Toward Explainable Cross-Lingual Adaptive NAS for Enhanced Tamil Medical Text Summarization.

Abstract

In the era of digital health, the demand for effective summarization of medical texts in low-resource languages, such as Tamil, is rapidly increasing. Traditional neural models often struggle with this task due to the scarcity of annotated data and the complexity of medical terminology. To address these challenges, we propose a novel Cross-Lingual Adaptive Neural Architecture Search (CLANAS) framework, specifically designed to enhance Tamil medical text summarization by leveraging cross-lingual transfer learning. The CLANAS framework integrates embedding alignment techniques with neural architecture search (NAS) to automatically design optimal models tailored to the target language. Our methodology involves pre-training on large-scale English medical datasets, including iCliniq, HealthCare Magic, MeQSum, MEDIQA, and BioLaySumm-2023, followed by fine-tuning on Tamil medical texts, ensuring semantic consistency across languages through embedding alignment. The framework was rigorously evaluated on these benchmark datasets, where CLANAS demonstrated significant performance improvements, achieving up to 9.3% enhancement in ROUGE-1 scores, 8.4% in BLEU, and 7.5% in Metric for Evaluation of Translation with Explicit ORdering (METEOR) compared to state-of-the-art models. Ablation studies further confirmed the effectiveness of each component within the framework. These results underscore the potential of CLANAS as a robust solution for improving the quality of medical text summarization in Tamil.