SPU_DeepSimplify at the CLEF 2026 SimpleText Track: A Planning-Based Approach to Scientific Text Simplification

SPU_DeepSimplify at the CLEF 2026 SimpleText Track: A Planning-Based Approach to Scientific Text Simplification

Summary

This paper presents our participation in Task 1 of the SimpleText 2026 Track on scientific text simplification. To address these challenges, we propose a planning-based framework that separates sentence analysis from text generation to improve controllability...

This paper presents our participation in Task 1 of the SimpleText 2026 Track on scientific text simplification. To address these challenges, we propose a planning-based framework that separates sentence analysis from text generation to improve controllability and meaning preservation. The system combines terminology-aware preprocessing, semantic planning, strategy selection, and controlled generation. A lightweight Llama 3.1 8B model is used for semantic planning and simplification strategy selection, providing efficient structured analysis, while a larger Llama 3.3 70B model is employed for final controlled generation to maximize simplification quality and fluency. The planner determines the most suitable operation, such as rephrasing, splitting, deleting, or keeping the sentence unchanged, before the generator produces the final simplified output. Experimental results show that the proposed approach achieves competitive performance, reaching a SARI score of 45.00. The findings suggest that explicit planning can improve the transparency and effectiveness of scientific text simplification.

Keywords

Scientific text simplification Semantic planning Controlled generation Large language models Readability

Researchers

Publication Files