Not stated in source 2026 Needs review

Catalan Text Simplification with Large Language Models

Visit original source

Abstract

Automatic text simplification (ATS) is crucial for accessibility, yet it remains underexplored for low-resource languages. This thesis investigates sentence-level ATS for Catalan using large language models (LLMs). A performance analysis of three LLMs is conducted using the iDEM corpus as a benchmark and SARI as the ATS evaluation metric. Diverse prompt engineering techniques are employed to examine their impact on simplification quality. This work introduces a novel reward function that combines SARI with specific penalty terms to achieve better alignment. The reward function is designed for Group Relative Policy Optimization (GRPO) to guide LLMs toward a targeted simplification style. The effectiveness of GRPO with this reward function is demonstrated by post-training IberianLLM-7B-Instruct on the ASSET dataset. Following post-training, the model’s ATS performance improves, while also successfully suppressing negative behaviors. Cross-lingual transfer learning is explored by translating ASSET into Catalan and Spanish and post-training the model on them.

Keywords

Not available

Specialisations

Similar theses