Linear-time Minimum Bayes Risk Decoding with Reference Aggregation

2024/02/06 by Jannis Vamvas, Rico Sennrich, Vamvas, Jannis +1 · 4 citations
Decision Sciences · Engineering · Computer Science · #Advanced Statistical Process Monitoring #Fault Detection and Control Systems #Bayesian Modeling and Causal Inference

paper · pdf · doi:10.48550/arxiv.2402.04251

Abstract

Minimum Bayes Risk (MBR) decoding is a text generation technique that has been shown to improve the quality of machine translations, but is expensive, even if a sampling-based approximation is used. Besides requiring a large number of sampled sequences, it requires the pairwise calculation of a utility metric, which has quadratic complexity. In this paper, we propose to approximate pairwise metric scores with scores calculated against aggregated reference representations. This changes the complexity of utility estimation from O(n²) to O(n), while empirically preserving most of the quality gains of MBR decoding. We release our source code at https://github.com/ZurichNLP/mbr

Cited by

• Agreement-Constrained Probabilistic Minimum Bayes Risk Decoding
• Structure-Conditional Minimum Bayes Risk Decoding
• Case-Based Decision-Theoretic Decoding with Quality Memories
• Multi-Hypothesis Distillation of Multilingual Neural Translation Models for Low-Resource Languages

Related

• Agreement-Constrained Probabilistic Minimum Bayes Risk Decoding
• Theoretical Guarantees for Minimum Bayes Risk Decoding
• Faster Minimum Bayes Risk Decoding with Confidence-based Pruning
• Efficient Minimum Bayes Risk Decoding using Low-Rank Matrix Completion Algorithms
• Mitigating Metric Bias in Minimum Bayes Risk Decoding