Evaluating Direct Preference Optimization for Personalizing German Automatic Text Simplifications for Persons with Intellectual Disabilities

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

A new framework evaluates Direct Preference Optimization (DPO) for personalizing German Automatic Text Simplification (ATS) systems, specifically for persons with intellectual disabilities. While large language models (LLMs) have improved ATS quality, existing systems lack personalization because they do not incorporate preference feedback during post-training. This work proposes an ATS personalization framework that post-trains LLM-based ATS models using human feedback collected directly from persons with intellectual disabilities, reflecting their preferences for paired text simplifications. The development pipeline includes data collection, model selection, supervised fine-tuning (SFT), DPO post-training, and result evaluation. Findings emphasize the critical need for active participation from target groups in designing personalized, inclusive AI solutions aligned with human preferences.

Key takeaway

For AI Scientists and NLP Engineers developing accessible language technologies, you should integrate Direct Preference Optimization (DPO) with direct human preference feedback. This approach is vital for creating truly personalized Automatic Text Simplification (ATS) systems, especially for persons with intellectual disabilities. Prioritize active participation from your target user groups throughout the design and post-training phases to ensure AI solutions genuinely align with their specific needs and preferences.

Key insights

DPO with human preference feedback is crucial for personalizing LLM-based text simplification for specific user groups.

Principles

• LLM-based ATS benefits from preference feedback.
• Personalization requires target group participation.
• DPO aligns models with human preferences.

Method

The pipeline involves collecting human preference data, selecting models, applying supervised fine-tuning (SFT), performing Direct Preference Optimization (DPO) post-training, and evaluating results.

In practice

• Collect human preference data from target users.
• Apply DPO to fine-tune LLM-based ATS models.
• Involve persons with disabilities in AI design.

Topics

• Automatic Text Simplification
• Direct Preference Optimization
• Large Language Models
• Intellectual Disabilities
• Human Feedback
• AI Personalization

Best for: AI Scientist, NLP Engineer, Research Scientist

Related on AIssential

• Direct Preference Optimization for Chatbot Fine-Tuning: An Empirical Study — DPO offers a computationally efficient and simplified pipeline for competitive LLM chatbot fine-tuning.
• Give Users the Wheel — Integrating LLMs with traditional recommenders via natural language prompts offers nuanced user control over content discovery.
• AFRILANGTUTOR: Advancing Language Tutoring and Culture Education in Low-Resource Languages with Large Language Models — Dictionary-based synthetic data generation and combined SFT+DPO significantly enhance LLM tutoring for low-resource African languages.
• Better Literary Translation: A Multi-Aspect Data Generation and LLM Training Approach — Multi-aspect iterative refinement generates superior literary translation data for LLM training.
• FSPO: Few-Shot Optimization of Synthetic Preferences Personalizes to Real Users — FSPO enables LLMs to personalize responses by meta-learning user preferences from few-shot synthetic data.
• Macro: Enhancing Multilingual Counterfactual Explanations through Alignment-as-Preference Optimization — Preference alignment, specifically DPO, effectively balances validity and minimality in multilingual counterfactual explanation generation for LLMs.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.