Agreement-Constrained Efficient Probabilistic Minimum Bayes Risk Decoding with Knowledge Distillation Metrics

Koki Natsumi; Deguchi Hiroyuki; Yusuke Sakai; Hidetaka Kamigaito; Taro Watanabe

doi:10.5715/jnlp.33.570

January 1, 2026Journal of Natural Language ProcessingOpen Access

Agreement-Constrained Efficient Probabilistic Minimum Bayes Risk Decoding with Knowledge Distillation Metrics

View Full Paper

Ask AI

Bookmark

Share

Key Points

The aim is to improve translation quality while reducing computation costs in probabilistic minimum Bayes risk decoding.
Introduced agreement-constrained PMBR decoding to enhance score matrix completion using knowledge distillation.
Evaluated the performance in the WMT’23 English-German translation task.
Compared the computational cost and translation quality to traditional probabilistic MBR decoding.
Achieved up to three times improvement in approximation error in matrix completion.
Maintained translation quality comparable to PMBR while matching its computational cost.

Abstract

最小ベイズリスク (Minimum Bayes Risk; MBR) 復号は，出力候補の期待効用を最大化することで高品質な翻訳を生成する手法であるが，候補集合内の全組み合わせに対してスコア（効用）を算出する必要があるため，候補数に対して二乗の計算時間を要する．この計算コストを削減するため，確率的最小ベイズリスク (Probabilistic MBR; PMBR)

Discussion

No takes yet. Share an insight, caveat, or question.

Cite This Study

Natsumi et al. (2026) studied this question.

synapsesocial.com/papers/6a3238e2d50b63ecad2049d3 https://doi.org/10.5715/jnlp.33.570

View Full Paper

Ask AI

Bookmark

Share

Agreement-Constrained Efficient Probabilistic Minimum Bayes Risk Decoding with Knowledge Distillation Metrics

Key Points

Abstract

Discussion

Cite This Study

Agreement-Constrained Efficient Probabilistic Minimum Bayes Risk Decoding with Knowledge Distillation Metrics

Key Points

Abstract

Discussion

Cite This Study

Keep reading

More related papers

More from Journal of Natural Language Processing

Keep reading

More related papers

More from Journal of Natural Language Processing