Year: 2026
Type: conference-paper
Abstract: When reading foreign-language literature, nonnative users often face significant challenges.Existing machine translation systems often obscure or mistranslate key terminology, while lay-oriented paraphrasing tends to oversimplify it, hindering readers from acquiring domainspecific technical vocabulary.To address this gap, we define a new task, Glossing-Oriented Academic Translation (GOAT), which aims to produce translations adapted to a reader's academic level.We then propose GlossaGen, a comprehensive framework to address this task.GlossaGen combines a multi-agent data synthesis pipeline that generates large-scale, level-specific training data with a training strategy based on dynamic adapter merging, which balances task-level generalization and reader-level specialization through a "generalist" adapter and a fine-grained "expert" one.We evaluate GlossaGen on a synthesized benchmark using automatic metrics and large language model (LLM)-based assessments at both reader levels, together with a human evaluation study on the undergraduate setting.Across these evaluations, our approach outperforms strong baselines on most metrics.Overall, GlossaGen provides a practical step toward making scientific literature more accessible to non-native readers through more accurate translations and pedagogically appropriate, levelspecific term explanations.We release our code and data to facilitate further research:GlossaGen.
Institutions: Renmin University of China, McGill University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Quality education
Open Access status: hybrid
OpenAlex ID: W7166830121