TLDR: Vicomtech developed an Automatic Post-Editing Cycles (APEC) method for adapting Spanish texts to Plain Language and Easy Read. This iterative approach uses Large Language Models to refine initial adaptations, focusing on improving readability and preserving meaning. Their system achieved first place in Plain Language and second place in Easy Read adaptation in the CLEARS challenge, demonstrating a robust way to make content more accessible.
Making information accessible to everyone is a crucial goal, especially for individuals with reading difficulties, cognitive disabilities, or low literacy. This is where text adaptation to simplified language varieties like Plain Language (PL) and Easy Read (ER) becomes essential. These adaptations aim to reduce linguistic complexity while preserving the original meaning, but they cater to different audiences with specific requirements.
Plain Language focuses on clear and concise communication, using common words and short sentences. Easy Read, on the other hand, targets individuals with intellectual disabilities, requiring even shorter sentences (often under 15 words), frequent vocabulary, explicit explanations of complex terms, and supportive visual elements. The CLEARS challenge, organized within the IberLEF 2025 conference, aimed to evaluate automated approaches to both PL and ER adaptation in Spanish.
Vicomtech’s Innovative Approach: Automatic Post-Editing Cycles (APEC)
Vicomtech participated in the CLEARS challenge with a unique framework based on Automatic Post-Editing Cycles (APEC). This approach leverages Large Language Models (LLMs) to generate initial text adaptations, which are then iteratively refined. The core idea is to continuously improve these adaptations until readability and similarity metrics indicate that no further successful refinement can be made.
Initial Adaptation Strategies
Before the APEC cycles, Vicomtech explored several methods for generating initial adaptations using LLMs:
- Zero-shot (ZS): Direct querying of the LLM with specific prompts.
- Few-shot (FS): Enhancing prompts with a few selected examples from the training data. This included variants like random sampling (FS RDM), lexical Retrieval-Augmented Generation (RAG) using BM25 indexing (FS BM25), and semantic similarity RAG with embedding indexing (FS SIM).
- Fine-tuning (FT): Supervised fine-tuning of the LLM on training data.
- Direct Preference Optimisation (DPO): Optimizing the LLM based on contrastive pairs, where human references were preferred over zero-shot generations.
After evaluating these initial methods, the BM25 and DPO variants were selected as the most robust options to feed into the APEC cycles for further refinement.
How APEC Works: Iterative Refinement
The APEC approach involves a cyclical process where an input text and its current adaptation are fed into an LLM. The LLM is prompted to act as both a judge and a post-editing agent. It analyzes the adaptation based on specific Plain Language or Easy Read guidelines and provided demonstrations, identifies issues, and then generates a new, refined adaptation. This new adaptation is then evaluated using two key metrics:
- Embedding Cosine Similarity: Measures how well the meaning of the original text is preserved in the adapted version.
- Fernández Huerta Readability Index: A metric adapted for Spanish, similar to Flesch-Kincaid, which assesses readability based on average sentence and syllable length.
If the average of these two metrics improves, the refined adaptation replaces the previous one. This process is repeated for a fixed number of cycles (five in their experiments), allowing the model to continuously seek improvements. This iterative self-correction mechanism helps address potential hallucinations or deviations from guidelines in initial LLM outputs.
Outstanding Results in the CLEARS Challenge
Vicomtech’s APEC approach demonstrated significant success in the CLEARS challenge. By averaging all official metrics, their submissions achieved:
- First place in Plain Language adaptation.
- Second place in Easy Read adaptation.
Notably, their method outperformed other participants in terms of the readability index for both tasks, showcasing the LLMs’ ability to guide adaptation towards shorter sentences and words while maintaining overall semantic similarity. While the approach sometimes resulted in lower lexical similarity (Bag-of-Words similarity) compared to human references, this was an expected trade-off, as the focus was on improving readability and preserving meaning rather than strictly mimicking reference adaptations.
Also Read:
- Decoding AI’s Writing Style: A Benchmark for Stylistic Variation in LLM-Generated Texts
- Assessing Foundation Models for Planning Assistance
Future Directions and Considerations
The research highlights that APEC relies heavily on automated metrics for assessing adaptation progress. While effective, exploring alternative metrics or incorporating human evaluation could provide even deeper insights into the quality of adaptations. The computational cost of iterative APEC cycles is also a consideration, making it potentially more suitable for offline text adaptation rather than real-time online usage where rapid adaptation is expected. However, the success of initial non-iterative methods like BM25 RAG or DPO-tuned models suggests viable alternatives for such scenarios.
Overall, Vicomtech’s work demonstrates the viability and effectiveness of using Automatic Post-Editing Cycles with Large Language Models for creating highly readable and accessible text in both Plain Language and Easy Read formats.
For more in-depth details, you can read the full research paper here.
