We present our CLEF 2025 SimpleText Task 1.1 submission (Lenguaje-Claro team), demonstrating that competitive sentence simplification of scientific text can be achieved with only a 'pinch' of high quality data. Our approach uses GPT-3.5-Turbo, o4-mini and T5-Efficient with zero-shot and three-shot prompt on three sample sentences, complemented by a lightweight ensemble and rule-based model simplifiers. Then, a unified LLM-based judge selects or, if necessary, regenerates outputs below a quality threshold. Experiments show that GPT-3.5-Turbo with a three-shot prompt outperforms all other modules, establishing a good baseline in data-scarce settings.
and three-shot prompting of
GPT-3.5-Turbo
[
• A rule-based model, • A lightweight ensemble (choosing the shorter simplified text) and,
In this section, we describe our methodology in more detail about our participation in SimpleText
shared-task [
We curate three representative pairs of complex-simple sentence samples for a few-shot prompting
with Microsoft Copilot tool [
Each input to GPT-3.5-Turbo and o4-mini is prefixed with the prompt listed in Listing 1. You a r e a h e l p f u l p l a i n l a n g u a g e a s s i s t a n t f o r c l e a r , e a s y , p l a i n , and s i m p l e t e x t s i m p l i f i c a t i o n .
S i m p l i f y t h e f o l l o w i n g s c i e n t i f i c − s t y l e s e n t e n c e i n t o a c o n c i s e , p l a i n − E n g l i s h s e n t e n c e t h a t p r e s e r v e s meaning .
Example 1 : Complex : " . . . " S i m p l i f i e d : " . . . " Example 2 : Complex : " . . . " S i m p l i f i e d : " . . . " Example 3 : Complex : " . . . " S i m p l i f i e d : " . . . " Now s i m p l i f y : Complex : " . . . " S i m p l i f i e d ( l i m i t t o max . 45 c h a r a c t e r s ) :
Listing 1: Prompt for GPT-based models.
On the other hand, for the T5-Efficient-small model we use a more simple prompt (see Listing 2). These minimal settings ensure that the model learns key simplification patterns without extensive ifne-tuning.
S i m p l i f y : I n p u t : " . . . " O u t p u t : " . . . " I n p u t : " . . . " O u t p u t : " . . . " I n p u t : " . . . " O u t p u t : " . . . " S i m p l i f y : I n p u t : " . . . " O u t p u t ( l i m i t t o max . 45 c h a r a c t e r s ) :
In addition to the aforementioned models, we implement a rule-based simplifier and ensemble model. Rule-based simplifier The rule-based simplifier removes parentheticals and splits at discourse markers. We used these markers {and, but, or, so, because} for the simple approach, and then add these for the complex approach {although, however, therefore, moreover, meanwhile, nevertheless, nonetheless, yet, still, then, thus, consequently}.
Ensemble simplifier To improve robustness and ensure output quality in low-data scenarios, we
implemented a lightweight heuristic ensemble strategy that selects the best simplification among
available candidates. Specifically, we collect output from the rule-based, T5-based, and optionally
GPTbased modules, discard any empty or whitespace-only strings, and then choose the shortest nonempty
output inspired by readability studies linking brevity to simplicity. In a quantitative analysis, 68% of the
shortest candidates retained the full factual content while removing peripheral clauses, justifying the
length as a proxy for simplicity [
In cases of ties, we apply a fixed priority: GPT > T5 > Rule. This simple yet efective selection mechanism favors brevity and leverages the higher performance of GPT-based simplifications when available. Our ensembler uses the logic shown in Appendix A.
The outputs with the three highest SARI metric scores (see the results in Table 1) are passed as candidates to our LLM-as-a-judge Unified Evaluator, which scores each simplification candidate in terms of fluency , adequacy, and simplicity on a scale from 1 to 5. This scoring is based on a consistent prompting format in which the original and simplified sentences are provided, and the model is asked to return three comma-separated numbers (one per criterion).
The evaluator is designed to be model-agnostic and supports both local Hugging Face models (e.g.,
LLaMA, Gemma, etc.) and remote deployments via Azure OpenAI [
P l e a s e s i m p l i f y t h i s s c i e n t i f i c s e n t e n c e i n t o c o n c i s e , p l a i n E n g l i s h ( l i m i t t o max . 4 5 c h a r a c t e r s ) : " . . . " .
Listing 3: Unified Evaluator’s fallback prompt.
The freshly generated simplification replaces the weaker candidates. The final selection always favors the highest-rated candidate or the regenerated version when necessary.
This unified evaluation pipeline enables automatic quality assessment and controlled generation in a consistent and scalable manner across diferent model settings.
We evaluated our systems and models on the oficial Task 1.1 test set. Additionally, we performed experiments with the text length of complex sentences, in order to define the adequate length of the simplified sentences ( Truncation in Table 1). We set the optimal length to 45 characters. Table 1 summarizes the performance.
Although absolute diferences are modest, GPT-3.5-Turbo with three-shot prompting consistently surpasses all other modules by +1.8 SARI average. Fallback generation recovers 0.3 SARI when initial candidates fail.
We demonstrate that high quality sentence simplification can be obtained with minimal and synthetic data via strategic prompt engineering and a modular ensemble. GPT-3.5-Turbo with three-shot prompting sets a new low-resource baseline for CLEF SimpleText Task 1.1. Future work will explore automated example selection, use no-synthetic few-shot data, try more setups, and domain adaptation.
This research was supported by cloud credits from Fujitsu’s Microsoft Azure subscription. The authors thank the organizers of CLEF 2025 and SimpleText Lab 2025 for creating the perfect ecosystem for the shared task.
During the preparation of this work, the authors used Microsoft Copilot and Writefull, in order to: Grammar and spelling check. Further, the authors used DeepL in order to: Translation. After using these tools and services, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content.
The corresponding logic is implemented in the ensemble component as shown below in Python
language:
" " "
s e l e c t s h o r t e s t o f non −empty o u t p u t s ,
t i e − b r e a k by p r i o r i t y [ ' gpt ' , ' t 5 ' , ' r u l e ' ]
" " "
c a n d i d a t e s = { ' r u l e ' : r _ o u t , ' t 5 ' : t 5 _ o u t , ' gpt ' : g p t _ o u t }
f i l t e r e d = { k : v f o r k , v i n c a n d i d a t e s . i t e m s ( ) i f v . s t r i p ( ) }
i f n o t f i l t e r e d : r e t u r n r _ o u t
p i c k s h o r t e s t , t h e n by p r i o r i t y
b e s t = min (
f i l t e r e d . i t e m s ( ) , key = lambda x : ( l e n ( x [