OUNLP at TSAR 2025 Shared Task: Multi-Round Text Simplifier via Code Generation

TL;DR

This work tackles readability-controlled text simplification for language learners by exploiting CEFR-guided prompts and multi-round generation. It introduces the notion of CEFR-Gap as a predictor of simplification difficulty and presents two GPT-4o-generated methods: MRS-Rule, a purely rule-based multi-round framework, and MRS-Joint, which combines rule-based steps with prompting to leverage both symbolic and generative strengths. Empirical results on the TSAR-2025 data show that multi-round approaches outperform single-step baselines, with MRS-Joint delivering the best CEFR alignment (lowest $RMSE$) while maintaining meaningful content. The study highlights the promise and limitations of code-generation-based NLP pipelines for controlled text simplification and points to future work on curriculum-informed strategies and expanded evaluations.

Abstract

This paper describes the OUNLP system submitted to the TSAR-2025 Shared Task (Alva-Manchego et al., 2025), designed for readability-controlled text simplification using LLM-prompting-based generation. Based on the analysis of prompt-based text simplification methods, we discovered an interesting finding that text simplification performance is highly related to the gap between the source CEFR (Arase et al., 2022) level and the target CEFR level. Inspired by this finding, we propose two multi-round simplification methods and generate them via GPT-4o: rule-based simplification (MRS-Rule) and jointly rule-based LLM simplification (MRS-Joint). Our submitted systems ranked 7 out of 20 teams. Later improvements with MRS-Joint show that taking the LLM simplified candidates as the starting point could further boost the multi-round simplification performance.

Figures (3)

• Figure 1: Workflow of MRS-Rule (Run 2)
• Figure 2: Workflow of MRS-Joint
• Figure 3: # Simplified sentences that reach the target level across retries for MRS-Rule vs. MRS-Joint.