Abstract
Cyber harassment represents a significant societal challenge, yet automated systems often fail to capture the nuanced dynamics of counseling dialogues. This research introduces a novel framework and dataset designed to distinguish between a victim's distress (Manifestations) and a perpetrator's specific actions (Determinants). We propose E-ConDuCT, a transformer-based model that integrates emotional intelligence and dual-attention mechanisms. Our results demonstrate that E-ConDuCT outperforms current state-of-the-art systems, including general-purpose LLMs like ChatGPT, in identifying these critical components within harassment-related conversations.
Introduction: Defining the Demarcation
Effective intervention in cyber harassment requires more than just identifying a "toxic" event; it necessitates understanding the specific roles of dialogue components. Current automated systems often treat harassment as a monolithic occurrence, lacking the granularity needed for legal or psychological support.
To address this, we define two distinct elements:
- Cyber Harassment Manifestations (CHMs): These are indicators of the victim’s emotional distress or vulnerability (e.g., "I am so terrified").
- Cyber Harassment Determinants (CHDs): These describe the specific behaviors or tactics employed by the harasser (e.g., "He has tried to take advantage of me").
By demarcating these two elements, virtual counselors can provide more targeted assistance, and legal entities can build clearer cases based on documented perpetrator behavior.
The E-ConDuCT Framework
We propose the Emotion-informed Contextual Dual attention Convolution Transformer (E-ConDuCT). This architecture is specifically designed to handle the emotional and contextual complexities of harassment dialogues.
Key Architectural Components
- Emotion-Informed Encoding: The model extracts "Adjective-Noun Pairs" (ANP) to capture latent emotional structures, going beyond standard semantic embeddings.
- Dual Attention (DuA): Unlike standard self-attention, our dual-attention mechanism ranks and selects the most contributory tokens, effectively pruning noise and focusing on core spans.
- EmoMCD Module: This "demarcation module" utilizes task-specific emotional knowledge (based on Ekman’s basic emotions) to differentiate the emotional weight of a manifestation from the descriptive nature of a determinant.
- Multi-Loss Optimization: The framework employs adversarial and zero-shot losses to align text with emotional features, ensuring the model accurately distinguishes between CHMs and CHDs.
Datasets and Experimental Setup
Our research introduces Cyber-MaD, a novel corpus of 996 dialogues (27,666 utterances) manually annotated for CHM and CHD spans. We also validated the model on the CARES CEASE-v2.0 dataset of suicide notes to test its robustness in general emotion-cause extraction.
Key Insights and Results
Our analysis revealed that CHMs and CHDs possess distinct emotional profiles. Manifestations are typically associated with Fear and Sadness, while Determinants often trigger Anger or even "positive" sentiments in contexts involving perpetrator manipulation. Leveraging these nuances significantly boosts extraction accuracy.
Performance Benchmarks
E-ConDuCT consistently outperformed six state-of-the-art baselines, including SpanBERT and BERT-HateXplain:
- Accuracy: On the Cyber-MaD dataset, it achieved a Full Match (FM) score of 35.62 for Manifestations and 33.19 for Determinants.
- Contextual Sensitivity: Performance peaks at a context length of 7 previous utterances, balancing the need for historical information against the risk of introducing noise.
- Comparison with LLMs: Qualitative assessments showed that E-ConDuCT is more reliable than ChatGPT 3.5, which often misclassifies perpetrator actions as victim manifestations.
Conclusion and Future Directions
While E-ConDuCT sets a new benchmark for dialogue comprehension in cyber harassment, several avenues for future work remain. Currently focused on English, the framework should be expanded to include diverse languages and cultural contexts. Additionally, incorporating multimodal inputs—such as images and voice notes—and increasing the granularity of determinant types (e.g., stalking vs. doxing) will further enhance the utility of automated counseling systems.
Summary updated