Research map / Large language models / Language, safety and interpretability
Benchmarking Cyber Harassment Dialogue Comprehension through Emotion-Informed Manifestations-Determinants Demarcation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces the Cyber-MaD dataset of cyber harassment counseling dialogues annotated with manifestation and determinant spans, and the E-ConDuCT model that uses emotion-informed adjective-noun features, dual-attention convolution-transformer layers, and a discriminator loss to extract them.
Problem
Distinguishing and extracting cyber harassment manifestations (victim distress or harassment indicators) from determinants (perpetrator actions) in victim-counselor chatbot dialogues, a distinction that was previously unclear.
Approach
A dual-attention convolution-transformer (ConDuCT) combined with emotion-informed ANP autoencoder features, BERT-based Ekman emotion embeddings, and an EmoMD2 demarcation module with a discriminator loss separates manifestation and determinant spans.
Key results
- On Cyber-MaD, E-ConDuCT reaches 35.62 full match for manifestations and 33.19 for determinants, above the best baseline CMSEKI (33.65 and 31.73).
- On CARES CEASE-v2.0 determinant extraction, E-ConDuCT achieves 35.24 full match and 25.71 partial match, versus 33.21 and 20.74 for CMSEKI.
- Ablation shows replacing dual attention with standard self-attention causes the largest drop across metrics.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ANP emotion features (ablation) | E-ConDuCT · Cyber-MaD — Manifestation | FM: 33.73 -> 35.62 | Helped |
| Use dual attention instead of traditional self-attention (ablation) | E-ConDuCT · Cyber-MaD — Manifestation | FM: 32.73 -> 35.62 | Helped |
| Add the EmoMD2 demarcation module (ablation) | E-ConDuCT · Cyber-MaD — Manifestation | FM: 33.41 -> 35.62 | Helped |
| Add discriminator loss LD (ablation) | E-ConDuCT · Cyber-MaD — Manifestation | FM: 34.44 -> 35.62 | Helped |
| Add adversarial loss Lad (ablation) | E-ConDuCT · Cyber-MaD — Manifestation | FM: 34.72 -> 35.62 | Helped |
| Add zero-shot loss Lzl (ablation) | E-ConDuCT · Cyber-MaD — Manifestation | FM: 34.40 -> 35.62 | Helped |
| Increase context length from 1 toward 7 | E-ConDuCT · Cyber-MaD | performance | Helped |
| Increase context length beyond 7 | E-ConDuCT · Cyber-MaD | performance | Hurt |
Surprising findings
- ChatGPT often misclassifies determinant spans as manifestations on cyber harassment dialogues.
Limitations
- Dataset is derived from a single counseling dataset (MHLCD) and covers only English dialogues.
- Gains over strongest baselines are modest on several metrics; ChatGPT comparison is qualitative on few examples.
Experimental setup
Datasets: Cyber-MaD CARES CEASE-v2.0 MHLCD
Benchmarks: Cyber-MaD CARES CEASE-v2.0
Metrics: Full match Partial match Hamming Distance Jaccard Similarity Ratcliff-Obershelp Similarity
Baselines: BiRNN-Attn CNN-GRU BiRNN-HateXplain SpanBERT BERT-HateXplain CMSEKI ChatGPT
Builds on: BERT Transformer Lite convolution (Wu et al. 2019) CMSEKI Vanilla GAN
Closest papers by meaning
- Thousand Voices of Trauma: A Large-Scale Synthetic Dataset for Modeling Prolonged Exposure Therapy Conversations (NeurIPS 2025 · similarity 0.83)
- EmoNet-Face: An Expert-Annotated Benchmark for Synthetic Emotion Recognition (NeurIPS 2025 · similarity 0.82)
- AffectGPT: A New Dataset, Model, and Benchmark for Emotion Understanding with Multimodal Large Language Models (ICML 2025 · similarity 0.81)
- A Multi-Task Benchmark for Abusive Language Detection in Low-Resource Settings (NeurIPS 2025 · similarity 0.81)
- VidEmo: Affective-Tree Reasoning for Emotion-Centric Video Foundation Models (NeurIPS 2025 · similarity 0.81)
- MME-Emotion: A Holistic Evaluation Benchmark for Emotional Intelligence in Multimodal Large Language Models (ICLR 2026 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.