Research map / Large language models / AI and society
Breaking Barriers: A Paradigm Shift in Technology Accessibility for Individuals with Physical Disabilities
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.
Builds PDCARE, a 6,796-dialogue physical-disability support dataset annotated with OCEAN persona, gender, age, politeness and empathy labels, and trains EDiSS (Phi-3-small with LoRA SFT, then PPO with classifier-based rewards) for persona-, politeness- and empathy-aware support dialogue.
Problem
Individuals with physical disabilities need personalized, polite and empathetic support, but existing conversational systems rarely tailor responses to user persona, gender, age, or use appropriate politeness and empathy strategies across a wide range of disabilities.
Approach
Fine-tune a small LLM on a persona-annotated disability dialogue dataset, then refine it with PPO using a reward built from transformer classifiers for persona-gender-age alignment, politeness strategy and empathy strategy, plus smoothness terms (perplexity and utterance similarity).
Key results
- EDiSS reaches 60.7% UPC, 81.5% PSA and 74.9% ESA on PDCARE test, above DSS (Phi-3-small SFT) at 57.9%, 79.1% and 71.5%.
- EDiSS achieves perplexity 3.60 versus 4.85 for the SFT-only DSS and 5.60 for Llama3-8B.
- In human evaluation by 10 raters on a 1-5 Likert scale, EDiSS scores 3.85 on empathy accuracy versus 3.18 for DSS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add PPO refinement with task-relevance and smoothness rewards | DSS (Phi3-small) · PDCARE test | UPC: 57.9% -> 60.7% | Helped |
| Add PPO refinement with task-relevance and smoothness rewards | DSS (Phi3-small) · PDCARE test | PSA: 79.1% -> 81.5% | Helped |
| Add PPO refinement with task-relevance and smoothness rewards | DSS (Phi3-small) · PDCARE test | ESA: 71.5% -> 74.9% | Helped |
| Add PPO refinement with task-relevance and smoothness rewards | DSS (Phi3-small) · PDCARE test | PPL: 4.85 -> 3.60 | Helped |
| Use no reward instead of both task-relevance and smoothness rewards (ablation) | EDiSS · PDCARE test | UPC: 60.7% -> 57.5% | Hurt |
| Remove task-relevance reward, retaining smoothness reward (ablation) | EDiSS · PDCARE test | UPC: 60.7% -> 58.4% | Hurt |
| Remove smoothness reward, retaining task-relevance reward (ablation) | EDiSS · PDCARE test | UPC: 60.7% -> 59.3% | Hurt |
| Use no reward instead of both task-relevance and smoothness rewards (ablation) | EDiSS · Human evaluation | UPC: 3.15 -> 2.72 | Hurt |
| Remove task-relevance reward, retaining smoothness reward (ablation) | EDiSS · Human evaluation | UPC: 3.15 -> 2.82 | Hurt |
| Remove smoothness reward, retaining task-relevance reward (ablation) | EDiSS · Human evaluation | UPC: 3.15 -> 3.01 | Hurt |
Limitations
- Persona misalignment and sub-optimal politeness or empathy strategies occur in some responses, partly due to limited training data.
- Classifier accuracy (80.2% PGA, 85.4% empathy) limits reward quality and may introduce bias.
- Dialogues are synthetic, generated by Llama3-70B and validated by a small group of annotators, and the evaluation is limited to automatic metrics and human ratings from 10 evaluators.
Experimental setup
Datasets: PDCARE
Benchmarks: PDCARE
Metrics: UPC PSA ESA PPL Rlen Nrep persona accuracy gender-age accuracy politeness accuracy empathy accuracy
Baselines: GPT2-large ARDM Phi-1.5 Llama2-7B Mistral-7B ZEPHYR-7B Phi-2 Mistral-8B Llama3-8B DSS
Builds on: PPO LoRA RoBERTa Phi-3-small
Closest papers by meaning
- Reinforce Trustworthiness in Multimodal Emotional Support System (AAAI 2026 · similarity 0.80)
- EmpathyChat: Structured Cognitive Reasoning in Empathetic Spoken Dialogue (NeurIPS 2026 · similarity 0.80)
- SocialSim: Towards Socialized Simulation of Emotional Support Conversation (AAAI 2025 · similarity 0.80)
- SPASCA: Social Presence and Support with Conversational Agent for Persons Living with Dementia (AAAI 2025 · similarity 0.79)
- Position: Beyond Assistance – Reimagining LLMs as Ethical and Adaptive Co-Creators in Mental Health Care (ICML 2025 · similarity 0.79)
- STAMPsy: Towards SpatioTemporal-Aware Mixed-Type Dialogues for Psychological Counseling (AAAI 2025 · similarity 0.79)
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.