Supported decoders

Supported decoders are listed below.

Note

All decoders classes can be imported from mbrs.decoders

Decoding strategies

List of supported decoding strategies

Strategy	CLI name --decoder=<name>	Class	Supported metrics	Reference
MBR decoding	mbr	DecoderMBR	all	(Eikema and Aziz, 2020; Eikema and Aziz, 2022)
N-best reranking	rerank	DecoderRerank	COMETkiwi	–

Expectation estimations

List of supported expectation estimations

Estimation	CLI option	Reference
Monte Carlo	–	(Eikema and Aziz, 2020; Eikema and Aziz, 2022)
Model-based	--reference_lprobs=lprobs.txt	(Jinnai et al., 2024)

Efficient MBR decoders

List of supported efficient MBR decoders

Decoder	CLI name --decoder=<name>	Class	Supported metrics	Reference
Confidence-based pruning	pruning_mbr	DecoderPruningMBR	all	(Cheng and Vlachos, 2023)
Reference aggregation	aggregate_mbr	DecoderAggregateMBR	BLEU: N-gram and length aggregation (DeNero et al., 2009) chrF: N-gram aggregation (Vamvas and Sennrich, 2024) COMET: Embedding aggregation (Vamvas and Sennrich, 2024; Deguchi et al., 2024)	(DeNero et al., 2009; Vamvas and Sennrich, 2024)
Centroid-based aggregation	centroid_mbr	DecoderCentroidMBR	COMET	(Deguchi et al., 2024)
Low-Rank Matrix Completion (Probabilistic MBR)	probabilistic_mbr	DecoderProbabilisticMBR	all	(Trabelsi et al., 2024)