mbrs: A library for minimum Bayes risk decoding

Hiroyuki Deguchi, Yusuke Sakai, Hidetaka Kamigaito, Taro Watanebe
Nara Institute of Science and Technology (NAIST)
Paper Code Docs arXiv Demo
mbrs

Abstract

Minimum Bayes risk (MBR) decoding is a decision rule of generation tasks that outperforms conventional maximum a posterior (MAP) decoding using beam search by selecting high-quality outputs rather than high-probability. Typically, it finds the most probable hypothesis from the set of hypotheses under the sampled pseudo-references. mbrs is a library of MBR decoding, which can flexibly combine various metrics, expectation estimations, and algorithmic variants. It is designed with a focus on scope-based speed measurement and calling count, transparency, reproducibility, and extensibility, which are essential for researchers and developers. We published our mbrs as an MIT-licensed open-source project, and the code is available on GitHub.

BibTeX

@misc{deguchi-2024-mbrs,
   title={mbrs: A Library for Minimum Bayes Risk Decoding},
   author={Hiroyuki Deguchi and Yusuke Sakai and Hidetaka Kamigaito and Taro Watanabe},
   year={2024},
   eprint={2408.04167},
   archivePrefix={arXiv},
   primaryClass={cs.CL},
   url={https://arxiv.org/abs/2408.04167},
}