The implementation of the Minimum Bayes Risk decoding method described in "Minimum Bayes Risk decoding and system combination based on a recursion for edit distance", Haihua Xu, Daniel Povey, Lidia Mangu and Jie Zhu, Computer Speech and Language, 2011 This is a slightly more principled way to do Minimum Bayes Risk (MBR) decoding than the standard "Confusion Network" method. More...

#include <sausages.h>

Collaboration diagram for MinimumBayesRiskOptions:

[legend]

Public Member Functions
	MinimumBayesRiskOptions ()

void	Register (OptionsItf *opts)

Public Attributes
bool	decode_mbr
	Boolean configuration parameter: if true, we actually update the hypothesis to do MBR decoding (if false, our output is the MAP decoded output, but we output the stats too (i.e. More...

bool	print_silence
	Boolean configuration parameter: if true, the 1-best path will 'keep' the <eps> bins,. More...

Detailed Description

The implementation of the Minimum Bayes Risk decoding method described in "Minimum Bayes Risk decoding and system combination based on a recursion for edit distance", Haihua Xu, Daniel Povey, Lidia Mangu and Jie Zhu, Computer Speech and Language, 2011 This is a slightly more principled way to do Minimum Bayes Risk (MBR) decoding than the standard "Confusion Network" method.

Note: MBR decoding aims to minimize the expected word error rate, assuming the lattice encodes the true uncertainty about what was spoken; standard Viterbi decoding gives the most likely utterance, which corresponds to minimizing the expected sentence error rate.

In addition to giving the MBR output, we also provide a way to get a "Confusion Network" or informally "sausage"-like structure. This is a linear sequence of bins, and in each bin, there is a distribution over words (or epsilon, meaning no word). This is useful for estimating confidence. Note: due to the way these sausages are made, typically there will be, between each bin representing a high-confidence word, a bin in which epsilon (no word) is the most likely word. Inside these bins is where we put possible insertions.

Definition at line 56 of file sausages.h.

Constructor & Destructor Documentation

◆ MinimumBayesRiskOptions()

MinimumBayesRiskOptions ( )

inline

Definition at line 64 of file sausages.h.

64 : decode_mbr(true), print_silence(false)

65 { }

kaldi::MinimumBayesRiskOptions::print_silence

bool print_silence

Boolean configuration parameter: if true, the 1-best path will 'keep' the <eps> bins,.

Definition: sausages.h:62

kaldi::MinimumBayesRiskOptions::decode_mbr

bool decode_mbr

Boolean configuration parameter: if true, we actually update the hypothesis to do MBR decoding (if fa...

Definition: sausages.h:60

Member Function Documentation

◆ Register()

void Register ( OptionsItf * opts )

inline

Definition at line 66 of file sausages.h.

References OptionsItf::Register().

Referenced by main().

                                   {
     opts->Register("decode-mbr", &decode_mbr, "If true, do Minimum Bayes Risk "
                    "decoding (else, Maximum a Posteriori)");
     opts->Register("print-silence", &print_silence, "Keep the inter-word '<eps>' "
                    "bins in the 1-best output (ctm, <eps> can be a 'silence' or a 'deleted' word)");
   }

Member Data Documentation

◆ decode_mbr

bool decode_mbr

Boolean configuration parameter: if true, we actually update the hypothesis to do MBR decoding (if false, our output is the MAP decoded output, but we output the stats too (i.e.

the confidences)).

Definition at line 60 of file sausages.h.

Referenced by MinimumBayesRisk::MbrDecode().