Abstract
Current text simplification (TS) systems often operate as "black boxes," attempting to simplify every input sentence regardless of its inherent complexity. This approach frequently leads to over-simplification, semantic loss, and a lack of transparency. This paper proposes a modular, explainable pipeline that prioritizes two critical preliminaries: predicting whether a sentence requires simplification and identifying the specific tokens contributing to its complexity. Our empirical study demonstrates that integrating these steps can significantly enhance the performance of state-of-the-art models, reducing errors by up to 50%.
Introduction
Text simplification is vital for making complex information accessible to diverse audiences, including children and non-native speakers. However, treating TS solely as a machine translation problem—mapping complex sentences to simple ones—introduces significant flaws. Models often "hallucinate" or distort meaning when processing already simple text. Furthermore, the lack of transparency in neural models makes it difficult to understand why specific edits are made. We argue that the field has overlooked essential preliminary steps: complexity prediction and explanation.
Proposed Framework
We decompose the text simplification task into a three-step pipeline designed to improve both accuracy and interpretability.
- Complexity Prediction (1A): A binary classifier determines if a sentence is complex enough to warrant simplification.
- Complexity Explanation (1B): The system highlights specific tokens (words or punctuation) that drive the complexity.
- Simplification Generation (2): The rewriting engine processes only those sentences identified as complex, guided by the highlighted rationales.
Experimental Setup
The study evaluates this framework across three diverse datasets: Newsela (news articles), WikiLarge (Wikipedia alignments), and Biendata (scientific papers matched with press releases). We compared traditional "shallow" models (e.g., Naive Bayes, Random Forests) against deep learning architectures (CNNs, LSTMs) and pre-trained transformers including BERT, XLNet, and ULMFiT. For the explanation component, we utilized interpretability tools like LIME and SHAP, along with a novel Extractive Adversarial Network.
Results and Key Insights
Our results indicate that pre-trained transformers are highly effective at the initial prediction task.
As shown in the table above, models like BERT and XLNet achieved approximately 80% accuracy on news and Wikipedia data, while reaching over 95% accuracy on the scientific Biendata set. The high performance on scientific text suggests that technical jargon serves as a strong indicator of complexity.
Crucially, using the complexity predictor as a "gatekeeper" significantly improved overall system reliability. By filtering out sentences that were already simple, we reduced errors—measured by Edit Distance and Translation Edit Rate—by 30% to 50%. This prevents common failures where models alter the meaning of simple sentences (e.g., changing "In Ethiopia, HIV disclosure is low" to "In Ethiopia, HIV is low").
Conclusion
This research highlights the necessity of moving beyond black-box simplification models. By incorporating explicit complexity prediction and explanation stages, we can create more robust, transparent, and accurate systems. Future work will focus on end-to-end integration, where the identified complexity rationales directly guide the generation process, and developing more nuanced human-centric evaluation metrics.
Summary updated