When translating complex technical jargon into plain language—whether for science journalism, patient health portals, or educational accessibility—the primary goal is to lower the cognitive barrier for the reader. In natural language processing, this task is known as Automatic Text Simplification (ATS).
Intuitively, making a text simpler requires reducing its linguistic complexity. However, a persistent challenge across this research direction is the simplicity-adequacy trade-off: as a text becomes easier to read, it frequently loses its factual precision. When models optimize heavily for readability, they often drop critical nuances, over-delete essential information, or introduce factual distortions and hallucinations. Understanding why this trade-off occurs—and how models fail or succeed in navigating it—is central to building reliable simplification systems.
The Mechanics of Information Loss: Over-Deletion and Generalization
To lower the reading grade level of a document, automated systems rely on operations like sentence splitting, phrase deletion, and lexical substitution. When unconstrained, these operations frequently lead to a phenomenon known as over-deletion, where the model prunes away the core meaning of the source text to satisfy readability constraints.
Sweta Agrawal et al. exposed the severity of this issue in Do Text Simplification Systems Preserve Meaning? A Human Evaluation via Reading Comprehension. By using human reading comprehension tests to evaluate state-of-the-art simplification models, the authors discovered that even top-performing systems render 14% to 16% of key facts completely unrecoverable. When models prioritize surface-level brevity over context, they remove the exact details a reader needs to answer basic questions about the text.
This aggressive pruning is often an artifact of how simplification models are optimized. For instance, Xingxing Zhang developed a reinforcement learning framework in Sentence Simplification with Deep Reinforcement Learning that rewards models for generating short, simple sentences. However, the author noted a distinct failure mode: heavily optimizing for simplicity rewards causes the model to sacrifice semantic adequacy, aggressively deleting information to achieve a lower readability score.
[Complex Source] ➔ [Optimization for Simplicity Rewards] ➔ [Aggressive Deletion] ➔ [Loss of Core Meaning]
Beyond deleting text, models lower complexity by swapping specialized terminology for common words. As Leonardo F. R. Ribeiro et al. demonstrate in Generating Summaries with Controllable Readability Levels, there is a measurable inverse relationship between text readability and information specificity. As models target simpler reading levels, they systematically replace precise technical terms with broad, general descriptions, eroding the factual density and nuance of the original document.
The Abstractive Dilemma: Hallucinations and Distortions
When models move past simple word-swapping and attempt to rewrite text abstractively, the risk shifts from omission to active distortion. To make a complex concept accessible to a lay audience, an abstractive model cannot merely copy text; it must synthesize explanations for technical terms. This process introduces a high risk of hallucination.
The mechanics of this tension were mapped out by Sebastian Joseph et al. through the FactPICO: Factuality Evaluation for Plain Language Summarization of Medical Evidence benchmark. The authors found that abstractive rewriting frequently introduces erroneous explanations of technical terms. While extractive methods preserve factual accuracy by maintaining the original phrasing, they fail to reduce linguistic complexity for lay audiences.
This explains why large language models (LLMs) often struggle with simplification compared to standard summarization. As Chantal Shaib et al. observe in Summarizing, Simplifying, and Synthesizing Medical Evidence using GPT-3 (with Varying Success), the simplification process is uniquely prone to introducing minor hallucinations. Because the model is actively attempting to bridge conceptual gaps for a non-expert audience, its generative behavior drifts away from strict source grounding, leading to unintended semantic shifts.
High-Stakes Consequences in Specialized Domains
In high-stakes domains like medicine or psychiatry, the simplicity-adequacy trade-off has real-world safety implications. A minor semantic shift or an omitted detail can transform a nuanced clinical finding into dangerous misinformation.
Ashwin Devaraj et al. highlighted this hazard in Evaluating Factuality in Text Simplification, noting that standard automated evaluation metrics like SARI prioritize linguistic accessibility and fail to detect factual distortions. When applied to clinical letters, Matthew Shardlow et al. showed in Neural Text Simplification of Clinical Letters with a Domain Specific Phrase Table that standard neural architectures risk deleting essential clinical prognosis details, which can lead to dangerous misunderstandings regarding a patient's medical status.
Furthermore, this cross-complexity mapping remains highly inconsistent even for advanced models. Evaluating LLMs on psychiatric materials, Ankit Aich demonstrated in Language models in digital psychiatry: challenges with simplification of healthcare materials that models exhibit high variance and a tendency to drift off-topic, underscoring the difficulty of preserving core semantic meaning when forcing text into a lower-complexity target domain.
Designing for Semantic Fidelity
To overcome the simplicity-adequacy trade-off, researchers are developing architectures that constrain generative models, ensuring that vocabulary reduction does not break semantic integrity.
One approach focuses on providing structured guidance over vocabulary transitions. Seyed Ali Bahrainian et al. demonstrate in Text Simplification via Adaptive Teaching that explicitly controlling lexical transitions leads to significant improvements in human-perceived correctness and fluency. By managing how technical terms map to simpler synonyms, systems can preserve the original meaning of a document without triggering unconstrained, hallucination-prone rewriting.
Another strategy abandons free-form generation entirely in favor of iterative editing. In Iterative Edit-Based Unsupervised Sentence Simplification, Dhruv Kumar et al. proposed an unsupervised method that simplifies text by applying discrete, targeted operations—such as word substitution and phrase deletion—directly to the source sentence.
[Complex Sentence] ➔ [Discrete Substitution / Deletion Edits] ➔ [Faithful Simple Sentence]
By framing simplification as an explicit editing task rather than generating text from scratch, the model avoids the over-deletion and hallucination traps of standard sequence-to-sequence models, proving that text can be made accessible while remaining strictly faithful to the source.
Key Foundational Work
Motivations and Methods for Text Simplification (1996) — contributed 3 concept elements to this direction.