Beyond Readability: Building Accessible Text with People and AI

by

Author: Denis Avetisyan

New research details a practical framework for generating easily understood text that combines the power of automated simplification with crucial human oversight.

The proposed HiTL-HoTL framework establishes a workflow designed to reconcile the theoretical benefits of human-in-the-loop learning with the practical realities of handling unforeseen production edge cases.
The proposed HiTL-HoTL framework establishes a workflow designed to reconcile the theoretical benefits of human-in-the-loop learning with the practical realities of handling unforeseen production edge cases.

This review proposes a Human-in/on-the-Loop framework leveraging accessibility standards and KPIs to optimize text for diverse cognitive needs.

While automated text simplification aims to improve readability, current pipelines often fail to fully account for user comprehension or established accessibility standards. This paper introduces ‘A Human-in/on-the-Loop Framework for Accessible Text Generation’, presenting a hybrid approach that integrates human expertise into both the generation and evaluation of plain language texts. By operationalizing accessibility guidelines and empirical evidence through checklists and automated triggers, the framework enables systematic human oversight and model adaptation. Could this human-centered approach establish a new standard for transparent, auditable, and truly inclusive natural language processing systems?

The Illusion of Accessibility: Standards and Their Discontents

The fundamental principle of effective communication necessitates adapting textual content to accommodate a wide spectrum of cognitive abilities; however, realizing this principle is hindered by a lack of unified standards. While the intention to broaden access to information is widespread, various guidelines and frameworks exist with differing levels of rigor and intended audiences. This fragmented landscape presents challenges for content creators, developers of accessibility tools, and organizations seeking to ensure inclusivity. Consequently, assessing and improving text accessibility becomes a complex undertaking, requiring careful consideration of which standards apply and how they intersect, ultimately impacting the reach and effectiveness of vital information for diverse populations.

Though both Plain Language (PL) and Easy-to-Read (ER) principles strive to make information accessible, they represent distinct approaches with varying degrees of simplification and intended audiences. Plain Language generally focuses on clarity and conciseness for a broad audience, including those with limited literacy or cognitive differences, but doesn’t necessarily reduce complexity to the same extent as Easy-to-Read. ER, conversely, prioritizes extreme simplification-shorter sentences, common vocabulary, and ample white space-specifically targeting individuals with significant cognitive impairments or learning disabilities. This difference in stringency means that a text adhering to PL guidelines might not automatically satisfy ER criteria, and vice versa, necessitating careful consideration of the target demographic when choosing which accessibility standards to apply.

The pursuit of universally accessible text is guided by evolving international standards, with Plain Language (PL) and Easy-to-Read (ER) approaches taking distinct, yet complementary, paths. ISO 24495-1:2023 establishes a comprehensive framework for PL, prioritizing clarity and conciseness through broad principles applicable across various contexts – it focuses on how to write clearly, rather than prescribing strict rules. In contrast, the UNE 153101:2018 standard delivers a more prescriptive methodology for ER texts, outlining specific linguistic criteria – such as sentence length, syllable count, and the use of passive voice – aimed at a narrower audience with demonstrably lower literacy levels. This difference reflects a spectrum of accessibility needs; while PL aims to improve comprehension for the general public, ER provides a targeted solution for individuals requiring significant simplification to fully engage with written information.

The development of truly inclusive textual communication hinges on a granular understanding of accessibility standards. Simply aiming for “easy to read” is insufficient; tools and workflows must be designed to differentiate between varying cognitive loads and cater to specific reader needs. Recognizing the distinctions between frameworks like Plain Language and Easy-to-Read – and leveraging the detailed criteria offered by standards such as ISO 24495-1:2023 and UNE 153101:2018 – allows for the creation of adaptable content. This nuanced approach moves beyond broad simplification towards precise tailoring, enabling the automatic assessment of reading difficulty and the generation of texts optimized for diverse audiences, ultimately fostering greater comprehension and engagement across the spectrum of cognitive abilities.

Iterative Simplification: Chasing a Moving Target

Robust text simplification systems are no longer limited to static, pre-defined rules or a single transformation step; instead, they incorporate adaptation and learning mechanisms to continuously improve performance. This shift from one-time transformations to iterative processes enables the simplification pipeline to refine its techniques based on ongoing evaluation and feedback. By analyzing the outcomes of previous simplifications, the system adjusts its parameters and strategies to better address the nuances of complex language and enhance readability for target audiences. This adaptive capability is essential for addressing the inherent variability in language and tailoring simplifications to specific user needs and cognitive abilities.

Continuous evaluation and feedback are central to refining text simplification techniques. This process involves assessing the output of simplification models against established readability metrics and, critically, gathering user feedback from target audiences. These evaluations identify areas where simplification fails to improve comprehension, such as instances of semantic distortion or the retention of complex syntactic structures. The gathered data is then used to adjust model parameters, refine simplification rules, or retrain the model with augmented datasets. This iterative cycle of evaluation, analysis, and refinement allows simplification systems to progressively improve their performance and achieve demonstrably higher levels of readability and comprehension for diverse user groups.

The iterative refinement of text simplification techniques is substantially enabled by the use of annotated corpora, with the EASIER Corpus serving as a key example. This resource provides a dataset of complex sentences paired with multiple simplified versions, each annotated with information regarding the specific linguistic features targeted for simplification – such as rare words, long sentences, or passive voice constructions. The availability of such data allows for supervised machine learning approaches, enabling models to learn the relationships between complex language and effective simplification strategies. This data-driven learning process is critical for improving the accuracy and relevance of simplification, as demonstrated by its impact on complex word identification recall rates for specific user groups.

Data-driven text simplification models demonstrate improved performance in complex word identification, as quantified by recall rates observed in user studies. Specifically, evaluations using the EASIER Corpus have shown a recall of 0.73 for older adults and 0.56 for participants with intellectual disabilities (ID) when identifying previously complex words within simplified text. These recall rates indicate a statistically significant increase in comprehension for both groups, demonstrating the model’s ability to effectively target and mitigate linguistic challenges presented by complex vocabulary. The data suggests that leveraging annotated corpora allows models to learn patterns of complexity and apply targeted simplification strategies, leading to measurable gains in readability and accessibility.

Modular Workflows: The Illusion of Control

A Multi-Agent Workflow for text simplification decomposes the overall task into discrete, specialized modules, or ‘agents’. This modularity allows for independent development, testing, and optimization of each component, contributing to system scalability. Agents are assigned specific sub-tasks, such as lexical simplification, syntactic simplification, or sentence segmentation, enabling parallel processing and efficient resource allocation. This approach contrasts with monolithic simplification systems and facilitates easier adaptation to diverse text types and user needs by allowing for the addition, removal, or modification of individual agents without impacting the entire workflow. The architecture supports both rule-based and machine learning-based agents, offering flexibility in implementation and allowing for integration of diverse simplification techniques.

The multi-agent workflow decomposes text simplification into discrete operations performed by specialized agents. These agents address specific linguistic challenges, including lexical simplification through synonym replacement, structural simplification via sentence splitting, and grammatical simplification targeting complex syntactic constructions. By isolating these tasks, each agent can be individually optimized for its specific function, leading to improved overall performance compared to monolithic simplification systems. This modularity allows for targeted improvements and facilitates the integration of new simplification techniques as they become available, enhancing both the efficiency and effectiveness of the process.

The multi-agent workflow incorporates an Adaptation & Learning process that utilizes evaluation data to iteratively improve the performance of individual agents. This refinement cycle directly impacts complex word identification, achieving a precision of 0.54 when tested with older adult participants and 0.58 with participants identified as having intellectual disabilities (ID). The system’s ability to learn from evaluation results allows for targeted adjustments to agent behavior, enhancing its effectiveness in tailoring text simplification for these specific user groups and improving comprehension of challenging vocabulary.

Evaluations demonstrate the framework’s effectiveness in synonym substitution, achieving a 67.5% acceptance rate among older adults and 58.9% among participants with intellectual disabilities. This indicates a substantial degree of user agreement with the lexical choices made by the simplification process. Furthermore, inter-annotator agreement for multi-word expression simplification was measured at 0.64 using Kappa statistic, signifying a strong level of consistency in how simplification decisions are made across different human evaluators, and thus, the reliability of the framework’s outputs.

The pursuit of fully automated text simplification, as explored in this framework, often overlooks a crucial point. One imagines elegant algorithms churning out perfectly accessible prose, yet production – real-world usage – inevitably introduces nuance and complexity. As Grace Hopper observed, “It’s easier to ask forgiveness than it is to get permission.” This rings true; striving for absolute pre-emptive accessibility can paralyze progress. The Human-in/on-the-Loop approach acknowledges this reality. It doesn’t promise a flawless, hands-off solution, but rather a pragmatic one-a system that anticipates the need for iterative human refinement, accepting that even the most sophisticated automated processes will require a touch of practical adaptation to truly meet accessibility standards. It’s not about avoiding errors, but about building a system robust enough to handle them.

What’s Next?

The pursuit of accessible text generation, even with the nuanced approach of combined human-in/on-the-loop systems, inevitably bumps against the hard realities of deployment. The framework detailed herein establishes a logical, even elegant, method for simplifying language according to established KPIs – but those KPIs themselves are, at best, imperfect proxies for genuine comprehension. It is all too easy to imagine edge cases, subtle ambiguities, and the sheer diversity of cognitive styles rendering even meticulously crafted plain language inaccessible to some. Every abstraction dies in production, and simplification is an abstraction of meaning.

Future work will undoubtedly focus on refining those KPIs, perhaps leveraging increasingly sophisticated metrics derived from neurocognitive studies. Yet, the deeper challenge remains: defining ‘accessibility’ itself. Is it merely a matter of readability scores, or does true accessibility demand a dynamic, personalized approach – text that adapts not just to general guidelines, but to the specific needs of each individual reader? That path, while theoretically appealing, introduces a complexity that threatens to overwhelm any system, however cleverly designed.

Ultimately, this framework – like all its predecessors – will become a foundation for future technical debt. The goalposts will shift. User expectations will evolve. And the relentless pressure to scale will inevitably lead to compromises. Still, it dies beautifully – a structured attempt to bridge the gap between intention and understanding, before the inevitable entropy sets in. Everything deployable will eventually crash.

Original article: https://arxiv.org/pdf/2603.18879.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-03-22 17:43