Research map / Detection and segmentation / Open-set and few-shot detection
Enhancing Detection of Relationship Abuse in Text with Multi-task Learning (Student Abstract)
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.
Proposes a multi-task learning DeBERTaV3 framework that jointly detects emotional abuse (UCC multi-label) and physical relationship abuse (Schrading Reddit binary), showing MTL improves emotional abuse detection in preliminary experiments.
Problem
Detecting relationship abuse in text, especially precursory emotional/psychological abuse that precedes physical violence, where NLP work on intimate partner violence is scarce and emotional-abuse labeled data is lacking.
Approach
A shared DeBERTaV3-large encoder with a shared dense feature extractor and task-specific heads, trained on emotional-abuse multi-label classification (UCC) and binary relationship-abuse classification with a weighted sum of losses (weight τ); MTL weightings are compared to single-task models to tune detection to abuse type.
Key results
- MTL Conf. reaches 0.713 test accuracy on task 1 (emotional abuse) versus 0.713 for STL1 and 0.631 for MTL Default; MTL Conf. overall test accuracy 0.754.
- All models reach at least 95% testing accuracy on task 2 (relationship abuse), with STL2 at 0.963.
- Only MTL τ=0.6 and STL correctly classified the manipulation example as abuse, with MTL τ=0.6 at probability 0.9307 vs STL 0.8224; only MTL models flagged the belittling example.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train jointly on emotional-abuse and relationship-abuse tasks instead of Task 1 only | DeBERTaV3 large · UCC test (Task 1) | accuracy: 0.713 -> 0.631 | Hurt |
| Train jointly on emotional-abuse and relationship-abuse tasks instead of Task 2 only | DeBERTaV3 large · Relationship abuse test (Task 2) | accuracy: 0.963 -> 0.957 | Hurt |
| Weight Task 2 more by setting τ=0.4 (ablation) | MTL Default · UCC test (Task 1) | accuracy: 0.631 -> 0.637 | No clear effect |
| Weight Task 2 more by setting τ=0.4 (ablation) | MTL Default · Relationship abuse test (Task 2) | accuracy: 0.957 -> 0.954 | No clear effect |
| Weight Task 1 more by setting τ=0.6 (ablation) | MTL Default · UCC test (Task 1) | accuracy: 0.631 -> 0.705 | Helped |
| Weight Task 1 more by setting τ=0.6 (ablation) | MTL Default · Relationship abuse test (Task 2) | accuracy: 0.957 -> 0.959 | No clear effect |
| Use only UCC training samples with crowdworker confidence at least 0.6 (ablation) | MTL Default · UCC test (Task 1) | accuracy: 0.631 -> 0.713 | Helped |
| Use only UCC training samples with crowdworker confidence at least 0.6 (ablation) | MTL Default · Relationship abuse test (Task 2) | accuracy: 0.957 -> 0.960 | No clear effect |
| Weight emotional-abuse Task 1 more with τ=0.6 rather than use STL2 (ablation) | Relationship abuse detector · One Love manipulation example | probability of abuse: 0.8224 -> 0.9307 | Helped |
| Use multi-task learning rather than single-task learning | Relationship abuse detector · One Love belittling example | correct abuse classification | Helped |
Surprising findings
- The model trained with the 'Confident' subset of UCC shows gains, and MTL models showed a significant gender bias in abuser portrayal, with higher abuse probability when the abuser is male.
Limitations
- Evaluation of emotional abuse detection relies on only three hand-picked custom examples without an expertly labeled emotional abuse dataset.
- Significant gender bias found in model predictions regarding abuser and victim gender.
- Results are preliminary with single runs and small-scale experiments.
Experimental setup
Datasets: Unhealthy Comments Corpus (UCC) Schrading et al. (2015) relationship abuse Reddit dataset
Benchmarks: UCC (task 1 testing accuracy) Schrading et al. (2015) relationship abuse task 2 test set custom inference examples
Metrics: training loss training accuracy testing accuracy predicted probability of abuse
Baselines: single-task learning (STL) models
Builds on: DeBERTaV3 Unhealthy Comments Corpus
Closest papers by meaning
- Distribution Matching for Multi-Task Learning of Classification Tasks: A Large-Scale Study on Faces & Beyond (AAAI 2024 · similarity 0.80)
- A Multi-Task Benchmark for Abusive Language Detection in Low-Resource Settings (NeurIPS 2025 · similarity 0.80)
- Causal Explanation of Quality of Parent-Child Interactions with Multimodal Behavioral Features (Student Abstract) (AAAI 2025 · similarity 0.79)
- Benchmarking Cyber Harassment Dialogue Comprehension through Emotion-Informed Manifestations-Determinants Demarcation (AAAI 2024 · similarity 0.79)
- Multi-Modal Prompting for Open-Vocabulary Video Visual Relationship Detection (AAAI 2024 · similarity 0.78)
- Thousand Voices of Trauma: A Large-Scale Synthetic Dataset for Modeling Prolonged Exposure Therapy Conversations (NeurIPS 2025 · similarity 0.78)
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.