Margin-space integration of MPE loss via differencing of MMI functionals for generalized error-weighted discriminative training

Erik McDermott, Shinji Watanabe, Atsushi Nakamura

Research output: Contribution to journalArticle

13 Citations (Scopus)

Abstract

Using the central observation that margin-based weighted classification error (modeled using Minimum Phone Error (MPE)) corresponds to the derivative with respect to the margin term of margin-based hinge loss (modeled using Maximum Mutual Information (MMI)), this article subsumes and extends margin-based MPE and MMI within a broader framework in which the objective function is an integral of MPE loss over a range of margin values. Applying the Fundamental Theorem of Calculus, this integral is easily evaluated using finite differences of MMI functionals; lattice-based training using the new criterion can then be carried out using differences of MMI gradients. Experimental results comparing the new framework with margin-based MMI, MCE and MPE on the Corpus of Spontaneous Japanese and the MIT OpenCourseWare/MIT-World corpus are presented.

Original languageEnglish
Pages (from-to)224-227
Number of pages4
JournalUnknown Journal
Publication statusPublished - 2009
Externally publishedYes

Fingerprint

Calculi
functionals
margins
education
integral calculus
Hinges
hinges
Derivatives
theorems
gradients

ASJC Scopus subject areas

  • Human-Computer Interaction
  • Signal Processing
  • Software
  • Sensory Systems

Cite this

Margin-space integration of MPE loss via differencing of MMI functionals for generalized error-weighted discriminative training. / McDermott, Erik; Watanabe, Shinji; Nakamura, Atsushi.

In: Unknown Journal, 2009, p. 224-227.

Research output: Contribution to journalArticle

@article{1676133ed0b64aeb9f88cef1169f97b2,
title = "Margin-space integration of MPE loss via differencing of MMI functionals for generalized error-weighted discriminative training",
abstract = "Using the central observation that margin-based weighted classification error (modeled using Minimum Phone Error (MPE)) corresponds to the derivative with respect to the margin term of margin-based hinge loss (modeled using Maximum Mutual Information (MMI)), this article subsumes and extends margin-based MPE and MMI within a broader framework in which the objective function is an integral of MPE loss over a range of margin values. Applying the Fundamental Theorem of Calculus, this integral is easily evaluated using finite differences of MMI functionals; lattice-based training using the new criterion can then be carried out using differences of MMI gradients. Experimental results comparing the new framework with margin-based MMI, MCE and MPE on the Corpus of Spontaneous Japanese and the MIT OpenCourseWare/MIT-World corpus are presented.",
author = "Erik McDermott and Shinji Watanabe and Atsushi Nakamura",
year = "2009",
language = "English",
pages = "224--227",
journal = "Nuclear Physics A",
issn = "0375-9474",
publisher = "Elsevier",

}

TY - JOUR

T1 - Margin-space integration of MPE loss via differencing of MMI functionals for generalized error-weighted discriminative training

AU - McDermott, Erik

AU - Watanabe, Shinji

AU - Nakamura, Atsushi

PY - 2009

Y1 - 2009

N2 - Using the central observation that margin-based weighted classification error (modeled using Minimum Phone Error (MPE)) corresponds to the derivative with respect to the margin term of margin-based hinge loss (modeled using Maximum Mutual Information (MMI)), this article subsumes and extends margin-based MPE and MMI within a broader framework in which the objective function is an integral of MPE loss over a range of margin values. Applying the Fundamental Theorem of Calculus, this integral is easily evaluated using finite differences of MMI functionals; lattice-based training using the new criterion can then be carried out using differences of MMI gradients. Experimental results comparing the new framework with margin-based MMI, MCE and MPE on the Corpus of Spontaneous Japanese and the MIT OpenCourseWare/MIT-World corpus are presented.

AB - Using the central observation that margin-based weighted classification error (modeled using Minimum Phone Error (MPE)) corresponds to the derivative with respect to the margin term of margin-based hinge loss (modeled using Maximum Mutual Information (MMI)), this article subsumes and extends margin-based MPE and MMI within a broader framework in which the objective function is an integral of MPE loss over a range of margin values. Applying the Fundamental Theorem of Calculus, this integral is easily evaluated using finite differences of MMI functionals; lattice-based training using the new criterion can then be carried out using differences of MMI gradients. Experimental results comparing the new framework with margin-based MMI, MCE and MPE on the Corpus of Spontaneous Japanese and the MIT OpenCourseWare/MIT-World corpus are presented.

UR - http://www.scopus.com/inward/record.url?scp=70450194926&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=70450194926&partnerID=8YFLogxK

M3 - Article

AN - SCOPUS:70450194926

SP - 224

EP - 227

JO - Nuclear Physics A

JF - Nuclear Physics A

SN - 0375-9474

ER -