❗️Warning: this document contains content that may be offensive or upsetting.

AmpleHate amplifies target-context relations for implicit hate speech detection, achieving 92.14% better performance than constrastive learning baselines.

AmpleHate amplifies target-context relationships for implicit hate speech detection. Unlike explicit hate, implicit hate is subtle and depends heavily on context rather than offensive language.

While existing models use contrastive learning, humans typically identify hate by first recognizing targets and then evaluating their context. Inspired by this process, AmpleHate uses a pre-trained Named Entity Recognition model to detect explicit targets and [CLS] tokens for implicit cues.

It then applies attention-based mechanisms to model relationships between targets and context, directly integrating these signals into the sentence representation. This approach significantly boosts detection accuracy and achieves state-of-the-art results.

AmpleHate mimics how humans detect implicit hate speech--by first identifying targets and then interpreting context. It improves model focus on key signals through three main steps:

1. Target Identification
   • Uses a pre-trained NER model to extract explicit targets (e.g., groups, places, events)
   • Utilizes the Transformer's [CLS] token to represent impilcit target (overall conetxt of the sentence)
2. Relation Computation
   • Computes attention-based interactions between targets and context
   • These interactions from a relation verctor that captures how target mentions influence the sentence's hateful tone
3. Direct Injection
   • Injects the relation vector directly into the sentence embedding ([CLS])
   • This amplifies relevant target-context signals while reducing noise, making predictions more precise and human-aligned.

AmpleHate outperforms contrastive learning baselines across multiple datasets (marco-F1).

• Consistently outperforms existing baseline in macro-F1 score
• Achievew the highest average marco-F1 scroe of 82.14, surpassing the best baselin (LAHN) by ~5.6%p
• Especially strong on datasets with subtle and context-dependent hate (e.g., White: +27%p over BERT)

These result confirm that amplifying target-context relationship is highly effective for detecting implicit haet speech.

This table shows how AmpleHate idenfities key implicit hate indicators at the token level mirroring human attention patterns.

• For each dataset, AmpleHate assigns higher attention signals to contextially important tokens
• These tokens often don't include explicit slurs, but are essential to understanding the underlying bias or streotype.
• Compared to BERT, AmpleHate places more precise focus on hate-relevant cues, even when no explicit target is tagged.

This demonstrates that AmpleHate's target-aware attention mechanism effectively captures subtle signals in implicit hate speech-making the model both inerpretable and accurate.

Dataset file route: ./data/{dataset_name}
We used the IHC, SBIC, DYNA, Hateval, Toxigen, White, and Ethos datasets.

Install the necessary dependencises using the provided requirements

$ pip install -r requirements.txt

Modify the config/train_config.py file.

$ python train.py

Modify the config/test_config.py file.

$ python test.py

@misc{lee2025amplehateamplifyingattentionversatile,
      title={AmpleHate: Amplifying the Attention for Versatile Implicit Hate Detection}, 
      author={Yejin Lee and Joonghyuk Hahn and Hyeseon Ahn and Yo-Sub Han},
      year={2025},
      eprint={2505.19528},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.19528}, 
}