Minimum Bayes Risk Decoding for Error Span Detection in Reference-Free Automatic Machine Translation Evaluation

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Lyu, Boxuan, Song, Haiyue, Kamigaito, Hidetaka, Ding, Chenchen, Tanaka, Hideki, Utiyama, Masao, Funakoshi, Kotaro, Okumura, Manabu
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866908739783622656
author Lyu, Boxuan
Song, Haiyue
Kamigaito, Hidetaka
Ding, Chenchen
Tanaka, Hideki
Utiyama, Masao
Funakoshi, Kotaro
Okumura, Manabu
author_facet Lyu, Boxuan
Song, Haiyue
Kamigaito, Hidetaka
Ding, Chenchen
Tanaka, Hideki
Utiyama, Masao
Funakoshi, Kotaro
Okumura, Manabu
contents Error Span Detection (ESD) extends automatic machine translation (MT) evaluation by localizing translation errors and labeling their severity. Current generative ESD methods typically use Maximum a Posteriori (MAP) decoding, assuming that the model-estimated probabilities are perfectly correlated with similarity to the human annotation, but we often observe higher likelihood assigned to an incorrect annotation than to the human one. We instead apply Minimum Bayes Risk (MBR) decoding to generative ESD. We use a sentence- or span-level similarity function for MBR decoding, which selects candidate hypotheses based on their approximate similarity to the human annotation. Experimental results on the WMT24 Metrics Shared Task show that MBR decoding significantly improves span-level performance and generally matches or outperforms MAP at the system and sentence levels. To reduce the computational cost of MBR decoding, we further distill its decisions into a model decoded via greedy search, removing the inference-time latency bottleneck.
format Preprint
id arxiv_https___arxiv_org_abs_2512_07540
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Minimum Bayes Risk Decoding for Error Span Detection in Reference-Free Automatic Machine Translation Evaluation
Lyu, Boxuan
Song, Haiyue
Kamigaito, Hidetaka
Ding, Chenchen
Tanaka, Hideki
Utiyama, Masao
Funakoshi, Kotaro
Okumura, Manabu
Computation and Language
Artificial Intelligence
Machine Learning
Error Span Detection (ESD) extends automatic machine translation (MT) evaluation by localizing translation errors and labeling their severity. Current generative ESD methods typically use Maximum a Posteriori (MAP) decoding, assuming that the model-estimated probabilities are perfectly correlated with similarity to the human annotation, but we often observe higher likelihood assigned to an incorrect annotation than to the human one. We instead apply Minimum Bayes Risk (MBR) decoding to generative ESD. We use a sentence- or span-level similarity function for MBR decoding, which selects candidate hypotheses based on their approximate similarity to the human annotation. Experimental results on the WMT24 Metrics Shared Task show that MBR decoding significantly improves span-level performance and generally matches or outperforms MAP at the system and sentence levels. To reduce the computational cost of MBR decoding, we further distill its decisions into a model decoded via greedy search, removing the inference-time latency bottleneck.
title Minimum Bayes Risk Decoding for Error Span Detection in Reference-Free Automatic Machine Translation Evaluation
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2512.07540