Abstract
Text simplification (TS) is essential for accessibility, yet current end-to-end neural models often suffer from a "trigger-happy" bias, over-simplifying already clear text and causing meaning loss. This paper proposes a pipeline approach that introduces two "missing preliminaries": Complexity Prediction (Task 1A) and Complexity Explanation (Task 1B). By filtering sentences through these explainable stages before generation, the researchers achieved a 30% to 50% reduction in edit errors across state-of-the-art models.
Introduction
Text simplification aims to make information accessible to diverse audiences, including children, non-native speakers, and individuals with reading disorders. However, the field's recent shift toward "black-box" neural models—treating simplification as a direct machine translation task (Complex → Simple)—has introduced significant flaws.
A primary issue is that these models lack a "null" option; because they are trained on parallel corpora where every sentence is simplified, they often mangle perfectly clear text. Furthermore, these models offer no explanation for their edits, making them difficult for human editors to trust. This research argues that we must first predict if a sentence needs simplifying and which parts are complex before attempting to rewrite it.
Methodology
The proposed framework decomposes the text simplification process into a structured pipeline, moving away from monolithic end-to-end generation.
- Complexity Prediction (Task 1A): A binary classifier determines if a sentence requires simplification. The authors tested a range of models, from classic baselines like Naive Bayes and Logistic Regression to modern pre-trained transformers including BERT, XLNet, and ULMFiT.
- Complexity Explanation (Task 1B): The system highlights specific tokens (words or phrases) that contribute to complexity. The researchers explored several "explainable AI" (XAI) techniques, such as LIME, SHAP, and a novel joint approach called Extractive Adversarial Networks (EAN), which learns to extract rationales and make predictions simultaneously.
- Simplification Generation: The actual rewriting phase, which can be human-driven, computer-assisted, or automated, is only triggered if the preliminary stages flag the text as complex.
Datasets
The study evaluated the framework across three diverse domains to test generalizability:
- Newsela: Professional news articles rewritten at different reading levels.
- WikiLarge: A standard benchmark of aligned sentences from English Wikipedia and Simple English Wikipedia.
- Biendata: A challenging set matching scientific research papers with their corresponding press releases.
Key Results and Insights
Experimental results demonstrate that adding an explainable preliminary step significantly improves the reliability of simplification systems.
- Classification Performance: Pre-trained models like ULMFiT and BERT achieved the highest accuracies for complexity prediction, reaching 80.8% on Newsela and 81.5% on WikiLarge. Notably, predicting complexity in scientific papers (Biendata) was significantly easier, reaching over 95% accuracy.
- Substantial Error Reduction: By using the complexity predictor as a "gatekeeper" to prevent the model from hallucinating edits on simple text, the researchers reduced the Edit Distance (ED) error by 30% to 50% across all datasets. For the ACCESS model on the Biendata set, the Translation Edit Rate (TER) was reduced by 55%.
- The "Simplify-All" Bias: The study confirmed that state-of-the-art models like ACCESS and DMLMTL typically change over 70% of sentences that are already simple. The proposed filter effectively mitigates this over-correction.
Conclusion and Future Work
Decomposing text simplification into explainable preliminaries addresses the inherent biases of black-box neural models. While this paper successfully utilizes complexity prediction as a binary filter, the authors suggest that future work should focus on "constrained generation." This would involve forcing simplification models to specifically target the words highlighted during the explanation phase, further improving the precision and trustworthiness of automated text simplification.
Summary updated