Translation can systematically alter text difficulty, particularly when moving into morphologically rich languages. This study examines whether readability-constrained Large Language Models (LLMs) can mitigate difficulty shifts observed in English–Romanian translation of children’s literature. We construct a paired four-condition corpus comprising English originals, published Romanian translations, readability-constrained LLM translations, and human readability adaptations (12 aligned passages; approx. 23,000 words). Readability is assessed using a Romanian grade-level index (LEMI) designed to be educationally comparable to Flesch–Kincaid Grade Level (FKGL), the cross-linguistic LIX metric, and morphologically informed measures derived from spaCy. Published Romanian translations are significantly more difficult than their English originals, showing higher LIX scores, grade-level estimates, and increased morphological variation. Readability-constrained LLM translation substantially reduces difficulty relative to the published versions (median delta approx. −1.46 grade levels), with significant decreases in LIX, morphological feature density, and lexical diversity (MTLD). Human adaptation yields a smaller reduction (median delta approx. −0.26). Although the direct comparison between LLM and human adaptation is marginal (p = .055, r = 0.64), LLM outputs generally produce larger reductions. These findings demonstrate that translation-induced difficulty shifts are measurable and that controllable LLM translation can modulate readability across structural, lexical, and morphological dimensions in multilingual educational contexts.

Large Language Models deployed for biomedical text simplification frequently produce overgeneration: extraneous content appended beyond the faithful simplification, including leaked model instructions, ungrounded medical claims, and repetitive text. Despite its prevalence, this failure mode remains largely unaddressed. We present a benchmark for document-level overgeneration detection, releasing two resources: SimpleOG-manual, 500 abstract-level examples with human-validated positive labels, and SimpleOG-auto, over 46,000 automatically labeled abstract-level examples derived from submissions to the CLEF 2025 SimpleText Track. Our method exploits the positional regularity of overgeneration in simplification output through sequence alignment, identifying trailing content that lacks a corresponding segment in the source. Human validation of 117 automatically flagged positives confirms ∼95% precision, with leaked model instructions accounting for 75.7% of confirmed cases. Analysis across teams and models reveals that overgeneration is primarily driven by system-level choices, such as prompting and post-processing, rather than by model architecture. We evaluate three detection paradigms and find that sentence similarity (F1 = 0.731, ROC-AUC = 0.915) surprisingly outperforms both NLI-based and LLM-based approaches, suggesting that overgenerated content occupies distinct semantic regions from source material.

This paper presents Conplext 1.0, a multilingual dataset designed for lexical complexity prediction in the context of second language (L2) learning. The resource covers 3,901 sentence contexts for 1,000 vocabulary items across five languages (English, French, Spanish, Swedish, and Dutch), each aligned with Common European Framework of Reference (CEFR) proficiency levels. Contexts were generated using a generative large language model and subsequently filtered for pedagogical suitability. A large-scale best–worst scaling (BWS) annotation experiment is being conducted with L2 learners to derive continuous, learner-informed lexical complexity values. The resulting dataset enables the development of context-aware word difficulty models that account for variation across both languages and learning stages. In addition to its primary use in lexical complexity prediction, Conplext provides valuable opportunities for research in word sense disambiguation, generative model evaluation, and adaptive language learning applications. By integrating computational and educational perspectives, this work advances the study of lexical difficulty in multilingual language learning environments.

Understanding specialized biomedical knowledge can be particularly challenging, posing significant barriers to the acquisition and use of medical information especially by patients. In this study, a methodology for drafting patient-centered explanations of concepts related to the gut-brain axis and related medical conditions is proposed. The explanations are specifically intended for patients affected by neurodegenerative diseases, who are experiencing cognitive decline. The methodology consists of the following steps: 1) the drafting of specialized definitions in the form of intensional definitions, which enable the structured representation of domain-specific knowledge, and 2) the simplification of specialized definitions into patient-centered explanations. In particular, explanations intended for patients are formulated using popular terms and plain language, considered as two complementary strategies aimed at enhancing the comprehension of specialized biomedical knowledge. This work lays the foundation for the future development of a terminology resource specifically designed to collect and systematically represent knowledge related to the gut-brain axis and associated health conditions.

Reliable Text Difficulty Assessment is a prerequisite for valid text simplification workflows and personalized learning applications. However, the development of robust assessment models is severely hindered by a critical bottleneck: the scarcity of expert-annotated corpora containing fine-grained difficulty levels (e.g., CEFR), particularly for lower-resource languages. This paper addresses this data scarcity problem in the context of a low-resource European language. We propose a cross-lingual data augmentation strategy that leverages machine translation to transfer labeled resources from high-resource languages to the target low-resource language. We train BERT-based regression models to predict difficulty scores and investigate whether synthetic, translated data can effectively supplement native training sets. Our experiments demonstrate that augmenting scarce native data with machine-translated corpora significantly improves the accuracy of difficulty estimation, offering a viable solution for languages lacking extensive expert annotations.

In the past few decades, graded readers have been valued within language education and have so much as extended onto the so-called classical (or ‘dead’) languages, such as Latin and Greek. The immersive reading and listening of adapted texts in these languages has been shown to increase students’ proficiency, independence and motivation. However, as of now there is only a small number of related resources as well as of classical languages represented. The present study will investigate the current potential for (semi-)automatic generation of adapted classical-language readers while focusing on the Old Church Slavonic language. From a Natural Language Processing (NLP) point of view, work with the language is challenging due to the variety of dialects and diachronic variations it encompasses. The following steps are taken within our study: 1) Representative measurable characteristics of professional classical-language readers, such as the Latin Lingua latina per se illustrata and the Greek Athenaze, are analysed. 2) Automatic generation of adapted Old Church Slavonic text is attempted through the use of a sequence-to-sequence model (mT5) as well as a Large Language Model (GPT-5) in a one-shot setting. 3) The derived texts’ quality is assessed through both human evaluation and a comparison of their textual characteristics with those of professional texts as defined in point 1). The edited versions of the GPT-based texts are shared for future reference and use.

Text difficulty prediction in educational contexts requires models that balance predictive performance, interpretability, calibration, and pedagogical alignment. While transformer-based approaches increasingly dominate text difficulty classification, educational applications demand transparent and linguistically grounded modeling. This paper presents work aimed at developing a workbench for CEFR-based text difficulty prediction. The proposed platform comprises three main components: (i) a tool for CEFR-aligned dataset preparation incorporating a pipeline for documenting, processing, and enriching textual data, (ii) CEFR-aligned datasets, and (iii) three alternative modeling approaches, namely a rule-based baseline, a feature-based Machine Learning (ML) classifier, and a fine-tuned BERT model. Our approach integrates linguistically informed feature engineering with data-driven modeling techniques, thereby balancing transparency and predictive performance. The proposed workbench has been designed as a language-agnostic infrastructure that can be extended to any language. In its current implementation, it has been applied to the creation of a German CEFR dataset, while its Greek counterpart is currently under development.

This study examines the integration of a bilingual Italian–Spanish concept-oriented terminological resource into a controlled large language model (LLM) translation workflow within the domain of Campanian gastronomy. The termbase encodes structured conceptual, linguistic, and translational metadata, including grammatical information, translation strategies, and genre-sensitive usage recommendations. Through a local Model Context Protocol (MCP) architecture, the resource is dynamically connected to locally deployed LLMs, enabling the automatic identification and retrieval of relevant terminological units prior to generation. The system combines in-context terminological injection with deterministic post-processing enforcement: genre-specific policies are injected into the model prompt prior to generation and verified through a rule-based post-processing layer that enforces surface-level terminological consistency in the output. Two open-weight models — Mistral 7B Instruct and Gemma3 4B — are evaluated across three conditions and three discursive genres on a dataset of authentic texts. The findings suggest that the combination of terminological injection and deterministic enforcement can improve terminological compliance in controlled, domain-specific settings, while also highlighting differences in instruction-following behavior across models and genres.

Text simplification requires reliable automatic evaluation, yet existing learnable metrics such as LENS and LENS-SALSA are specialized and costly to develop. Moreover, it remains unclear how these metrics compare to using large language models (LLMs) as evaluators. Exploring this question is important because LLM-based evaluation could make simplification research and deployment more flexible and easier to adapt than training new task-specific metrics for each setting. In this work, we empirically compare several small, open-weight instruction-tuned LLMs with LENS and LENS-SALSA in both reference-based and reference-free evaluation settings. We measure their alignment with human judgments across multiple datasets. Our results provide insight into when small LLMs can serve as effective evaluators and when specialized metrics remain preferable, informing the design of future evaluation pipelines for text simplification and related text generation tasks.