Imagine reading a sentence where every individual word is familiar, yet the overall meaning remains entirely out of reach. This happens when simple vocabulary is trapped inside a labyrinth of nested relative clauses, passive voice constructions, and long-range grammatical dependencies. To make such a text accessible, a reader does not need a dictionary; they need the sentence architecture to be dismantled and rebuilt.
This distinction highlights the two primary axes of automatic text simplification (ATS): lexical simplification (swapping difficult words for easier synonyms) and syntactic simplification (altering complex grammatical structures). While modern natural language processing models excel at word-level substitutions, they consistently struggle with structural transformations. This friction is known as the syntactic simplification gap.
The Lexical Bias of Sequence Models
Why do automated systems default to word swapping? The root of the problem lies in how text simplification has historically been framed and optimized. Early approaches treated simplification as a monolingual machine translation task (Wang et al.) or relied heavily on word-level databases (Glavaš and Štajner). Because sequence-to-sequence models naturally optimize for local token transitions, they find it far easier to learn word-level substitutions than global structural rearrangements.
As a result, modern models frequently "cheat" by prioritizing simple vocabulary shifts over deep grammatical restructuring (Scalercio et al.). This creates a major usability bottleneck: technical jargon or complex medical instructions might be translated into common words, but those words remain constrained within confusing, clunky syntactic frameworks that fail to lower the reader's actual cognitive load (Sakakini et al.; Hazim et al.).
[Original Complex Sentence]
"The patient, who had previously exhibited resistance to standard therapies,
was administered the compound by the clinical staff."
│
▼ (Lexical-Only Simplification)
"The patient, who had previously shown resistance to normal therapies,
was given the mix by the clinic staff."
(The words are easier, but the passive voice and nested clause remain intact.)
Why Sentence Splitting Fails
The most critical operation in syntactic simplification is sentence splitting—the act of breaking a single, multi-clause sentence into multiple independent, shorter sentences. Human editors rely heavily on this strategy to make text readable. However, neural architectures struggle to execute it reliably.
To understand why, consider how edit-based models operate. Many systems simplify text by predicting token-level operations like KEEP, DELETE, or ADD. As Dong et al. demonstrate, these word-level edit operations are inherently blind to sentence structure; they lack the mechanism required to duplicate subjects, fix broken pronouns, or manage the division of independent clauses. Even when models attempt to drop subtrees to shorten a sentence, they often fail to rephrase or split the remaining components, leading to ungrammatical or disjointed fragments (Bingel and Søgaard).
Furthermore, this structural rigidity is reinforced by a lack of training data. Standard sentence simplification benchmarks are heavily biased toward lexical substitution and content deletion (Narayan et al.). Because parallel corpora contain very few examples of true structural decomposition, neural models rarely learn how to break down complex information without losing critical content or destroying the logical flow (Surya et al.; Jonathan Mallinson).
The Decoupling of Words and Grammar
A core insight of recent research is that lexical and syntactic complexity are entirely distinct dimensions of language that do not naturally correlate. A text can be syntactically complex while using elementary vocabulary, or syntactically simple while using highly advanced words.
Daiki Nishihara demonstrated this decoupling by showing that models forced to shorten sentences often fail to replace advanced vocabulary, creating sentences with simple structures that still contain inappropriately advanced words. Conversely, the authors of Controllable Sentence Simplification via Operation Classification found that when models successfully split a complex sentence, they often fail to simplify the vocabulary within those newly created segments.
This mismatch runs counter to how humans read. Expert human simplifiers perform nested, sequential operations, balancing structural and lexical variety simultaneously (Alva-Manchego et al.). For developing readers, children, or language learners, addressing only one side of the equation is insufficient; simplifying individual words does nothing if the underlying sentence structure remains beyond their developmental level (Valentini et al.; Habash et al.).
Blind Spots in Metrics and Evaluation
If the gap between structure and vocabulary is so pronounced, why has it persisted? The answer lies in how we evaluate readability and model performance.
Historically, automated readability assessments have relied on surface-level proxies like average sentence length or vocabulary frequency thresholds (Schwarm and Ostendorf; Matthew Shardlow). These metrics fail to capture the underlying syntactic density of a text, meaning they cannot distinguish between a long but grammatically straightforward sentence and a short, structurally dense one (Heilman et al.; Juan Piñeros Liberato).
This blind spot extends to generation metrics. Standard evaluation frameworks rely heavily on token-level edit distance to score model outputs. However, token-matching metrics completely fail to account for macro-structural transformations like sentence splitting or clause reordering (Vásquez-Rodríguez et al.). A model that perfectly splits a long sentence into two fluent, simple sentences might receive a penalizingly low score simply because its word order diverges from a token-matching reference.
Bridging the Gap
To overcome the syntactic simplification gap, systems must move away from treating simplification as a uniform sequence-to-sequence problem. Progress requires architectures that explicitly model text structure before generating words.
This includes frameworks that plan a text's macro-structure or syntactic "sketch" before filling in vocabulary (Junyi Li), pipelines that explicitly classify the required structural operation before generation (Controllable Sentence Simplification via Operation Classification), and evaluation datasets that explicitly track multiple simultaneous transformations (Alva-Manchego et al.). Only by treating grammar and vocabulary as distinct, interacting challenges can automatic text simplification truly bridge the gap to human-level readability.
Key Foundational Work
Reading Level Assessment Using Support Vector Machines and Statistical Language Models (2005) — contributed 1 concept elements to this direction.
Combining Lexical and Grammatical Features to Improve Readability Measures for First and Second Language Texts (2007) — contributed 1 concept elements to this direction.
Appropriate Microplanning Choices for Low-Skilled Readers (2005) — contributed 1 concept elements to this direction.