Tools

LEADER 04054nam a2200541 i 4500

001 on1609158634;

003 OCoLC;

005 20260729122319.0;

006 m o d ;

007 cr unu||||||||;

008 260729s2026 ncu obm 000 0 eng d;

040 a| EREb| enge| rdac| ERE;

035 a| (OCoLC)1609158634;

049 a| EREE;

100 1 a| Sarambo, Stephanie Elizabethe| author;

245 10 a| Uncertainty-Aware Adaptation for Simultaneous Machine Translation under Domain Shift;

264 1 a| [Greenville, N.C.] :b| [East Carolina Universityc| 2026.];

300 a| 66 pages;

336 a| textb| txt2| rdacontent;

337 a| computerb| c2| rdamedia;

338 a| online resourceb| cr2| rdacarrier;

347 a| text fileb| application/pdfc| 711.51 KB2| rda;

538 a| System requirements: Adobe Reader.;

538 a| Mode of access: World Wide Web.;

502 c| East Carolina Universityd| 2026.;

504 a| Includes bibliographical references.;

520 3 a| Simultaneous Machine Translation (SiMT) uses advancements in Natural Language Processing (NLP) to provide seamless, real-time communication between languages, while balancing translation quality and latency. However, these systems are highly sensitive to domain shift, where discrepancies between training and deployment data can significantly degrade performance. This paper proposes an uncertainty-aware SiMT framework that detects and adapts to domain shift in real-time. The system uses a neural machine translation model called StreamSpeech integrated with an uncertainty module that computes token-level entropy from the model distribution, then uses an adaptation policy to improve translation quality. The model is trained and evaluated on general-domain speech datasets commonly used in large-scale speech translation research, and its robustness is assessed using an out-of-domain dataset to induce controlled domain shift. The datasets used in this paper are the general-domain datasets CoVoST2 and Microsoft Speech Language Translation (MSLT), and the medical domain dataset Multilingual Medical Speech Translation (MultiMed-ST). Uncertainty behavior is analyzed by logging entropy values during streaming translation and examining the correlation between uncertainty spikes and translation errors. System performance is analyzed using BLEU scores for translation quality, and Average Lagging (AL) for latency.Finally, the proposed framework introduces an uncertainty-triggered adaptive waiting strategy that dynamically increases translation delay only when uncertainty falls within a predefined threshold. Experimentation results suggest that not only does uncertainty-aware adaptive waiting improve translation quality under domain-shift, but it also improves translation quality in the general-domain, while maintaining low latency, offering a practical approach to enhancing robustness in simultaneous translation systems.;

653 a| Simultaneous Machine Translation, Domain Shift, Uncertainty Estimation, Adaptive Decoding, Speech Translation, Entropy-Based Adaptation, Real-Time Translation, Neural Machine Translation,;

653 a| Simultaneous Machine Translation;

653 a| Domain Shift;

653 a| Uncertainty Estimation;

653 a| Adaptive Decoding;

653 a| Uncertainty-Aware Adaptation;

653 a| Adaptive Waiting;

653 a| Entropy Estimation;

653 a| Streaming Speech Translation;

653 a| Neural Machine Translation;

653 a| Low-Latency Translation;

700 1 a| Herndon, Nice| degree supervisor.;

710 2 a| East Carolina University;

856 40 z| Access via ScholarShipu| http://hdl.handle.net/10342/14727;

949 a| Click on web addressw| ASISh| JOYNER101;

949 a| Click on web addressw| ASISh| HSL111;

994 a| C0b| ERE;

596 a| 1 4;

998 a| 6795478;

999 a| CLICK ON WEB ADDRESSw| ASISc| 1i| 6795478-1001l| JNETm| JOYNERr| Ys| Yt| JNE3ETDu| 7/29/2026x| ETDz| JERESOURCE;

999 a| CLICK ON WEB ADDRESSw| ASISc| 1i| 6795478-2001l| HSLELECm| HSLr| Ys| Yt| HEETDu| 7/29/2026x| ETDz| HERESOURCE;

Uncertainty-Aware Adaptation for Simultaneous Machine Translation under Domain Shift

Author/creator	Sarambo, Stephanie Elizabeth author
Other author	Herndon, Nic degree supervisor.
Other author	East Carolina University
Format	Theses and dissertations
Publication	[Greenville, N.C.] : [East Carolina University 2026.]
Description	66 pages
Supplemental Content	Access via ScholarShip

Summary	Simultaneous Machine Translation (SiMT) uses advancements in Natural Language Processing (NLP) to provide seamless, real-time communication between languages, while balancing translation quality and latency. However, these systems are highly sensitive to domain shift, where discrepancies between training and deployment data can significantly degrade performance. This paper proposes an uncertainty-aware SiMT framework that detects and adapts to domain shift in real-time. The system uses a neural machine translation model called StreamSpeech integrated with an uncertainty module that computes token-level entropy from the model distribution, then uses an adaptation policy to improve translation quality. The model is trained and evaluated on general-domain speech datasets commonly used in large-scale speech translation research, and its robustness is assessed using an out-of-domain dataset to induce controlled domain shift. The datasets used in this paper are the general-domain datasets CoVoST2 and Microsoft Speech Language Translation (MSLT), and the medical domain dataset Multilingual Medical Speech Translation (MultiMed-ST). Uncertainty behavior is analyzed by logging entropy values during streaming translation and examining the correlation between uncertainty spikes and translation errors. System performance is analyzed using BLEU scores for translation quality, and Average Lagging (AL) for latency.Finally, the proposed framework introduces an uncertainty-triggered adaptive waiting strategy that dynamically increases translation delay only when uncertainty falls within a predefined threshold. Experimentation results suggest that not only does uncertainty-aware adaptive waiting improve translation quality under domain-shift, but it also improves translation quality in the general-domain, while maintaining low latency, offering a practical approach to enhancing robustness in simultaneous translation systems.
Dissertation note	East Carolina University 2026.
Bibliography note	Includes bibliographical references.
Technical details	System requirements: Adobe Reader.
Technical details	Mode of access: World Wide Web.

Availability

Library	Location	Call Number	Status	Item Actions
	Electronic Resources	Access Content Online	✔ Available

Uncertainty-Aware Adaptation for Simultaneous Machine Translation under Domain Shift

Click here for more information about this title

Availability