Designing a Modular Transformer-Based Framework for Low-Resource Language Text Summarization
Smart Human-centered Emerging Research in Machine Intelligence · 2026
DOI: 10.68012/shem.v2i2.271 ↗Kurz gefasst
Fachartikel in der Zeitschrift Smart Human-centered Emerging Research in Machine Intelligence, 2026. Verfasst von Shofiyul Millah und 3 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Text summarization for low-resource languages remains constrained by limited training corpora and the computational demands of large transformer models. This paper presents a modular transformer-based framework designed to address both constraints through architectural decomposition rather than data expansion. The framework separates the summarization pipeline into three interchangeable modules: a lightweight multilingual encoder pretrained on cross-lingual embeddings, a compression-aware attention mechanism that reduces parameter overhead by sharing weights across decoder layers, and a fine-tuning adapter layer that allows rapid domain adaptation with minimal labeled data. Each module operates independently, allowing researchers to swap components based on the target language's available resources without retraining the entire model. The design prioritizes transfer learning from high-resource languages, using adapter-based fine-tuning to bridge the performance gap that typically affects low-resource settings. We detail the architectural specifications, including layer configurations, attention head allocation, and the adapter injection strategy, alongside a reference implementation in Python using standard deep learning libraries. Rather than reporting benchmark results from empirical data collection, this study focuses on the design rationale, computational efficiency projections, and scalability considerations for languages with sparse digital text resources. The discussion extends to the framework's potential impact: reduced infrastructure barriers for NLP research in underrepresented languages, faster deployment cycles for organizations working in multilingual contexts, and a contribution toward more equitable access to AI-driven language technology. The proposed architecture offers a practical blueprint for researchers and practitioners aiming to build summarization systems where labeled data and computational resources are scarce.
Zitieren
APA
Millah, S., Sean, R. S., Vaher, K., & Julianingsih, D. (2026). Designing a Modular Transformer-Based Framework for Low-Resource Language Text Summarization. Smart Human-centered Emerging Research in Machine Intelligence. https://doi.org/10.68012/shem.v2i2.271
Vancouver
Millah S, Sean RS, Vaher K, Julianingsih D. Designing a Modular Transformer-Based Framework for Low-Resource Language Text Summarization. Smart Human-centered Emerging Research in Machine Intelligence. 2026. doi:10.68012/shem.v2i2.271
BibTeX
@article{Millah2026,
author = {Millah, Shofiyul and Sean, Rio Squire and Vaher, Kristina and Julianingsih, Dwi},
title = {Designing a Modular Transformer-Based Framework for Low-Resource Language Text Summarization},
journal = {Smart Human-centered Emerging Research in Machine Intelligence},
year = {2026},
doi = {10.68012/shem.v2i2.271}
}
EUVIMED