Imagine reading an essay where every individual sentence is perfectly simple, yet the overall text feels completely disjointed. You encounter pronouns with no clear targets, sudden jumps in logic, and an erratic narrative flow.
This is the fundamental pitfall of traditional automatic text simplification. For years, natural language processing treated simplification as a sentence-by-sentence translation task. However, language operates on a global scale. True simplification requires looking beyond the sentence boundary to preserve the cohesive thread of a document.
The Sentence-Level Breakdown
When a model simplifies a sentence in isolation, it lacks the context of the surrounding discourse. This blind spot introduces several critical failures across a text:
- Discourse Cohesion and Narrative Flow: Splitting or deleting sentences can inadvertently destroy the logical transitions that guide a reader. As Sarah E. Petersen demonstrated in Text simplification for language learners: a corpus analysis, human editors carefully manage structural thresholds, whereas automated sentence-level edits frequently degrade narrative flow. Sanja Štajner et al. in Sentence Simplification Capabilities of Transfer-Based Models and Kaijie Mo et al. in ExpertEase: A Multi-Agent Framework for Grade-Specific Document Simplification with Large Language Models emphasize that without a document-level perspective, individual rewrites inevitably conflict with the broader narrative.
- Coreference and Anaphora Resolution: If a sentence containing a subject is modified or deleted, subsequent sentences containing pronouns (like "he" or "it") become orphaned. Xingxing Zhang et al. in Sentence Simplification with Deep Reinforcement Learning and Shachar Mirkin et al. in Confidence-driven Rewriting for Improved Translation note that isolated processing breaks these critical referential links. Even simple operations like sentence splitting can leave fragments ambiguous if the subject isn't explicitly repeated, a problem highlighted by Yongsik Seo et al. in Make Compound Sentences Simple to Analyze: Learning to Split Sentences for Aspect-based Sentiment Analysis.
- Stylistic and Complexity Inconsistencies: Swapping a single word for a simpler alternative can create a jarring stylistic clash if the rest of the sentence remains highly complex. Lucia Specia in UOW-SHEF: SimpLex - Lexical Simplicity Ranking based on Contextual and Psycholinguistic Features and Sunil Kumar Jauhar et al. in SemEval-2012 Task 1: English Lexical Simplification argue that lexical simplification must be part of a holistic adjustment to avoid these artificial mismatches. Furthermore, Arthur Mariano Rocha De Azevedo Scalercio et al. point out in Enhancing Sentence Simplification in Portuguese: Leveraging Paraphrases, Context, and Linguistic Features that ignoring document context makes it nearly impossible to maintain a uniform level of reading complexity across a text.
The Mechanics of Document-Level Operations
Document-level simplification isn't just sentence simplification with a larger context window; it introduces entirely new editing operations that are fundamentally impossible at the sentence level.
Sentence-Level (1-to-1):
[Complex Sentence A] -------> [Simple Sentence A]
[Complex Sentence B] -------> [Simple Sentence B]
Document-Level (Many-to-Many & Global Curation):
[Complex Sentence A] ---\---> [Merged & Simplified Sentence AB]
[Complex Sentence B] ---/
[Complex Sentence C] -------> [Deleted (Globally Non-Essential)]
Information Redistribution and Merging
Professional editors rarely translate sentences one-to-one. Instead, they perform complex many-to-many transformations—merging ideas from multiple sentences or moving content across paragraph boundaries to improve clarity. Sofia Blinova et al. address this in SIMSUM: Document-level Text Simplification via Simultaneous Summarization by combining summarization and simplification to handle content reordering and structural transformations.
Strategic Content Deletion
To reduce cognitive load, non-essential background information must be pruned. However, determining what is "safe to delete" requires a global understanding of the text's core message. Bohan Zhang et al. showed in Predicting Sentence Deletions for Text Simplification Using a Functional Discourse Structure that incorporating the functional discourse structure of a document significantly improves a model's ability to predict which sentences to delete, while Reno Kriz et al. emphasize in Complexity-Weighted Loss and Diverse Reranking for Sentence Simplification that surrounding discourse is vital to ensure deletions do not compromise essential meaning.
New Benchmarks and Evaluation Paradigms
To transition the field toward holistic document processing, researchers have developed new datasets and metrics that move away from rigid sentence-to-sentence comparisons.
Document-Scale Corpora
Datasets like SWiPE (Philippe Laban et al. in SWiPE: A Dataset for Document-Level Simplification of Wikipedia Pages), D-Wikipedia (Renliang Sun et al. in Document-Level Text Simplification: Dataset, Criteria and Baseline), and VTechAGP (Ming Cheng et al. in VTechAGP: An Academic-to-General-Audience Text Paraphrase Dataset and Benchmark Models) capture real-world human editing strategies. These include paragraph restructuring, background elaboration, and the preservation of scientific narratives across full technical abstracts.
Context-Aware Evaluation
Traditional metrics like SARI evaluate text locally, often penalizing models when content is simply relocated rather than deleted. Ashwin Devaraj et al. demonstrated in Evaluating Factuality in Text Simplification that sentence-level evaluation frequently misidentifies valid information redistribution as factual errors.
To fix this, learnable metrics like LENS (Mounica Maddela et al. in LENS: A Learnable Evaluation Metric for Text Simplification) and human reading comprehension frameworks (Sweta Agrawal et al. in Do Text Simplification Systems Preserve Meaning? A Human Evaluation via Reading Comprehension) are being used to measure how well meaning, factual integrity, and coherence are preserved across entire paragraphs and documents.
Key Foundational Work
Text simplification for language learners: a corpus analysis (2007) — contributed 2 concept elements to this direction.
Experimenting with the Interaction between Aggregation and Text Structuring (2000) — contributed 1 concept elements to this direction.