition, multi-perspective candidate generation, dynamic arbitration, and sample-level reassembly in an auditor-guided framework for bilingual hate parsing. The pipeline breaks document-level parsing into explicit intermediate stages, which helps long texts, multi-target cases, implicit attacks, and culturally coded slang. 3.1 Task Formulation and Output Contracts Given an input document D, the system must predict a set of target-level structured tuples E=e1,e2,…,en.E=\e_1,e_2,…,e_n\. (1) The Chinese and English benchmarks follow different original annotation contracts. STATE-ToxiCN uses quadruples ezhraw=(target,argument,group,hateful),e_zh^raw=(target,argument,group,hateful), (2) where group denotes the attacked group category. TBO uses triples enraw=(target,argument,harmful).e_en^raw=(target,argument,harmful). (3) These schemas correspond respectively to the Target–Argument–Hateful–Group annotation contract in STATE-ToxiCN and the target–argument–harmfulness contract in TBO Bai et al. (2025); Zampieri et al. (2023). To construct the bilingual main track at field level, the Chinese main track removes group and uses a three-field output ezhmain=(target,argument,label).e_zh^main=(target,argument,label). (4) For unified notation, each tuple on the aligned bilingual main tracks is written as ei=(ti,ai,ℓi)e_i=(t_i,a_i, _i). Here tit_i denotes the attacked target, aia_i the attack argument, and ℓi∈0,1 _i∈\0,1\ the target-level harmfulness label. It corresponds to harmful in English and hateful in Chinese. The Chinese four-field extension retains group as an additional field. The Chinese three-field task supports the bilingual main comparison. A four-field extension retains group and tests adaptation to the original language-specific schema. Separate prompt templates and output constraints preserve both contracts. 3.2 Framework Overview SPAR-Hate has four stages: Segment, Perspective-Guided Role Generation, Arbitrate, and Reassemble, corresponding to Phase 1–4. The system splits a document into local focus units, generates structured candidates from three perspectives, arbitrates them with a dynamic-beacon procedure, and reassembles benchmark-aligned sample-level outputs. Figure 2 presents the full pipeline. Figure 2: Overview of SPAR-Hate. Bilingual input is segmented into local focus units, processed by three role-conditioned generators, arbitrated under evidence constraints, and reassembled into sample-level outputs. Distillation uses the earlier phases to construct teacher traces for student training. 3.3 Phase 1: Local Focus-Unit Segmentation Long texts often contain multiple targets, local stances, and interfering attack fragments. Direct extraction from the full document can therefore produce target–argument mismatches, overly wide arguments, and merged events. Divide-and-conquer frameworks for document-level sentiment parsing suggest the same advantage for structured extraction Wang et al. (2026). Phase 1 therefore uses an LLM-based segmenter to map document D to a set of local decision units G=g1,g2,…,gk.G=\g_1,g_2,…,g_k\. (5) Each local unit gig_i stores sample-level and local indices (group_id, local_group_id), local text (group_text), a focus-marked local context (focus_text), and a soft target anchor (canonical_target). A focus unit is defined around a target and local intent while retaining enough context for argument grounding. Its boundary need not coincide with a syntactic boundary. 3.4 Phase 2: Perspective-Guided Role Generation Perspective shapes hate judgements Waseem and Hovy (2016); Sap et al. (2020). The task-motivated role set covers affected-group harm, platform-governance boundaries, and community interpretation of implicit or culturally coded hate. For each local unit gig_i, the framework instantiates three role-conditioned generators R=rvictim,rmoderator,rbystander.R=\r_victim,r_moderator,r_bystander\. (6) Victim emphasises felt harm, exclusion, and microaggressions. Moderator emphasises platform-governance boundaries and explicit rule violations. Cultural Bystander emphasises local cultural context, community slang, pragmatic history, and coded expression. Each role generates one or more structured candidates over the same focus_text, yielding the role-specific candidate set Hr(gi)=hr(1),hr(2),….H_r(g_i)=\h_r^(1),h_r^(2),…\. (7) SPAR-Hate imposes a strict JSON schema and requires argument to be a supporting substring inside focus_text. Every judgement includes local textual evidence. The Chinese Cultural Bystander receives entries from the auxiliary STATE-ToxiCN lexicon. Its 829 term/category/definition records contain no sample identifiers, gold tuples, or instance labels. Retrieved entries provide weak context for the current focus_text and canonical_target. Candidates remain subject to focus-span grounding, and lexicon-only singleton candidates are discarded. Lexicon hits occur in 10.1% of ZH-main groups and 9.6% of ZH-quadruple groups. Chinese toxicity research motivates this support for slang, homophones, and emoji cloaking Lu et al. (2023); Bai et al. (2025); Xiao et al. (2024). Appendix B gives the prompt templates. 3.5 Phase 3: Local Candidate Clustering and Auditor Arbitration Phase 3 filters, clusters, and ranks the role candidates for each local unit. Given ℋi=⋃r∈RHr(gi),H_i= _r∈ RH_r(g_i), (8) the procedure applies quality scoring, cluster formation, triage, and deterministic selection. Candidate Quality Scoring The system first assigns each candidate c∈ℋic _i a quality score q(c)q(c): q(c)=∑jαjϕj(c).q(c)= _j _j\, _j(c). (9) Here ϕj(c) _j(c) includes grounding, target explicitness, consistency with canonical_target, structural completeness, label validity, and, in the Chinese four-field setting, group–label coherence. Chinese bystander candidates can receive a small lexicon bonus. Candidates below quality 0.600.60 are filtered before clustering. An ungrounded or missing argument fails validation. Cluster Formation Candidates are then clustered by similarity over target, argument, harmful/hateful, and, in the Chinese four-field extension, group. Similarity between a candidate and a cluster is written as sim(c,C)= (c,C)= βtst(c,C)+βasa(c,C) _t\,s_t(c,C)+ _a\,s_a(c,C) (10) +βysy(c,C)+βgsg(c,C), + _y\,s_y(c,C)+ _g\,s_g(c,C), where st,sa,sy,sgs_t,s_a,s_y,s_g denote similarity in target, argument, label, and group. On the three-field main tracks, βg=0 _g=0. The reported field weights are 0.45/0.45/0.05/0.050.45/0.45/0.05/0.05, respectively. Each cluster is then compressed into a canonical tuple and assigned a cluster score based on member quality, the number of supporting roles, and any lexicon-aware bonus: Score(C)=∑c∈Cq(c)+λ1|supp(C)|+λ2lex(C),Score(C)= _c∈ Cq(c)+ _1|supp(C)|+ _2I_lex(C), (11) where supp(C)supp(C) is the set of roles supporting that cluster. The mainline support-role and lexicon-cluster bonuses are 0.120.12 and 0.080.08; the candidate-level lexicon bonus is 0.050.05. Triage and Dynamic Soft Beacons After clustering, the system routes each local unit into one of three lanes according to role agreement, the margin of the top cluster, and signals of structural failure or missing roles. A clearly dominant top cluster with no obvious defect enters consensus. Missing roles or a lack of valid grounded candidates enters defective. The remaining cases enter conflict. The system then applies grounding-aware, lexicon-aware, and quote-aware soft beacons to adjust cluster ranking. If some roles are marked defective and regeneration is enabled, one bounded regeneration round is allowed. Highly conflicting cases can trigger a second auditor refinement. Deterministic Fallback and Local Refinement Final ranking uses a deterministic cluster-level fallback. The selected primary cluster undergoes local refinement of target and argument boundaries. The reported configuration retains at most two clusters, requires score ≥1.05≥ 1.05 and at least two supporting roles, and allows one regeneration round for a defective role. Across retained runs, 55.9–61.7% of focus units enter the conflict lane and 65.1–70.7% invoke auditor refinement. Appendix D reports the complete scoring, routing, and boundedness diagnostics. 3.6 Phase 4: Sample-Level Multi-Target Reassembly In this task setting, local focus-unit decisions are made internally while the benchmark expects sample-level predictions. Phase 4 therefore restores earlier decisions to the output format required by evaluation. The system first aggregates all local tuples by sample_id, producing E~(x)=⨄gi∈G(x)Y(gi), E(x)= _g_i∈ G(x)Y(g_i), (12) where Y(gi)Y(g_i) is the set of final tuples output by Phase 3 for local unit gig_i. The main configuration uses a concatenate-first strategy, preserving local provenance and not forcing exact deduplication. The default exported sample-level prediction is therefore E^(x)=E~(x), E(x)= E(x), (13) whereas under optional exact deduplication E^(x)=Dedup(E~(x)). E(x)=Dedup\! ( E(x) ). (14) 4 Experiments 4.1 Datasets and Evaluation Protocol Datasets Experiments are conducted on Chinese STATE-ToxiCN and English TBO Bai et al. (2025); Zampieri et al. (2023). STATE-ToxiCN keeps its official train/test split. The public release of TBO provides only a test split, so the public test set is deterministically shuffled with seed=20260415 and re-divided into 3200/800 train/test subsets. Beyond this repartition, processing is limited to field cleaning, schema normalisation, and split freezing. No extra relabelling is introduced, and multi-target documents are not split at this stage. The paper reports three evaluation tracks, ZH-main, EN-main, and ZH-quadruple, all defined exactly as in Section 3.1. Dataset composition appears in Table 1. TBO numbers are controlled within-paper comparisons on this deterministic split only; they are not directly comparable to studies evaluated on the original public test-only release. 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. Evaluation metrics The original evaluation protocols of both benchmarks are kept. No mixed cross-lingual total score is constructed. For Chinese, following STATE-ToxiCN, both Hard and Soft Macro-F1 are reported: Hard requires exact agreement with gold spans and field combinations, while Soft credits overlapping spans under the same target-centred structure Bai et al. (2025). On the Chinese tracks, Target and Argument evaluate target and argument field parsing. T-A Pair requires the target to be correctly bound to its argument. T-A-H Tri and Quad require the local structure to remain correct after adding the hateful label and, in the four-field task, the group field. For English, the paper follows TBO’s tuple-level evaluation: Target and Argument measure field-level parsing; Targeted requires joint parsing of target and harmfulness; NTA and TA require exact match on (target,argument)(target,argument) and (target,argument,harmful)(target,argument,harmful) respectively; Harm evaluates harmfulness on the predicted target tuples Zampieri et al. (2023). Chinese metrics therefore emphasise hard/soft field parsing, whereas English metrics emphasise exact tuple consistency Bai et al. (2025); Zampieri et al. (2023). 4.2 Experimental Setup and Baselines Experimental setup The main local experiments use Qwen3-14B under a dual-24GB-class GPU budget Yang et al. (2025). Phase 1–3 share the same backbone, with lexicon injection used only for the Chinese Cultural Bystander. Long runs support checkpoint-resume and OOM split-retry. Phase 4 is deterministic aggregation. Sample-level outputs follow the concatenate-first strategy of Section 3.6. Appendix A gives the runtime configuration and artifact map; Appendix D gives the arbitration settings. Baselines The baselines cover direct extraction, task-adapted SynChain and Dance, and SPAR with local and API backbones Fan et al. (2025); Wang et al. (2026); DeepSeek-AI (2026); OpenAI (2026). Table 2 reports the added full-test API zero-shot averages. Prompt templates and baseline adaptations appear in Appendices B and C. API zero-shot model EN ZH DeepSeek-v4-Pro 23.18 22.80 GPT-5.4 21.01 28.13 Table 2: Full-test average scores for the added API zero-shot baselines. Backbone Track Integrated SPAR DeepSeek-v4-Pro EN NTA / TA 4.87 / 3.89 19.81 / 13.33 GPT-5.4 EN NTA / TA 11.41 / 8.18 18.84 / 15.81 DeepSeek-v4-Pro ZH TA / TAH-H 6.31 / 4.40 17.27 / 13.01 GPT-5.4 ZH TA / TAH-H 14.08 / 10.56 17.42 / 13.32 Table 4: Full-test fixed-Phase-1 integrated-prompt controls. The integrated condition retains the three lens descriptions (and Chinese lexicon access) but removes separated candidates, clustering, and auditor arbitration. (a) ZH-main 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 (b) EN-main Setting Tgt NTA TA Harm Avg. Qwen3-14B (mainline) 43.45 20.96 18.68 59.31 37.51 w/o P1 segmenter 46.95 20.65 18.04 57.21 37.34 w/o P2 victim 44.70 21.18 17.20 57.21 37.02 w/o P2 moderator 44.29 21.34 17.51 59.11 37.43 w/o P3 arbitration 43.66 20.42 17.81 57.61 36.72 Table 5: 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 25. (a) ZH-main Model Target Argument T-A Pair T-A-H Tri. Avg. Hard Soft Hard Soft Hard Soft Hard Soft Local LLMs 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 few-shot-DeepSeek-v4-Pro 52.69 63.08 17.86 61.03 13.62 39.53 10.75 31.09 36.21 few-shot-GPT-5.4 54.17 65.83 18.87 62.10 13.82 43.90 10.60 31.47 37.60 SynChain-DeepSeek-v4-Pro 38.92 48.68 11.54 49.57 5.09 31.89 3.75 23.75 26.65 SynChain-GPT-5.4 34.89 43.15 11.52 45.63 5.38 30.28 4.04 21.98 24.61 Dance-DeepSeek-v4-Pro 53.19 62.77 19.34 57.40 15.75 41.52 12.55 32.47 36.87 Dance-GPT-5.4 51.99 60.73 23.89 57.72 19.32 44.56 13.63 33.99 38.23 SPAR-DeepSeek-v4-Pro 56.80 65.96 24.03 62.06 17.27 44.13 13.01 32.88 39.52 SPAR-GPT-5.4 52.62 61.94 23.12 56.62 17.42 45.48 13.32 34.38 38.11 (b) EN-main Model Target Argument Targeted NTA TA Harm Avg. Local LLMs 0-shot-Qwen3-14B 32.41 46.48 25.06 14.91 11.06 47.61 29.59 SPAR-Qwen3-14B (mainline) 43.45 47.96 34.68 20.96 18.68 59.31 37.51 SPAR-Qwen3.5-35B 46.49 50.22 35.43 20.28 16.72 60.11 38.21 API Models few-shot-DeepSeek-v4-Pro 36.26 42.40 22.12 3.09 1.73 53.98 26.60 few-shot-GPT-5.4 38.94 44.02 13.50 4.00 2.61 48.16 25.21 SynChain-DeepSeek-v4-Pro 38.70 41.22 18.26 2.85 2.58 58.00 26.94 SynChain-GPT-5.4 36.60 37.52 15.56 4.76 3.52 51.92 24.98 Dance-DeepSeek-v4-Pro 45.67 36.08 27.43 2.92 2.61 56.44 28.53 Dance-GPT-5.4 46.15 42.39 24.92 8.88 7.75 52.62 30.45 SPAR-DeepSeek-v4-Pro 51.31 50.36 30.02 19.81 13.33 56.19 36.84 SPAR-GPT-5.4 42.15 47.00 29.24 18.84 15.81 56.78 34.97 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. 4.3 Main Results Table 3 reports the complete bilingual main-track comparison. Main bilingual triplet results Under a fixed local 14B backbone, SPAR raises the average score on ZH-main from 29.8329.83 to 32.8232.82 and on EN-main from 29.5929.59 to 37.5137.51. The larger gains occur on stricter joint structural metrics. Qwen3.5-35B records 38.2138.21 on EN-main and 32.5832.58 on ZH-main. Among API systems, SPAR-DeepSeek-v4-Pro records the highest ZH-main average (39.5239.52), while SPAR-GPT-5.4 records EN NTA/TA scores of 18.84/15.8118.84/15.81. Chinese quadruple extension The Chinese four-field extension (ZH-quadruple) retains the group field as a test of adaptation to language-specific schema. Full results appear in Appendix Table 25. Cross-model observations The largest differences occur on Targeted, NTA, and TA for English and on T-A Pair, T-A-H Tri, and Quad for Chinese. Appendix F provides additional outputs and cases. 4.4 Framework Analysis Integrated-prompt controls Four full-test integrated-prompt controls fix the Phase 1 focus units and place all three perspective descriptions in one call. The Chinese controls retain the same lexicon access. Table 4 reports lower strict tuple-binding scores after removing separated role outputs, candidate clustering, and auditor arbitration. GPT-5.4 Target scores are 46.0146.01 for integrated prompting and 42.1542.15 for SPAR on EN, with corresponding ZH scores of 54.4054.40 and 52.6252.62. Component ablations Table 5 summarises the most diagnostic hard and exact metrics; complete results are deferred to Appendix 25. Phase-wise ablations Removing the segmenter raises Chinese Target Hard F1 from 48.7348.73 to 52.3452.34 and lowers T-A-H Tri Hard F1 from 8.718.71 to 7.027.02. On English, Target F1 rises from 43.4543.45 to 46.9546.95, while TA and Harm fall. Removing Phase 3 lowers the ZH and EN averages to 31.7131.71 and 36.7236.72. Role-wise ablations Removing the bystander changes the Chinese average from 32.8232.82 to 32.8532.85 and lowers T-A Pair Hard and T-A-H Tri Hard. Removing the moderator lowers English TA from 18.6818.68 to 17.5117.51, while NTA changes from 20.9620.96 to 21.3421.34. Diagnostic validation Supplementary exact-set and blinded semantic diagnostics appear in Appendix E. They are reported as secondary checks, not primary ranking metrics. The attacked-group breakdown in the same appendix locates a remaining multi-group binding error. 4.5 Distillation Results Table 6 compares the distilled 4B student with the zero-shot and SPAR-Qwen3-14B systems on all three evaluation tracks. (a) ZH-main 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 (b) EN-main Model TA Harm Avg. 0-shot-Qwen3-14B 11.06 47.61 29.59 SPAR-Qwen3-14B (mainline) 18.68 59.31 37.51 Distillation-Qwen3-4B 18.50 59.05 37.33 (c) ZH-quadruple Model Quad-H Quad-S Avg. 0-shot-Qwen3-14B 6.87 19.42 25.70 SPAR-Qwen3-14B (mainline) 6.18 22.37 28.95 Distillation-Qwen3-4B 7.04 25.05 30.38 Table 6: Distillation summary across the three evaluation tracks. The distilled student improves over SPAR-Qwen3-14B on both Chinese tracks and remains close to its teacher on EN-main, despite using a substantially smaller backbone. Appendix G reports teacher-trace construction, filtering, and student training. 4.6 Qualitative Analysis Case study Table 7 traces three Phase 1 local focus units from a Chinese multi-target example. Phase 2 produces conflicting labels for g2g_2 and g3g_3, and Phase 3 selects the gold label for both units. For g1g_1, SPAR retains the full argument “poor; populous”, which SynChain truncates. Error analysis The main residual error is argument boundary over-expansion. In Table 8, SPAR-Hate predicts the correct target and harmfulness but retains additional contextual material, so the prediction receives no tuple credit under exact match. Appendix F provides further cases. 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) Table 8: A remaining SPAR-Hate error: overly wide argument boundaries. 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 g1g_1: (Henan, “poor; populous”, 1) g2g_2: (Northeast China, one of the world’s lowest birth rates, 1) g3g_3: (Xinjiang, public security may not be very good, 1) Phase 1 g1g_1: Henan: poor; populous g2g_2: Northeast China: one of the world’s lowest birth rates g3g_3: Xinjiang: public security may not be very good Phase 2 g1g_1: only hateful candidates remain. g2g_2: (Northeast China, one of the world’s lowest birth rates, 1/0). g3g_3: (Xinjiang, public security may not be very good, 1/0). Phase 3 g1g_1: (Henan, “poor; populous”, 1) g2g_2: (Northeast China, one of the world’s lowest birth rates, 1) g3g_3: (Xinjiang, public security may not be very good, 1) Baselines g1g_1: Dance is correct. SynChain outputs (Henan, poor, 0), missing “populous”. g2g_2: Dance and SynChain both assign label 0. g3g_3: Dance and SynChain both assign label 0. Table 7: Phase-wise processing of three local focus units. g1g_1, g2g_2, and g3g_3 denote Phase 1 units. One unit may yield several sample-level tuples after Phase 4. The Chinese input is shown in English translation. 5 Conclusion SPAR-Hate combines local focus-unit segmentation, role-conditioned generation, constrained auditor arbitration, and sample-level reassembly. Its largest gains occur on stricter joint structural metrics under the controlled benchmark settings. Integrated-prompt controls record lower tuple-binding scores after collapsing perspective generation and arbitration into one call. The Chinese four-field extension and distillation results cover language-specific schema and structured teacher traces. Limitations Evaluation covers Chinese and English and omits a full comparison with task-specific fine-tuned systems. The task-motivated role set has not been exhaustively searched, and the lexicon has no complete removal ablation. Weak lexical context can still over-flag dialectal, culturally loaded, or reclaimed terms. The threshold replay, blind audit, and attacked-group breakdown have limited scope and cannot support broad claims about statistical optimality, fairness, or cultural validity. TBO results use the deterministic internal split and do not support direct comparison with results on the original public test-only release. Ethical Considerations This study uses public Chinese and English benchmark datasets to examine structured hate speech parsing. Their contents include abusive, discriminatory, and potentially traumatic expressions. The paper retains only representative examples needed for the academic argument and recommends a minimal-exposure principle in data processing, visualisation, and manual analysis to reduce secondary harm to researchers, annotators, and readers. Such systems also carry misuse risks, including over-censorship, automated punishment, and large-scale opinion monitoring. Earlier work shows that false positives can arise from ambiguous boundaries between offensive language and hate speech, spurious correlations around minority-group mentions, and model fragility on functional phenomena Davidson et al. (2017); Hartvigsen et al. (2022); Röttger et al. (2021). Structured outputs and traceable intermediate processes increase transparency, but they may also increase the operational reach of deployed moderation systems. Real-world deployment therefore requires human review, appeals, threshold calibration, error auditing, and continuing bias evaluation across languages, groups, and cultural settings. References Bai et al. (2025) Z. Bai, L. Yang, S. Yin, J. Lu, J. Zeng, H. Zhu, Y. Sun, and H. Lin STATE ToxiCN: a benchmark for span-level target-aware toxicity extraction in Chinese hate speech detection. In Findings of the Association for Computational Linguistics: ACL 2025, W. Che, J. Nabende, E. Shutova, and M. T. Pilehvar (Eds.), Vienna, Austria, p. 10206–10219. External Links: Link, Document, ISBN 979-8-89176-256-5 Cited by: Table 25, §1, §1, §2.1, §3.1, §3.4, §4.1, §4.1, §4.1. Basile et al. (2019) V. Basile, C. Bosco, E. Fersini, D. Nozza, V. Patti, F. M. Rangel Pardo, P. Rosso, and M. Sanguinetti SemEval-2019 task 5: multilingual detection of hate speech against immigrants and women in Twitter. In Proceedings of the 13th International Workshop on Semantic Evaluation, J. May, E. Shutova, A. Herbelot, X. Zhu, M. Apidianaki, and S. M. Mohammad (Eds.), Minneapolis, Minnesota, USA, p. 54–63. External Links: Link, Document Cited by: §2.1. Davidson et al. (2017) T. Davidson, D. Warmsley, M. Macy, and I. Weber Automated Hate Speech Detection and the Problem of Offensive Language. Proceedings of the International AAAI Conference on Web and Social Media 11 (1), p. 512–515. External Links: Link, Document Cited by: §2.1, Ethical Considerations. DeepSeek-AI (2026) DeepSeek-AI DeepSeek-v4: towards highly efficient million-token context intelligence. Note: Technical report External Links: Link Cited by: §4.2. Dettmers et al. (2023) T. Dettmers, A. Pagnoni, A. Holtzman, and L. Zettlemoyer QLoRA: efficient finetuning of quantized llms. External Links: 2305.14314, Link Cited by: §G.4. Du et al. (2023) Y. Du, S. Li, A. Torralba, J. B. Tenenbaum, and I. Mordatch Improving factuality and reasoning in language models through multiagent debate. External Links: 2305.14325, Link Cited by: §2.3. ElSherief et al. (2021) M. ElSherief, C. Ziems, D. Muchlinski, V. Anupindi, J. Seybolt, M. De Choudhury, and D. Yang Latent hatred: a benchmark for understanding implicit hate speech. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, M. Moens, X. Huang, L. Specia, and S. W. Yih (Eds.), Online and Punta Cana, Dominican Republic, p. 345–363. External Links: Link, Document Cited by: §1, §2.2. Fan et al. (2025) R. Fan, S. Li, T. He, and Y. Liu Aspect-based sentiment analysis with syntax-opinion-sentiment reasoning chain. In Proceedings of the 31st International Conference on Computational Linguistics, O. Rambow, L. Wanner, M. Apidianaki, H. Al-Khalifa, B. D. Eugenio, and S. Schockaert (Eds.), Abu Dhabi, UAE, p. 3123–3137. External Links: Link Cited by: Appendix C, §4.2. Fortuna and Nunes (2018) P. Fortuna and S. Nunes A survey on automatic detection of hate speech in text. ACM Comput. Surv. 51 (4). External Links: ISSN 0360-0300, Link, Document Cited by: §1, §2.1. Hartvigsen et al. (2022) T. Hartvigsen, S. Gabriel, H. Palangi, M. Sap, D. Ray, and E. Kamar ToxiGen: a large-scale machine-generated dataset for adversarial and implicit hate speech detection. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), S. Muresan, P. Nakov, and A. Villavicencio (Eds.), Dublin, Ireland, p. 3309–3326. External Links: Link, Document Cited by: §2.2, Ethical Considerations. Hong et al. (2024) S. Hong, M. Zhuge, J. Chen, Y. Xiawu, et al. MetaGPT: meta programming for a multi-agent collaborative framework. External Links: 2308.00352, Link Cited by: §2.3. Hsieh et al. (2023) C. Hsieh, C. Li, C. Yeh, H. Nakhost, Y. Fujii, A. Ratner, R. Krishna, C. Lee, and T. Pfister Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes. In Findings of the Association for Computational Linguistics: ACL 2023, A. Rogers, J. Boyd-Graber, and N. Okazaki (Eds.), Toronto, Canada, p. 8003–8017. External Links: Link, Document Cited by: §2.3. Hu et al. (2021) E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen LoRA: low-rank adaptation of large language models. External Links: 2106.09685, Link Cited by: §G.4. Kwon et al. (2023) W. Kwon, Z. Li, S. Zhuang, Y. Sheng, L. Zheng, C. H. Yu, J. E. Gonzalez, H. Zhang, and I. Stoica Efficient memory management for large language model serving with pagedattention. External Links: 2309.06180, Link Cited by: Appendix A. Lin et al. (2026) D. Lin, T. Lan, J. Zhu, J. Li, W. Chen, X. Liu, Aruukhan, X. Su, H. Hou, and G. Gao Exploring the capability boundaries of llms in mastering of chinese chouxiang language. External Links: 2604.15841, Link Cited by: §1. Lu et al. (2023) J. Lu, B. Xu, X. Zhang, C. Min, L. Yang, and H. Lin Facilitating fine-grained detection of Chinese toxic language: hierarchical taxonomy, resources, and benchmarks. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), A. Rogers, J. Boyd-Graber, and N. Okazaki (Eds.), Toronto, Canada, p. 16235–16250. External Links: Link, Document Cited by: §1, §3.4. Mathew et al. (2021) B. Mathew, P. Saha, S. M. Yimam, C. Biemann, P. Goyal, and A. Mukherjee HateXplain: a benchmark dataset for explainable hate speech detection. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 35, p. 14867–14875. External Links: Link, Document Cited by: §1, §2.1, §2.2. Naous et al. (2024) T. Naous, M. J. Ryan, A. Ritter, and W. Xu Having beer after prayer? measuring cultural bias in large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), L. Ku, A. Martins, and V. Srikumar (Eds.), Bangkok, Thailand, p. 16366–16393. External Links: Link, Document Cited by: §1. Nozza (2021) D. Nozza Exposing the limits of zero-shot cross-lingual hate speech detection. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), C. Zong, F. Xia, W. Li, and R. Navigli (Eds.), Online, p. 907–914. External Links: Link, Document Cited by: §1, §2.2. Ocampo et al. (2023) N. B. Ocampo, E. Sviridova, E. Cabrio, and S. Villata An in-depth analysis of implicit and subtle hate speech messages. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, A. Vlachos and I. Augenstein (Eds.), Dubrovnik, Croatia, p. 1997–2013. External Links: Link, Document Cited by: §1, §2.2. OpenAI (2026) OpenAI Introducing GPT-5.4. Note: Product announcement External Links: Link Cited by: §4.2. Ousidhoum et al. (2019) N. Ousidhoum, Z. Lin, H. Zhang, Y. Song, and D. Yeung Multilingual and multi-aspect hate speech analysis. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), K. Inui, J. Jiang, V. Ng, and X. Wan (Eds.), Hong Kong, China, p. 4675–4684. External Links: Link, Document Cited by: §2.1. Pavlopoulos et al. (2021) J. Pavlopoulos, J. Sorensen, L. Laugier, and I. Androutsopoulos SemEval-2021 task 5: toxic spans detection. In Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), A. Palmer, N. Schneider, N. Schluter, G. Emerson, A. Herbelot, and X. Zhu (Eds.), Online, p. 59–69. External Links: Link, Document Cited by: §1, §2.1. Röttger et al. (2021) P. Röttger, B. Vidgen, D. Nguyen, Z. Waseem, H. Margetts, and J. Pierrehumbert HateCheck: functional tests for hate speech detection models. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), C. Zong, F. Xia, W. Li, and R. Navigli (Eds.), Online, p. 41–58. External Links: Link, Document Cited by: §2.2, Ethical Considerations. Sap et al. (2020) M. Sap, S. Gabriel, L. Qin, D. Jurafsky, N. A. Smith, and Y. Choi Social bias frames: reasoning about social and power implications of language. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, D. Jurafsky, J. Chai, N. Schluter, and J. Tetreault (Eds.), Online, p. 5477–5490. External Links: Link, Document Cited by: §2.2, §3.4. Schmidt and Wiegand (2017) A. Schmidt and M. Wiegand A survey on hate speech detection using natural language processing. In Proceedings of the Fifth International Workshop on Natural Language Processing for Social Media, L. Ku and C. Li (Eds.), Valencia, Spain, p. 1–10. External Links: Link, Document Cited by: §1, §2.1. Shridhar et al. (2023) K. Shridhar, A. Stolfo, and M. Sachan Distilling reasoning capabilities into smaller language models. External Links: 2212.00193, Link Cited by: §2.3. Spears (2021) R. Spears Social influence and group identity. Annual Review of Psychology 72, p. 367–390. External Links: Document, Link Cited by: §1. Wang et al. (2026) L. Wang, M. Huang, and E. Dragut DanceHA: A Multi-Agent Framework for Document-Level Aspect-Based Sentiment Analysis. Proceedings of the AAAI Conference on Artificial Intelligence 40 (35), p. 29714–29722. External Links: Link, Document Cited by: Appendix C, §3.3, §4.2. Wang et al. (2024) N. Wang, Z.y. Peng, H. Que, J. Liu, W. Zhou, Y. Wu, H. Guo, R. Gan, Z. Ni, J. Yang, M. Zhang, Z. Zhang, W. Ouyang, K. Xu, W. Huang, J. Fu, and J. Peng RoleLLM: benchmarking, eliciting, and enhancing role-playing abilities of large language models. In Findings of the Association for Computational Linguistics: ACL 2024, L. Ku, A. Martins, and V. Srikumar (Eds.), Bangkok, Thailand, p. 14743–14777. External Links: Link, Document Cited by: §2.3. Wang et al. (2023) X. Wang, J. Wei, D. Schuurmans, Q. Le, E. Chi, S. Narang, A. Chowdhery, and D. Zhou Self-consistency improves chain of thought reasoning in language models. External Links: 2203.11171, Link Cited by: §2.3. Waseem and Hovy (2016) Z. Waseem and D. Hovy Hateful symbols or hateful people? predictive features for hate speech detection on Twitter. In Proceedings of the NAACL Student Research Workshop, J. Andreas, E. Choi, and A. Lazaridou (Eds.), San Diego, California, p. 88–93. External Links: Link, Document Cited by: §2.1, §3.4. Wei et al. (2023) J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. Chi, Q. Le, and D. Zhou Chain-of-thought prompting elicits reasoning in large language models. External Links: 2201.11903, Link Cited by: §2.3. Wu et al. (2023) Q. Wu, G. Bansal, J. Zhang, Y. Wu, B. Li, E. Zhu, L. Jiang, X. Zhang, S. Zhang, J. Liu, A. H. Awadallah, R. W. White, D. Burger, and C. Wang AutoGen: enabling next-gen llm applications via multi-agent conversation. External Links: 2308.08155, Link Cited by: §2.3. Xiao et al. (2024) Y. Xiao, Y. Hu, K. T. W. Choo, and R. K. Lee ToxiCloakCN: evaluating robustness of offensive language detection in Chinese with cloaking perturbations. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Y. Al-Onaizan, M. Bansal, and Y. Chen (Eds.), Miami, Florida, USA, p. 6012–6025. External Links: Link, Document Cited by: §1, §3.4. Yang et al. (2025) A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lv, C. Zheng, D. Liu, F. Zhou, F. Huang, F. Hu, H. Ge, H. Wei, H. Lin, J. Tang, J. Yang, J. Tu, J. Zhang, J. Yang, J. Yang, J. Zhou, J. Zhou, J. Lin, K. Dang, K. Bao, K. Yang, L. Yu, L. Deng, M. Li, M. Xue, M. Li, P. Zhang, P. Wang, Q. Zhu, R. Men, R. Gao, S. Liu, S. Luo, T. Li, T. Tang, W. Yin, X. Ren, X. Wang, X. Zhang, X. Ren, Y. Fan, Y. Su, Y. Zhang, Y. Zhang, Y. Wan, Y. Liu, Z. Wang, Z. Cui, Z. Zhang, Z. Zhou, and Z. Qiu Qwen3 technical report. External Links: 2505.09388, Link Cited by: Appendix A, §G.4, §4.2. Yang et al. (2023) Y. Yang, J. Kim, Y. Kim, N. Ho, J. Thorne, and S. Yun HARE: explainable hate speech detection with step-by-step reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2023, H. Bouamor, J. Pino, and K. Bali (Eds.), Singapore, p. 5490–5505. External Links: Link, Document Cited by: §2.2. Zampieri et al. (2023) M. Zampieri, S. Morgan, K. North, T. Ranasinghe, A. Simmmons, P. Khandelwal, S. Rosenthal, and P. Nakov Target-based offensive language identification. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), A. Rogers, J. Boyd-Graber, and N. Okazaki (Eds.), Toronto, Canada, p. 762–770. External Links: Link, Document Cited by: §1, §2.1, §3.1, §4.1, §4.1. Appendix A Implementation and Artifact Details Table 9 gives the retained Qwen3-14B configuration Yang et al. (2025). Phase 1–3 share the same local backbone, and Phase 4 is deterministic. The compact profile uses vLLM Kwon et al. (2023). 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 Phase 2 3 roles; micro-batch 64; Chinese bystander with lexicon injection Phase 3 quality threshold 0.60; cluster similarity 0.66; auditor enabled; max retained clusters 2 Runtime safeguards checkpoint resume; OOM split-retry Compact service profile vLLM backend; auditor batch 1; regeneration disabled Phase 4 deterministic reassembly Table 9: Main local configuration for the reported Qwen3-14B mainline. A.1 Runtime and Call Diagnostics Retained EN, ZH-main, and ZH-quadruple runs take 499.0, 617.6, and 623.3 minutes. Phase 2 makes 4,524, 6,978, and 6,933 role calls at 0.25–0.28 role-groups per second. Auditor refinement is triggered for 982, 1,645, and 1,610 groups. These figures describe retained runs; incomplete billing logs preclude a monetary cost estimate. A.2 Artifact Map and Recoverability The artifact roots SPAR/, SPAR_anonymous_submission/, and transfer_bundles/ contain the deterministic TBO split and Phase 0 preparation path, Phase 1–3 prompts and configurations, Phase 3 scoring and fallback code, direct and API baseline outputs, SynChain and Dance adaptations, diagnostic summaries, and teacher/student exports. Historical bilingual distillation records are partly archival. Preserved reports and official evaluation exports support the reported construction and evaluation statistics, while the Chinese student bundle retains the independently recoverable training example. Appendix B Prompt Templates The three Phase 2 roles share the output contract, Target Resolution, Non-empty Argument, Output Sequence, Short Rationale, and Current Task blocks. Figure 4 gives the Victim instruction body and example headers. The Moderator and Cultural Bystander figures retain their role-specific additions. Repeated example titles are omitted. Phase 1 Segmenter # Role You are the SPAR-Hate Segment Agent specializing in English text. Your ONLY output is a strict JSON array. # Rules 1. Divide by target and intent: Split the text into separate groups if different targets are attacked or distinct intents exist. 2. Context isolation: group_text MUST be copied EXACTLY from the original text as a meaningful local context. Do not paraphrase, summarize, or alter punctuation/emojis. 3. Target normalization: canonical_target is the direct entity mention. If the target is strictly absent, unmentioned, or null, output “IMPLICIT_TARGET”. 4. Output scope: ONLY output local_group_id, group_text, and canonical_target. # Examples Example 1 (Multi-Target Segmentation) Example 2 (Null/Implicit Targets & General Venting) Example 3 (Non-harmful / Object Target with Emojis) # Current Task Original text: TEXT Output: [ Figure 3: Excerpt of the English Phase 1 segmenter prompt used in the bilingual main tracks. Victim # Role You are the “Victim Persona” Agent in a Hate Speech Evaluation system specializing in English text. You are highly sensitive, empathetic to marginalized groups, and acutely aware of emotional harm, exclusion, stereotyping, slurs, and microaggressions. Your ONLY output is a strict JSON array. # Rules 1. Victim’s Lens: Put yourself in the shoes of the attacked group. If the text inside the focus area makes you feel unsafe, dehumanized, degraded, or stereotyped, flag it as harmful (harmful: 1). 2. Attention Lock (CRITICAL): Your evaluation area is STRICTLY limited to the text inside the <focus> and </focus> tags. You MUST NOT penalize offensive behaviors or slurs that exist outside these tags. 3. Target Resolution: prioritize the target inside <focus>; use outside context or Target hint only when the focus text is fragmentary; do not override an explicit in-focus target. If the focus only contains an insult or adjective but no addressee, output “IMPLICIT_TARGET”. 4. Argument Extraction: argument MUST be an exact substring from INSIDE the <focus> tags and should be the shortest decisive span that carries the insult, stereotype, exclusion, or offensive action. 5. Multi-Argument Extraction (CRITICAL): Split independent attacks into separate objects. Do NOT merge two insults or predicates into one long argument just because they are joined by and, &, commas, or nearby context. 6. Non-empty Argument (CRITICAL): argument must NEVER be empty. Even for harmless outputs (harmful: 0), choose the shortest exact substring inside <focus> that best supports the harmless judgment. 7. Output Sequence: each object must strictly follow rationale, target, argument, harmful. 8. Short Rationale: rationale must be brief (under 30 words), single-line only, and must not contain raw newline characters inside the JSON string. # Examples Example 1 (Missing Subject Resolved via Context) Example 2 (Multiple Arguments on the Same Target within Focus) Example 3 (Multi-Argument Extraction on Permissible/Non-Harmful Text) # Current Task Target hint (soft only; use only when focus lacks a subject): "TARGET_HINT" Input: "FOCUS_TEXT" Output: [ Figure 4: Main instruction body of the English Victim prompt used in Phase 2. Moderator # Role You are the “Moderator Persona” Agent in a Hate Speech Evaluation system specializing in English text. Your duty is to enforce social media platform Terms of Service (TOS) and community safety guidelines. You look for actionable offenses: severe slurs, targeted harassment, incitement of violence, and hate speech. You tolerate general profanity and subjective disagreements if they do not cross into targeted abuse. # Role-specific rules 1. Moderator’s Lens: Evaluate if the text INSIDE the focus area violates safety policies. Explicit slurs, hostile degradation, threats of violence, or targeted abuse must be flagged as harmful (harmful: 1). 4. Argument Extraction: argument should be the shortest decisive span that contains the actionable abuse, slur, threat, or hostile phrase. # Example headers Example 1 (Severe Violation with Pronoun Resolution) Example 2 (Explicit Slurs and Threats) Example 3 (Harmless/No Violation inside Focus) Figure 5: Role-specific English Moderator prompt used in Phase 2. Shared blocks with Victim are omitted. Cultural Bystander # Role You are the “Cultural Bystander Persona” Agent in a Hate Speech Evaluation system specializing in English text. You act as an internet anthropologist. You focus on the sociological nuances of language: distinguishing reclaimed slang from derogatory usage, identifying group generalization templates, recognizing coded derogation, and differentiating between quoting/irony and genuine hostility. You assess the speaker’s stance and tone. # Role-specific rules 1. Cultural Lens: Evaluate the societal and cultural implications of the text INSIDE the focus area. Recognize when a slur is reclaimed by the community versus when it is weaponized. Flag generalizations and coded derogation as harmful if they enforce negative societal stereotypes. Treat reclaimed slang as non-harmful only when the focus itself shows clear affiliative or celebratory usage. Negative complaints, commands, mockery, or group generalizations are not reclaimed evidence. 4. Argument Extraction: argument should be the shortest decisive span that captures the key cultural attack, slang, stereotype, or behavior. 5. Multi-Target/Argument Extraction (CRITICAL): Split independent attacks into separate objects. # Example headers Example 1 (Reclaimed Slang & Performative Tone) Example 2 (Group Generalization) Figure 6: Role-specific English Cultural Bystander prompt used in Phase 2. Shared blocks with Victim are omitted. Few-shot API Baseline You are a content review expert. Please identify hate speech in the sentence and extract one or more triplets. Your ONLY output is a strict JSON array. Each object must contain: rationale, target, argument, harmful. target: the discussed or attacked entity; if there is no explicit target, output “IMPLICIT_TARGET”. argument: the core evidence span. harmful: must be 1 or 0. Output JSON array only. Do not output any explanation. Positive example: “The men where I live are trash.” Output: (target=men, argument=trash, harmful=1) Negative example: “What kind of dogs are the ugly gross ones with bloodshot eyes and long ass ears” Output: (target=dogs, argument=ugly gross ones with bloodshot eyes and long ass ears, harmful=0) Input: "TEXT" Output: return the answer starting exactly with [ and ending with ] Figure 7: English few-shot baseline prompt shared by the DeepSeek-v4-Pro and GPT-5.4 API baselines. Phase 3 Auditor ## Role You are the SPAR-Hate Phase-3 Arbitrate Auditor for English. You do not perform crude majority vote. You arbitrate between role evidence, candidate clusters, triage signals, and dynamic beacons to select the final tuple most likely to match the English ground truth. Your ONLY output is a strict JSON object. ## English Task Rules 1. Ground the final answer in focus_text. final_argument must be an exact or near-exact substring from the focus span. 2. Prefer final_target from inside the focus span. Only use canonical_target or outside context when the focus span is fragmentary, pronominal, or missing the subject. 5. Distinguish reclaimed slang, quotation, refutation, irony, and performative speech from genuine targeted harm. 7. If one role misses coded derogation or over-penalizes a reclaimed or quoted expression, say so compactly in fused_rationale and correction_trace. 9. If several candidates differ mainly in argument length, prefer the shortest span that still independently expresses the attack core. ## Few-shot example headers Example 1: clear group generalization, final harmful Example 2: reclaimed slang, final non-harmful Example 3: shrink an over-broad argument span Figure 8: Rendered excerpt of the English Phase 3 auditor prompt used in the mainline configuration. Appendix C Baseline Adaptation SynChain and Dance are modified only enough to produce hate tuples compatible with the benchmarks. The aim is to preserve each method’s high-level inductive bias rather than rewrite it into a new multi-stage system Fan et al. (2025); Wang et al. (2026). General adaptation. Across all baselines, the adaptation layer makes only three minimal changes. First, raw outputs are normalised into strict JSON readable by the official evaluation scripts. Second, field names and label spaces are aligned: English uses harmful, Chinese uses hateful, and group is enabled only on the Chinese four-field track. Third, lightweight post-processing is applied only where required by the evaluation interface, including boundary cleaning, duplicate tuple removal, and unified sample-level export. None of these operations adds candidate arbitration or reranking. SynChain. The syntax-aware extraction bias is retained. The adapted workflow still generates structure-sensitive candidates first, then resolves target, argument, and label, and finally maps intermediate candidates back to scorable text fields. Beyond alignment to the output contract, no extra multi-role generation, auditor arbitration, or lexicon weighting is added. Dance. The divide-and-conquer skeleton of grouping, per-group inference, and merge is retained. The adapted system still performs inference over local groups before returning to sample-level outputs. The only change is that the internal extraction target is rewritten from aspect/opinion/sentiment-style fields into hate tuples. As with SynChain, no extra SPAR-style arbitration or lexicon-aware reranking is added. Appendix D Auditor Scoring and Bounded Arbitration Phase 3 scores, clusters, routes, and selects role candidates under grounding constraints. Table 10 gives the main settings, and Table 11 gives the scoring signals. 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 Auditor batch 2 (main 14B); 1 (compact vllm) Regeneration enabled (main 14B); disabled (compact vllm) Cluster retention max 2; score floor 1.05; min support roles 2 Decoding profile bfloat16; TP=2; context 8192 Final selection deterministic fallback after optional refinement 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 Table 11: Main feature groups used by the Phase 3 candidate scorer. The quality threshold of 0.60 filters weak or ungrounded candidates before clustering. The similarity threshold of 0.66 limits merges between candidates with divergent target or argument fields. Both values are reused across the reported languages and benchmarks. Step Operation Bound Validate Check JSON, label, argument grounding, and target anchor quality ≥0.60≥ 0.60 Cluster Compare target, argument, label, and optional group similarity ≥0.66≥ 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 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% Table 13: Boundedness and routing diagnostics across the retained runs. Appendix E Robustness and Semantic Validation E.1 Threshold Sensitivity The actual auditor is replayed on fixed seed-10947 subsets of 200 complete samples per language, covering 385 EN and 281 ZH focus groups per setting. Quality varies over 0.55/0.60/0.650.55/0.60/0.65 at similarity 0.660.66. Similarity varies over 0.60/0.66/0.720.60/0.66/0.72 at quality 0.600.60. All other settings remain fixed. Table 14 reports the retained summary. Language Observed range Default EN 34.97–35.08 35.08 ZH 26.28–27.12 26.82 Table 14: Four-hard-metric averages in the fixed-subset threshold replay. The results are not full-test significance estimates. E.2 Blind Semantic Audit Two independent bilingual or native-speaker reviewers assess 60 stratified Chinese focus units with gold labels hidden. Separate final-tuple and Cultural-Bystander-rationale rubrics yield 120 judgements. Table 15 gives acceptance and agreement. Consensus cases reach 70% strict overall acceptance; disagreement-heavy lexicon–auditor cases receive lower acceptance. Object R1 R2 Agree. κ Final tuple 73.3 71.7 81.7 0.540 Bystander rationale 68.3 58.3 86.7 0.716 Table 15: Acceptance rates, percentage agreement, and Cohen’s κ in the blind semantic audit. E.3 Strict Sample Exact-Set Summary 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 Table 16: Sample exact-set summary. Some ZH few-shot baselines exceed the SPAR-DeepSeek value. E.4 Attacked-Group Breakdown Table 17 reports ZH-quadruple scores by gold attacked group. The Racism+Sexism group has the largest TAH-to-Quad drop, locating the main loss in multi-group binding. The breakdown measures robustness across the annotated groups; fairness requires separate evidence. 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 Table 17: ZH-quadruple attacked-group diagnostic. Scores are Soft F1 and groups are defined from gold labels. Appendix F Qualitative Cases and Error Analysis The Chinese case isolates a target conflict. The English case traces a conflict-lane decision within a multi-target sample. 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 Beacons grounding; consensus Correction trace moderator: downweighted for misreading the victim mention as the attacking target Final tuple (Beijingers, saying that Henan people steal manhole covers, 1) Table 18: Chinese case study showing target-conflict resolution in Phase 3. In this case, Victim and Cultural Bystander both identify the attacking target as “Beijingers”, whereas Moderator shifts the target to the mentioned victim group. The auditor follows the grounded consensus on the focus text and downweights the moderator reading. 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 Beacons grounding, victim_microaggression Correction trace moderator: downweighted; cultural_bystander: downweighted Final tuple (shit, shit, 1) Table 19: English case study showing conflict-lane arbitration inside a multi-target sample. The current text is focus unit g2g_2 in a sample containing g1g_1, g2g_2, and g3g_3. Phase 4 retains two harmful tuples from g1g_1, (shit, shit, 1) from g2g_2, and one non-harmful tuple from g3g_3. One focus unit can emit several tuples. For g2g_2, the auditor selects the Victim reading and downweights the non-harmful Moderator and Cultural Bystander candidates. Appendix G Distillation Data and Student Training G.1 Teacher-trace Schema 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 Supervision messages system instruction, user evidence package, assistant supervision target Table 20: Schema of the final constructed distillation rows. The distillation target includes more than the final tuple. Each training row also preserves intermediate decomposition, role candidates, conflict diagnosis, and the final arbitration result, so that the student learns a structured decision process rather than only the flat output surface. G.2 Filtering and Dataset Statistics 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 Teacher JSON-only plus 3,194 final JSON-aligned variant Validation checks all pass duplicate id 0; schema error 0; invalid final tuple 0 Table 21: Summary statistics for the teacher-trace construction pipeline. The bilingual statistics come from the main distillation construction run. The pipeline has three steps: first, retain train rows with recoverable identifiers and serialisable teacher packages; second, build a strict HQ reasoning subset using groundedness, final tuple validity, and audit confidence; third, expand it under the same grounding constraints into reasoning-plus and JSON-only variants for curriculum training. G.3 A Real Constructed Training Example 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. Cultural Bystander tuple: (men, act like simps, harmful); rationale: [Lexicon Hit] “simp” is used here to belittle men and suggest a lack of dignity. Conflict diagnosis lane=consensus; high-risk field=argument; error tags=lexicon_hit, auditor_span_refined Teacher final tuple target=men; argument=act like simps; group=unknown; harmful=1; direction=harmful Assistant target format: reasoning+JSON; final JSON: target=men; argument=act like simps; group=unknown; harmful=1; direction=harmful Table 22: A condensed real training row from the constructed distillation data. Table 22 shows one real final training row from the released Chinese student bundle. The sample comes from a multi-target Chinese input, but the current training row supervises only the second focus unit. The three role rationales are retained so that the student learns argument anchoring when targets agree but arguments compete. G.4 Student Training Setup Item Value Student backbone Qwen3-4B-Instruct-2507 Prompt template Qwen3 no-think Training split train only Supervision regime teacher-only Curriculum stages teacher reasoning HQ; teacher reasoning plus; teacher JSON-only plus Epochs per stage 1.0 / 1.0 / 1.0 Sequence length 4096 Per-device batch size 1 Gradient accumulation 8 Optimizer and schedule learning rate 1e-4; cosine; warmup ratio 0.05 Precision bf16 Adaptation method QLoRA (4-bit BnB) LoRA setting rank 32; alpha 32; dropout 0.0 Table 23: Student-model training setup used in the distillation pipeline. Student training follows a teacher-only curriculum. Stage 1 uses the strict HQ reasoning subset to establish the structured decision process. Stage 2 expands coverage with reasoning-plus. Stage 3 aligns the model to strict JSON output. Under this protocol, training uses only the train split, ground truth is never used as assistant supervision, evaluation is limited to teacher reasoning and teacher JSON, and manual prefill is disallowed. The Qwen3-4B backbone, 4-bit QLoRA, and LoRA rank/alpha settings follow the corresponding parameter-efficient fine-tuning methods Yang et al. (2025); Dettmers et al. (2023); Hu et al. (2021). H Full Bilingual Ablation Results Table 25 gives the complete bilingual ablations underlying the compact analysis in Section 4.4. (a) ZH-main 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 w/o phase3 46.54 56.64 17.03 51.42 10.67 36.58 7.68 27.10 31.71 (b) EN-main Setting Target Arg. Targeted NTA TA Harm Avg. Qwen3-14B (mainline) 43.45 47.96 34.68 20.96 18.68 59.31 37.51 w/o phase1 segmenter 46.95 47.11 34.06 20.65 18.04 57.21 37.34 w/o phase2 victim 44.70 49.17 32.68 21.18 17.20 57.21 37.02 w/o phase2 moderator 44.29 48.25 34.07 21.34 17.51 59.11 37.43 w/o phase2 bystander 44.49 47.24 33.10 20.64 18.37 59.22 37.18 w/o phase3 43.66 46.77 34.07 20.42 17.81 57.61 36.72 Table 24: Complete bilingual Qwen3-14B ablations. I Full Results on ZH-quadruple Table 25 reports the full four-field results used for the Chinese extension discussed in Section 4.3. Model Target Argument T-A Pair T-A-H Tri. Quad. Avg. Hard Soft Hard Soft Hard Soft Hard Soft Hard Soft Local LLMs 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 LLaMA3-70B* 30.87 41.45 14.80 46.68 8.29 25.38 7.40 22.58 4.72 13.14 21.53 Claude-3.5-Sonnet* 41.45 54.06 15.80 55.80 10.43 36.96 9.28 33.04 7.10 25.22 28.91 few-shot-DeepSeek-v4-Pro 54.86 65.70 18.62 60.50 14.39 41.02 11.34 32.51 9.70 27.99 33.66 few-shot-GPT-5.4 52.87 64.98 17.09 60.28 12.43 42.17 9.45 31.54 8.15 27.50 32.65 SPAR-DeepSeek-v4-Pro 52.66 61.29 23.73 65.66 16.45 43.12 12.70 32.25 10.82 27.65 34.63 SPAR-GPT-5.4 52.15 60.63 21.69 55.34 16.37 43.71 13.15 33.40 11.63 29.11 33.72 Table 25: Full Chinese four-field results. Starred rows are historical STATE-ToxiCN values Bai et al. (2025).