Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Hyperparameter-Free Approach for Faster Minimum Bayes Risk Decoding

Yuu Jinnai, Kaito Ariu

TL;DR

This work addresses the latency challenge of Minimum Bayes-Risk decoding by introducing Adaptive Minimum Bayes-Risk (AMBR), a hyperparameter-free method that reframes MBR as a medoid identification problem and solves it with Correlated Sequential Halving under a user-specified budget. AMBR eliminates dev-set hyperparameter tuning and directly enforces budget constraints, achieving performance on par with Oracle Confidence-Based Pruning and delivering 4–8× speed-ups across machine translation, text summarization, and image captioning with modest quality loss. A theoretical guarantee is established for a related AMBR-Replace variant, lending formal support to the approach, while empirical results demonstrate robust scalability with the number of samples given sufficient budget. Overall, AMBR offers a practical, broadly applicable acceleration for MBR decoding in latency-sensitive settings, reducing the need for extensive hyperparameter tuning while maintaining high-quality outputs.

Abstract

Minimum Bayes-Risk (MBR) decoding is shown to be a powerful alternative to beam search decoding for a wide range of text generation tasks. However, MBR requires a huge amount of time for inference to compute the MBR objective, which makes the method infeasible in many situations where response time is critical. Confidence-based pruning (CBP) (Cheng and Vlachos, 2023) has recently been proposed to reduce the inference time in machine translation tasks. Although it is shown to significantly reduce the amount of computation, it requires hyperparameter tuning using a development set to be effective. To this end, we propose Approximate Minimum Bayes-Risk (AMBR) decoding, a hyperparameter-free method to run MBR decoding approximately. AMBR is derived from the observation that the problem of computing the sample-based MBR objective is the medoid identification problem. AMBR uses the Correlated Sequential Halving (CSH) algorithm (Baharav and Tse, 2019), the best approximation algorithm to date for the medoid identification problem, to compute the sample-based MBR objective. We evaluate AMBR on machine translation, text summarization, and image captioning tasks. The results show that AMBR achieves on par with CBP, with CBP selecting hyperparameters through an Oracle for each given computation budget.

Hyperparameter-Free Approach for Faster Minimum Bayes Risk Decoding

TL;DR

This work addresses the latency challenge of Minimum Bayes-Risk decoding by introducing Adaptive Minimum Bayes-Risk (AMBR), a hyperparameter-free method that reframes MBR as a medoid identification problem and solves it with Correlated Sequential Halving under a user-specified budget. AMBR eliminates dev-set hyperparameter tuning and directly enforces budget constraints, achieving performance on par with Oracle Confidence-Based Pruning and delivering 4–8× speed-ups across machine translation, text summarization, and image captioning with modest quality loss. A theoretical guarantee is established for a related AMBR-Replace variant, lending formal support to the approach, while empirical results demonstrate robust scalability with the number of samples given sufficient budget. Overall, AMBR offers a practical, broadly applicable acceleration for MBR decoding in latency-sensitive settings, reducing the need for extensive hyperparameter tuning while maintaining high-quality outputs.

Abstract

Minimum Bayes-Risk (MBR) decoding is shown to be a powerful alternative to beam search decoding for a wide range of text generation tasks. However, MBR requires a huge amount of time for inference to compute the MBR objective, which makes the method infeasible in many situations where response time is critical. Confidence-based pruning (CBP) (Cheng and Vlachos, 2023) has recently been proposed to reduce the inference time in machine translation tasks. Although it is shown to significantly reduce the amount of computation, it requires hyperparameter tuning using a development set to be effective. To this end, we propose Approximate Minimum Bayes-Risk (AMBR) decoding, a hyperparameter-free method to run MBR decoding approximately. AMBR is derived from the observation that the problem of computing the sample-based MBR objective is the medoid identification problem. AMBR uses the Correlated Sequential Halving (CSH) algorithm (Baharav and Tse, 2019), the best approximation algorithm to date for the medoid identification problem, to compute the sample-based MBR objective. We evaluate AMBR on machine translation, text summarization, and image captioning tasks. The results show that AMBR achieves on par with CBP, with CBP selecting hyperparameters through an Oracle for each given computation budget.
Paper Structure (27 sections, 1 theorem, 13 equations, 8 figures, 6 tables)

This paper contains 27 sections, 1 theorem, 13 equations, 8 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Minimum Bayes-Risk (MBR) Decoding
  4. Computational Complexity of MBR Decoding
  5. Confidence-Based Pruning (CBP)
  6. Adaptive Minimum Bayes Risk (AMBR) Decoding
  7. Analytical Result
  8. Experiments
  9. Machine Translation
  10. AMBR is on par with Oracle CBP.
  11. AMBR scales with the number of samples given enough budget.
  12. Text Summarization
  13. C2F can surpass the score of standard MBR under conditions
  14. Image Captioning
  15. Related Work
  16. ...and 12 more sections

Key Result

Lemma 1

Assuming $T \geq N \log N$, AMBR replacing Line 7 with Eq. eq:guaranteed (AMBR-Replace) correctly identifies $\mathbf{h}^{\mathrm{MC}}$ with probability at least $1 - \log N \exp(- \frac{T}{\log N} C)$ where $C$ is an instance dependent variable determined by $u$ and $\mathcal{H}$.

Figures (8)

  • Figure 1: COMET-20 score on WMT'21 De-En and Ru-En using the WMT 21 X-En model. The shaded regions show the minimum and the maximum values over five runs. The horizontal axis shows the reduction in the number of evaluations compared to the standard MBR with all samples.
  • Figure 2: Evaluation of AMBR with a varying number of samples with a fixed evaluation budget on WMT'21 De-En with COMET-20 score using the M2M100 418M model. The shaded regions show the minimum and the maximum values over five runs.
  • Figure 3: (\ref{['fig:samsum']}) InfoLM score, (\ref{['fig:samsum-rouge']}) ROUGE-L score, and (\ref{['fig:samsum-err']}) error rate on SAMSum dataset. (\ref{['fig:xsum']}) InfoLM score, (\ref{['fig:xsum-rouge']}) ROUGE-L score, and (\ref{['fig:xsum-err']}) error rate on XSum dataset. (\ref{['fig:mscoco']}) RefCLIPScore, (\ref{['fig:mscoco-bleu']}) BLEU score, and (\ref{['fig:mscoco-err']}) error rate on MS COCO dataset. The shaded regions show the minimum and the maximum values over five runs. The error rate is defined as the ratio of selecting a hypothesis different from the one selected by standard MBR using all the samples ($N=64$).
  • Figure 4: Average reward according to OASST (gold reference reward) on Alpaca Eval dataset.
  • Figure 5: The error rate on WMT'21 De-En and Ru-En using the WMT 21 X-En model. The shaded regions show the minimum and the maximum values over five runs. The error rate is the ratio of selecting a hypothesis different from the standard MBR using all samples. The horizontal axis shows the reduction in the number of evaluations compared to the standard MBR with all samples.
  • ...and 3 more figures

Theorems & Definitions (1)

  • Lemma 1