Supported MMI FOR MODEL AND FEATURE-SPACE DISCRIMINATIVE TRAINING

0
0
1744 days ago, 568 views
PowerPoint PPT Presentation
Layout. MMI reviewBMMIimprovements to Model-Space updatesExperiments Resultsconclusion. MMI survey. Supported MMI. Uphold a delicate edge that is corresponding to the quantity of mistakes in a guessed sentenceBoost the probability of the sentences that have more blunders in this way creating more confusable datab : boosting variable : precision of a sentence s given the reference Sr

Presentation Transcript

Slide 1

﻿Helped MMI FOR MODEL AND FEATURE-SPACE DISCRIMINATIVE TRAINING Daniel Povey, Dimitri Kanevsky, Brian Kingsbury, Bhuvana Ramabhadran, George Saon and Karthik Visweswariah IBM T.J. Watson Research Center Yorktown Heights, NY, USA Presented by Yueng-Tien ,Lo

Slide 2

Outline MMI survey BMMI enhancements to Model-Space redesigns Experiments Results conclusion

Slide 3

MMI audit

Slide 4

BOOSTED MMI Enforce a delicate edge that is corresponding to the quantity of blunders in a speculated sentence Boost the probability of the sentences that have more mistakes hence creating more confusable information b : boosting component : exactness of a sentence s given the reference Sr

Slide 5

IMPROVEMENTS TO MODEL-SPACE UPDATES(1/2) Canceled measurements scratch off the insights aggregated on each edge from the numerator and denominator The main impact this crossing out has is to change the Gaussian-particular learning-rate consistent

Slide 6

IMPROVEMENTS TO MODEL-SPACE UPDATES(2/2) I-smoothing to past cycle back off to the MMI evaluate (in light of one emphasis of EBW beginning from the ebb and flow emphasis' insights). Rules for getting to be the biggest of: ( I ) the ordinary case: ( I )twofold the littlest Lead to being certain definite] I-smoothing to the past emphasis can be superior to anything I-smoothing to the ML gauge

Slide 7

Experimental conditions

Slide 8

EBN50: MPE versus (B)MMI, fourth iter, =0.053 0.6% from scratching off measurements 0.9% from boosting The pattern is MPE I-smoothed to MMI (MPE - > MMI).

Slide 9

EBN700 setup: MPE versus (B)MMI BMMI-c->ML is superior to MPE->MMI 0.3% from boosting ,0.7% from scratching off of insights

Slide 10

EBN50: I-smoothing - > ML versus - > past cycle ( =0.053) I-smoothing to the past emphasis as opposed to the ML appraise Smoothing to the past emphasis is better, by 0.2% for MPE and 0.4% for BMMI-c

Slide 11

determination We have Introduced… An adjustment to the MMI target work - Error-boosting An alteration to the MMI preparing methodology – the crossing out of measurements We have tried different things with an absolutely MMI-based model-space discriminative preparing system and found that it at times yet not generally prompts to significant upgrades over our past MPE-based technique