Reproduce “When to Use Extra Context: Evidence-Grounded Terminology Adaptation for Simultaneous Speech Translation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zeyu Yang, Satoshi Nakamura
Researchers at The Chinese University of Hong Kong, Shenzhen, developed Evidence-Grounded Terminology Adaptation (EGTA), a lightweight, inference-time framework for simultaneous speech translation (SimulST). EGTA selectively applies document-level terminology to technical talks, improving terminology fidelity (e.g., +0.212 NER_clean recall for En Zh on MCIF-dev) and overall translation quality while maintaining or reducing latency.
Key insights
- The most substantial benefits from external document context in SimulST for technical talks arise from the precise recovery of paper-specific terminology, rather than general semantic enhancement.
- Selective, evidence-grounded activation of terminology is crucial for effective context utilization, preventing over-biasing and hallucination, as demonstrated by diagnostic stress tests and controls.
- A lightweight, inference-time framework combining ASR/speech-side and decoder-side adaptation for specific terms significantly improves terminology fidelity and overall quality in real-time.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.