AIWizards at MULTIPRIDE: A Hierarchical Approach to Slur Reclamation Detection

1. Problem Definition

The paper addresses Subtask B of the MultiPRIDE shared task at EVALITA 2026, which focuses on detecting slur reclamation in social media.

• The Core Challenge: Determining whether a slur (specifically LGBTQ+ related terms) is used in a reclamatory manner (in-group, empowering, neutral) or a non-reclamatory manner (derogatory, hate speech).
• The Ambiguity: Identical lexical items can convey either abuse or solidarity depending on the speaker's identity, social context, and intent. Text-only models often fail because they cannot infer speaker intent without extra-linguistic context.
• The Input: Unlike standard hate speech detection, this task provides user biographies alongside tweets, allowing models to leverage author-level context to resolve ambiguity.
• Data Characteristics: The dataset exhibits significant class imbalance (approx. 80% non-reclamatory vs. 20% reclamatory) in both Italian and Spanish.

2. Methodology

The authors propose a hierarchical, dual-encoder architecture that decomposes the task into two stages to explicitly model the relationship between user identity and language use.

A. Core Hypothesis

The system is built on the assumption that members of the LGBTQ+ community are statistically more likely to use certain slurs in a reclamatory manner. Therefore, inferring a user's "affiliation" or identity signals can aid in classifying the intent of the slur.

B. The Two-Stage Hierarchical Approach

Stage 1: Weakly Supervised User Profiling (The Proxy Task)

• Objective: Train a User Encoder to learn latent representations of user identity.
• Data Labeling: Since ground-truth identity labels are unavailable, the authors use LLM-assisted weak supervision. They use an instruction-tuned LLM (DeepSeek-V3.2) to generate binary "proxy labels" (Affiliated vs. Not Affiliated) based on the concatenation of the tweet and user biography.
• Training: The User Encoder is fine-tuned to predict these proxy labels. This forces the model to learn sociolinguistic signals (pronouns, self-descriptors, community markers) without claiming to verify actual protected attributes.

Stage 2: Slur Reclamation Detection (The Main Task)

• Architecture: A Dual-Encoder system with Gated Fusion.
  1. Text Encoder: A BERT-like model fine-tuned directly on the slur reclamation task (Tweet + Bio).
  2. User Encoder: The model trained in Stage 1, processing the same input (Tweet + Bio) but optimized for the identity proxy task.
• Fusion Mechanism: The final hidden states of both encoders ($h_{text}$ and $h_{user}$) are combined using a learned gating mechanism:
  • A gating vector $g$ is computed via a neural network: $g = \sigma(W_2 \cdot \tanh(W_1 \cdot [h_{text} :: h_{user}]))$.
  • The fused representation is: $h_{fused} = g \odot h_{text} + (1-g) \odot h_{user}$.
  • This allows the model to dynamically balance reliance on the text content versus user context for each specific instance.

C. Training Strategy

• Linear Probing–Fine-Tuning (LPFT): The fusion and classification layers are trained first with frozen encoders, followed by joint fine-tuning of the entire architecture.
• Loss Function: Standard Cross-Entropy was used (Focal Loss did not yield consistent improvements).
• Languages: Separate models were built for Italian and Spanish using language-specific pre-trained backbones (SetFit for Italian, XLM-RoBERTa for Spanish).

3. Key Contributions

1. Hierarchical Modeling Framework: Proposes a novel architecture that separates textual linguistic patterns from user-oriented sociolinguistic signals, fusing them via a learned gate.
2. Weakly Supervised Identity Proxying: Demonstrates a method to leverage LLMs to generate "soft" identity labels for training a user encoder, avoiding the need for sensitive ground-truth data while still capturing community alignment signals.
3. Modular Extensibility: The architecture is designed to be modular, allowing for future integration of multimodal data (images) or social network structures without retraining the entire system from scratch.
4. Ethical Handling: Explicitly frames the user profiling as a contextual cue for NLP tasks rather than a tool for verifying personal identity, mitigating risks of stereotyping.

4. Experimental Results

• Performance:
  • Italian: The dual-encoder model achieved an F1 score of 0.88 (vs. baseline 0.90), with no statistically significant difference ($p=0.28$).
  • Spanish: The dual-encoder model achieved an F1 score of 0.64 (vs. baseline 0.67), with no statistically significant difference ($p=0.17$).
• Comparison to Baseline: The proposed approach achieved performance statistically equivalent to a strong BERT-based baseline. It did not outperform the baseline in aggregate metrics.
• Language Disparity: A significant performance gap existed between Italian and Spanish. The authors attribute this to the baseline architectures (Italian used a specialized SetFit model pre-trained on hate speech, while Spanish used a general-purpose multilingual model) rather than linguistic differences.

5. Significance and Discussion

• Validation of Framework: While the method did not immediately boost accuracy, it validates that incorporating user-level context via hierarchical modeling does not degrade performance. This is crucial for a field where adding complexity often leads to overfitting or noise.
• Error Analysis: The authors identified that the model occasionally misclassifies overtly hostile tweets as "reclamatory" when specific slur terms are present. This suggests the user encoder's signals can sometimes override strong local indicators of abuse, highlighting a limitation in current fusion mechanisms.
• Future Directions: The paper argues that the true value lies in the modularity of the approach. Future work could improve performance by:
  • Refining the fusion mechanism (e.g., contrastive objectives).
  • Integrating richer contextual signals (multimodal data, social graphs).
  • Improving the proxy labeling process to better align with reclamation nuances.

Conclusion

The paper presents a sophisticated, modular approach to a difficult NLP problem. While it did not surpass the strong baselines in this specific iteration, it successfully demonstrates a viable pathway for integrating sociolinguistic user context into hate speech detection systems, offering a flexible foundation for future research into reclaimed language.