Sledgehammer or Scalpel? A Fine-grained Adaptive Framework for Implicit Hate Speech Detection
Organizations: School of Information and Control Engineering, China University of Mining and Technology, Xuzhou, China,221100
Abstract
Unlike explicit attacks with obvious profanity, implicit hate speech hides malice within seemingly compliant expressions through metaphors and contextual hints, making its detection in online content review challenging. Many PLM- or LLM-based methods apply the same processing pipeline across samples, which may overlook differences in information requirements and introduce unnecessary computation. We observe that online hate speech is not monolithic but manifests in varied forms. We therefore define three fine-grained categories: Shallow, Targeted, and Context-Dependent. Accordingly, we propose Fine-grained Adaptive Implicit Hate speech Detection (FAID), a novel framework that first performs fine-grained classification and then adapts to specific categories. Specifically, for Shallow samples with surface-identifiable intents, the framework adopts lightweight Prompt-Tuning for rapid classification; for Targeted comments that bind malicious intent to concealed targets, we design knowledge augmentation to iteratively refine the model and reveal hidden targets; for Context-Dependent comments lacking background information, we utilize an agentic framework that automatically generates prompts to evolve context, infer missing background information and identify ambiguous malicious intents. This adaptive architecture focuses computational resources on complex implicit samples while avoiding redundant reasoning for shallow samples. Experiments on four benchmark datasets demonstrate that FAID significantly outperforms SOTA baselines.
Figures & tables
| No. | Example | Categories |
| 1 | Black people should just go back to the cotton fields. | Shallow |
| 2 | The Atlantic is actually a big cup of bubble tea. | Targeted |
| 3 | Look at how ‘brave’ she is for wearing that. | Context-Dependent |
| Datasets | SBIC | LHd | State_ToxiCN | ProsCons | ||||
| Acc | F1 | Acc | F1 | Acc | F1 | Acc | F1 | |
| SKS | ||||||||
| HateBERT | ||||||||
| ConPrompt | ||||||||
| AmpleHate | ||||||||
| Soft | ||||||||
| Variants | Datasets | ||||
| SBIC | LHd | State_ToxiCN | ProsCons | ||
| w/o Shallow Strategy (Test on Cat.0) | Replace w/ Iterative | 75.24 | 52.62 | 72.39 | 77.40 |
| Replace w/ ACE | 88.83 | 71.00 | 83.77 | 76.27 | |
| FAID (Ours) | 81.60 | 63.70 | 79.48 | 89.77 | |
| w/o Targeted Strategy (Test on Cat.1) | Replace w/ PT | 66.63 | 60.35 | 69.89 | 87.70 |
| Replace w/ ACE | 81.64 | 72.53 | 72.58 | 71.72 | |
| Dataset | Cohen’s (A1–A2) | Cohen’s (A1–A3) | Cohen’s (A2–A3) | Fleiss’ (A1–A2–A3) |
| SBIC | 0.6572 | 0.6140 | 0.5881 | 0.6196 |
| LHd | 0.6507 | 0.6396 | 0.6397 | 0.6433 |
| State_ToxiCN | 0.6973 | 0.6137 | 0.6198 | 0.6435 |
| ProsCons | 0.6993 | 0.6977 | 0.7195 | 0.7049 |
| Dataset | Routing method | Router classification | Accuracy of Hate speech detection | ||||
| Acc | Macro-F1 | Overall | Shallow | Targeted | Context- Dependent | ||
| SBIC | BERT | 49.39 | 48.71 | 51.88 | 39.23 | 47.05 | 69.73 |
| Prompt-Tuning (Ours) | 81.29 | 79.48 | 82.32 | 81.60 | 84.45 | 80.93 | |
| Gold-category | – | – | 82.89 | 85.95 | 86.28 | 78.84 | |
| LHd | BERT | 45.86 | 35.57 | 53.83 | 50.82 | 58.75 | 61.88 |
| Prompt-Tuning (Ours) | 85.13 | 83.71 | 69.57 | 63.70 | 81.22 | 68.10 | |
| Datasets | ConPrompt | Soft | DuPL | Ours |
| SBIC | 0.05s | 0.03s | 6.40s | 1.84s |
| ProsCons | 0.06s | 0.04s | 5.33s | 3.57s |
| Category | Examples | True Label | FAID Pred. |
| Shallow | He only got the job because he is black. | Hate | Hate( ✓ ) |
| 这哥们儿打比赛真tm是个疯子,操作太变态了! (This dude plays like a f***ing maniac, his mechanics are absolutely sick!) | No-hate | Hate( ✗ ) | |
| Targeted | Just look at the neighborhood’s crime rate since they moved in. | Hate | Hate( ✓ ) |
| 她晋升得这么快,平时肯定很懂得怎么跟男领导们沟通。 (She got promoted so fast; she must really know how to communicate with male bosses.) | Hate | No-hate( ✗ ) | |
| Context- Dependent | 不愧是摘棉花的,动作确实熟练。 (As expected from a cotton picker, those movements are indeed proficient.) | Hate | Hate( ✓ ) |
| I’m sure you’ll enjoy the new camp facilities. | Hate | No-hate( ✗ ) |
| Datasets | SBIC | LHd | State_ToxiCN | ProsCons | ||||
| Acc | F1 | Acc | F1 | Acc | F1 | Acc | F1 | |
| LLaMA3-8B | 69.07 | 69.03 | 61.94 | 64.04 | 68.35 | 64.65 | 79.35 | 73.38 |
| Qwen-2.5-7B | 72.40 | 72.35 | 64.68 | 65.43 | 69.22 | 65.78 | 80.68 | 80.00 |
| Gemini-2.5 | 72.86 | 73.82 | 61.51 | 63.89 | 65.98 | 60.97 | 74.22 | 77.16 |
| Deepseek-V3 | 76.68 | 78.74 | 63.20 | 64.74 | 74.70 | 73.04 | 91.34 | 91.35 |
| GPT-4 | 82.32 | 82.43 | 69.57 | 71.48 | 69.99 | 66.40 | 81.33 | 81.74 |
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 108 papers