A Severity-Based Curriculum Learning Strategy for Arabic Medical Text Generation
Ahmed Alansary, Molham Mohamed, Ali Hamdi · 2026
Read A Severity-Based Curriculum Learning Strategy for Arabic Medical Text Generation by Ahmed Alansary, Molham Mohamed, Ali Hamdi (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-06365.txt.
Introduction
Alansary, Mohamed, and Hamdi introduce severity-based curriculum learning for Arabic medical text generation, tackling uniform sample weight. Training moves slowly from less severe to more critical conditions, splitting the dataset into risk stages and showing the model harder cases during fine-tuning. The model first learns basic patterns before tackling cases. Tests use a MAQA subset of Arabic medical questions and replies, tagged Mild, Moderate, and Critical through rules. It always lifts scores across all models. Gains run 4 to 7 percent over bases and 3 to 6 percent over usual tuning.
Catalog record
| CREATOR | Ahmed Alansary, Molham Mohamed, Ali Hamdi |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 23242 |
| LICENSE | http://creativecommons.org/licenses/by-sa/4.0/ |
| RIGHTS | CC BY SA |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.06365
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
A Severity-Based Curriculum Learning Strategy for Arabic Medical Text Generation Ahmed Alansary, Molham Mohamed, Ali Hamdi Abstract Arabic medical text generation is increasingly needed to help users interpret symptoms and access general health guidance in their native language. Nevertheless, many existing methods assume uniform importance across training samples, overlooking differences in clinical severity. This simplification can hinder the model’s ability to properly capture complex or high-risk cases. To overcome this issue, this work introduces a Severity-based Curriculum Learning Strategy for Arabic Medical Text Generation , where the training process is structured to move gradually from less severe to more critical medical conditions. The approach divides the dataset into ordered stages based on severity and incrementally exposes the model to more challenging cases during fine-tuning, allowing it to first learn basic medical patterns before addressing more complex scenarios. The proposed method is evaluated on a subset of the Medical Arabic Question Answering (MAQA) dataset, which includes Arabic medical questions describing symptoms alongside corresponding responses. In addition, the dataset is annotated with three severity levels (Mild, Moderate, and Critical) using a rule-based method developed in this study. The results demonstrate that incorporating severity-aware curriculum learning leads to consistent performance improvements across all tested models, with gains of around +4% to +7% over baseline models and +3% to +6% compared with conventional fine-tuning approaches. † † institute: October University for Modern Sciences & Arts, Egypt 1 Introduction Medical text generation and understanding are increasingly central to modern healthcare, where clinical documents such as electronic medical records (EMRs), radiology reports, and physician notes contain essential information about diagnoses, treatments, and patient outcomes. These records support both clinical decision-making and large-scale medical analysis. However, working with medical text remains challenging due to its unstructured nature, specialized terminology, and wide variation in complexity. Clinical descriptions can range from routine observations to rare and highly complex cases, making automated processing difficult for natural language processing (NLP) systems 15 . Moreover, generating reliable medical text requires models to capture not only frequent clinical patterns but also infrequent yet critical abnormalities, further increasing the complexity of the task 11 . Advances in deep learning and large language models have led to substantial progress in NLP, including medical text generation and understanding. Architectures such as recurrent neural networks, convolutional neural networks, and transformer-based models have been widely adopted in this domain. Through large-scale pretraining and contextual representation learning, these models can capture complex linguistic structures and domain-specific knowledge from extensive corpora 15 . Despite these improvements, most training approaches still rely on uniform sampling, where all training instances are treated equally during optimization. This assumption does not align well with medical data, which are often highly imbalanced. Common conditions dominate clinical datasets, while rare or abnormal cases are underrepresented. Consequently, models tend to produce generic outputs that reflect frequent patterns, while struggling to accurately represent clinically important but infrequent cases 11 . In addition, ignoring differences in sample difficulty may limit the model’s ability to learn more complex medical scenarios effectively 15 . Curriculum learning offers a structured alternative by introducing training samples in a meaningful sequence, typically moving from simpler to more complex examples. Inspired by human learning, this approach allows models to build foundational knowledge before tackling more difficult instances. Prior work has shown that such strategies can improve convergence behaviour and overall performance by guiding optimization more effectively in the early stages of training 5 . In text generation, the order of training data has been shown to influence both learning efficiency and output quality, highlighting the potential of curriculum-based methods for neural generation tasks 5 . In the context of medical NLP, recent studies have begun to explore curriculum learning as a way to better handle the inherent complexity and imbalance of clinical data, particularly for rare and challenging cases 15 . Building on these insights, this work proposes a severity-based curriculum learning strategy for Arabic medical text generation. Instead of treating all samples uniformly, the approach organizes training data according to levels of medical severity and introduces them progressively during fine-tuning. The model is first exposed to simpler cases, allowing it to learn basic medical patterns, before gradually incorporating more complex and severe conditions. This structured training process helps the model better capture the layered nature of medical knowledge and improves its ability to generate accurate and contextually appropriate outputs across a wide range of clinical scenarios. 2 Related Work Natural language processing has become increasingly important in healthcare due to the growing availability of digital medical data such as electronic medical records, clinical notes, and online medical discussions 18 ; 6 ; 16 . These resources enable automated systems to assist in tasks such as diagnosis support, medical question answering, and report generation. However, medical text processing remains challenging because clinical language is complex, domain-specific, and often unstructured 14 . In addition, the biomedical domain contains specialized terminology and limited annotated datasets, which makes it difficult for general language models to perform reliably without domain adaptation 9 . Recent studies have therefore explored adapting large language models to medical applications through fine-tuning, domain-specific corpora, and specialized training strategies. Several works have focused on Arabic medical language processing due to the scarcity of resources and the linguistic complexity of the Arabic language. The rich morphology of Arabic, the diversity of dialects, and the limited availability of annotated datasets create additional challenges for developing reliable medical NLP systems 1 . To address these issues, researchers have proposed various approaches such as dataset construction, ensemble learning, and specialized preprocessing techniques. For example, ensemble transformer models have been applied to Arabic biomedical question classification to improve prediction stability and accuracy 13 . Other studies investigated text preprocessing techniques such as summarization, refinement, and named entity recognition to improve disease classification and symptom severity prediction from Arabic medical texts 12 ; 8 . Additionally, large language models have been fine-tuned on real-world medical conversations to generate medical advice and assist patients in healthcare systems 4 . These approaches demonstrate the potential of transformer-based models in Arabic medical applications. Another research direction focuses on intelligent healthcare systems that combine classification and generation tasks to assist medical decision-making. For example, transformer-based systems have been developed for medical triage and severity prediction using symptom descriptions provided by patients 2 . Open-domain question answering systems have also been proposed to provide reliable health information in Arabic by integrating knowledge extraction and language generation models 3 . In addition, synthetic data generation has been explored to address the scarcity of clinical data in privacy-sensitive domains 10 . These studies highlight the increasing role of generative models in healthcare applications. Curriculum learning has emerged as an effective training strategy for improving neural models by organizing the learning process from easier samples to more difficult ones. Previous research demonstrated that curriculum learning can improve both convergence speed and model performance in data-to-text generation tasks 5 . In the medical domain, curriculum-based frameworks have been applied to complex tasks such as medical report generation, where models progressively learn from simpler training instances before handling more challenging cases 11 . Curriculum learning has also been applied to other generation tasks such as emotional dialogue systems and machine translation