discriminative training
play

Discriminative Training February 19, 2013 Tuesday, February 19, 13 - PowerPoint PPT Presentation

Oct 20, 2023 •158 likes •983 views

Discriminative Training February 19, 2013 Tuesday, February 19, 13 Noisy Channels Again p ( e ) English source Tuesday, February 19, 13 Noisy Channels Again p ( g | e ) p ( e ) English source German Tuesday, February 19, 13 Noisy

  1. Discriminative Training February 19, 2013 Tuesday, February 19, 13

  2. Noisy Channels Again p ( e ) English source Tuesday, February 19, 13

  3. Noisy Channels Again p ( g | e ) p ( e ) English source German Tuesday, February 19, 13

  4. Noisy Channels Again p ( g | e ) p ( e ) English source German decoder e ∗ = arg max p ( e | g ) e p ( g | e ) × p ( e ) = arg max p ( g ) e = arg max p ( g | e ) × p ( e ) e Tuesday, February 19, 13

  5. Noisy Channels Again e ∗ = arg max p ( e | g ) e p ( g | e ) × p ( e ) = arg max p ( g ) e = arg max p ( g | e ) × p ( e ) e Tuesday, February 19, 13

  6. Noisy Channels Again e ∗ = arg max p ( e | g ) e p ( g | e ) × p ( e ) = arg max p ( g ) e = arg max p ( g | e ) × p ( e ) e = arg max log p ( g | e ) + log p ( e ) e  1 � >  log p ( g | e ) � = arg max 1 log p ( e ) e | {z } | {z } w > h ( g , e ) Tuesday, February 19, 13

  7. Noisy Channels Again e ∗ = arg max p ( e | g ) e p ( g | e ) × p ( e ) = arg max p ( g ) e = arg max p ( g | e ) × p ( e ) e = arg max log p ( g | e ) + log p ( e ) e  1 � >  log p ( g | e ) � = arg max 1 log p ( e ) e | {z } | {z } w > h ( g , e ) Tuesday, February 19, 13

  8. Noisy Channels Again e ∗ = arg max p ( e | g ) e p ( g | e ) × p ( e ) = arg max p ( g ) e = arg max p ( g | e ) × p ( e ) e Does this look familiar? = arg max log p ( g | e ) + log p ( e ) e  1 � >  log p ( g | e ) � = arg max 1 log p ( e ) e | {z } | {z } w > h ( g , e ) Tuesday, February 19, 13

  9. The Noisy Channel -log p ( g | e ) -log p ( e ) Tuesday, February 19, 13

  10. As a Linear Model -log p ( g | e ) ~ w -log p ( e ) Tuesday, February 19, 13

  11. As a Linear Model -log p ( g | e ) ~ w -log p ( e ) Tuesday, February 19, 13

  12. As a Linear Model -log p ( g | e ) ~ w -log p ( e ) Tuesday, February 19, 13

  13. As a Linear Model -log p ( g | e ) Improvement 1: change to find better translations ~ ~ w w g -log p ( e ) Tuesday, February 19, 13

  14. As a Linear Model -log p ( g | e ) ~ w -log p ( e ) Tuesday, February 19, 13

  15. As a Linear Model -log p ( g | e ) ~ w -log p ( e ) Tuesday, February 19, 13

  16. As a Linear Model -log p ( g | e ) ~ w -log p ( e ) Tuesday, February 19, 13

  17. As a Linear Model -log p ( g | e ) Improvement 2: Add dimensions to make points separable ~ w -log p ( e ) Tuesday, February 19, 13

  18. Linear Models e ⇤ = arg max w > h ( g , e ) e • Improve the modeling capacity of the noisy channel in two ways • Reorient the weight vector • Add new dimensions ( new features ) • Questions • What features? h ( g , e ) • How do we set the weights? w Tuesday, February 19, 13

  19. beißt Hund Mann 18 Tuesday, February 19, 13

  20. beißt Hund Mann x BITES y 18 Tuesday, February 19, 13

  21. beißt Hund Mann x BITES y Mann beißt Hund Mann beißt Hund bites cat man chase dog man Mann beißt Hund Mann beißt Hund bite cat bite dog man man Mann beißt Hund Mann beißt Hund dog bites man bites dog man 19 Tuesday, February 19, 13

  22. beißt Hund Mann x BITES y Mann beißt Hund Mann beißt Hund bites cat man chase dog man Mann beißt Hund Mann beißt Hund bite cat bite dog man man Mann beißt Hund Mann beißt Hund dog bites man bites dog man 19 Tuesday, February 19, 13

  23. beißt Hund Mann x BITES y Mann beißt Hund Mann beißt Hund bites cat man chase dog man Mann beißt Hund Mann beißt Hund bite cat bite dog man man Mann beißt Hund Mann beißt Hund dog bites man bites dog man 19 Tuesday, February 19, 13

  24. beißt Hund Mann x BITES y Mann beißt Hund Mann beißt Hund bites cat man chase dog man Mann beißt Hund Mann beißt Hund bite cat bite dog man man Mann beißt Hund Mann beißt Hund dog bites man bites dog man 19 Tuesday, February 19, 13

  25. beißt Hund Mann x BITES y Mann beißt Hund Mann beißt Hund bites cat man chase dog man Mann beißt Hund Mann beißt Hund bite cat bite dog man man Mann beißt Hund Mann beißt Hund dog bites man bites dog man 19 Tuesday, February 19, 13

  26. Feature Classes Lexical Are lexical choices appropriate? bank = “River bank” vs. “Financial institution” 20 Tuesday, February 19, 13

  27. Feature Classes Lexical Are lexical choices appropriate? bank = “River bank” vs. “Financial institution” Configurational Are semantic/syntactic relations preserved? “Dog bites man” vs. “Man bites dog” 20 Tuesday, February 19, 13

  28. Feature Classes Lexical Are lexical choices appropriate? bank = “River bank” vs. “Financial institution” Configurational Are semantic/syntactic relations preserved? “Dog bites man” vs. “Man bites dog” Grammatical Is the output fluent / well-formed? “Man bites dog” vs. “Man bite dog” 20 Tuesday, February 19, 13

  29. What do lexical features look like? Mann beißt Hund bites cat man 21 Tuesday, February 19, 13

  30. What do lexical features look like? Mann beißt Hund bites cat man 21 Tuesday, February 19, 13

  31. What do lexical features look like? Mann beißt Hund bites cat man First attempt: score ( g , e ) = w > h ( g , e ) ( 1 , ∃ i, j : g i = Hund , e j = cat h 15 , 342 ( g , e ) = 0 , otherwise 21 Tuesday, February 19, 13

  32. What do lexical features look like? Mann beißt Hund bites cat man First attempt: score ( g , e ) = w > h ( g , e ) ( 1 , ∃ i, j : g i = Hund , e j = cat h 15 , 342 ( g , e ) = 0 , otherwise But what if a cat is being chased by a Hund ? 21 Tuesday, February 19, 13

  33. What do lexical features look like? Mann beißt Hund bites cat man Latent variables enable more precise features: score ( g , e , a ) = w > h ( g , e , a ) ( 1 , if g i = Hund , e j = cat X h 15 , 342 ( g , e , a ) = 0 , otherwise ( i,j ) 2 a 22 Tuesday, February 19, 13

  34. Standard Features • Target side features • log p(e) [ n -gram language model ] • Number of words in hypothesis • Non-English character count • Source + target features • log relative frequency e | f of each rule [ log #(e,f) - log #(f) ] • log relative frequency f | e of each rule [ log #(e,f) - log #(e) ] • “lexical translation” log probability e | f of each rule [ ≈ log p model1 (e|f) ] • “lexical translation” log probability f | e of each rule [ ≈ log p model1 (f|e) ] • Other features • Count of rules/phrases used • Reordering pattern probabilities 23 Tuesday, February 19, 13

  35. Parameter Learning 24 Tuesday, February 19, 13

  36. Hypothesis Space h 1 h 2 25 Tuesday, February 19, 13

  37. Hypothesis Space h 1 Hypotheses h 2 25 Tuesday, February 19, 13

  38. Hypothesis Space h 1 References h 2 26 Tuesday, February 19, 13

  39. Preliminaries We assume a decoder that computes: h e ⇤ , a ⇤ i = arg max h e , a i w > h ( g , e , a ) And K -best lists of, that is: { � e ⇤ i , a ⇤ h e , a i w > h ( g , e , a ) i ⇥ } i = K i =1 = arg i th - max Standard, efficient algorithms exist for this. 27 Tuesday, February 19, 13

  40. Learning Weights • Try to match the reference translation exactly • Conditional random field • Maximize the conditional probability of the reference translations • “Average” over the different latent variables • Max-margin • Find the weight vector that separates the reference translation from others by the maximal margin • Maximal setting of the latent variables 28 Tuesday, February 19, 13

  41. Learning Weights • Try to match the reference translation exactly • Conditional random field • Maximize the conditional probability of the reference translations • “Average” over the different latent variables • Max-margin • Find the weight vector that separates the reference translation from others by the maximal margin • Maximal setting of the latent variables 28 Tuesday, February 19, 13

  42. Problems • These methods give “full credit” when the model exactly produces the reference and no credit otherwise • What is the problem with this? • There are many ways to translate a sentence • What if we have multiple reference translations? • What about partial credit? 29 Tuesday, February 19, 13

  43. Problems • These methods give “full credit” when the model exactly produces the reference and no credit otherwise • What is the problem with this? • There are many ways to translate a sentence • What if we have multiple reference translations? • What about partial credit? 29 Tuesday, February 19, 13

  44. Cost-Sensitive Training • Assume we have a cost function that gives a score for how good/bad a translation is � (ˆ e , E ) ⇥� [0 , 1] • Optimize the weight vector by making reference to this function • We will talk about two ways to do this 30 Tuesday, February 19, 13

  45. K-Best List Example h 1 ~ w h 2 31 Tuesday, February 19, 13

  46. K-Best List Example #2#1 h 1 #3 ~ w #6 #5 #4 #7 #8 #9 #10 h 2 31 Tuesday, February 19, 13

Download Presentation
Download Policy: The content available on the website is offered to you 'AS IS' for your personal information and use only. It cannot be commercialized, licensed, or distributed on other websites without prior consent from the author. To download a presentation, simply click this link. If you encounter any difficulties during the download process, it's possible that the publisher has removed the file from their server.

Recommend

HMM-based acoustic model adaptation and discriminative training Steven Wegmann ICSI 11 April

HMM-based acoustic model adaptation and discriminative training Steven Wegmann ICSI 11 April

HMM-based acoustic model adaptation and discriminative training Steven Wegmann ICSI 11 April 2012 HMM-based adaptation and discriminative training are important techniques for improving accuracy Both procedures start with HMMs ML

745 views • 58 slides
CoFiGAN: Collaborative Filtering by Generative and Discriminative Training for One-Class

CoFiGAN: Collaborative Filtering by Generative and Discriminative Training for One-Class

CoFiGAN: Collaborative Filtering by Generative and Discriminative Training for One-Class Recommendation Jixiong Liu a , b , c , Weike Pan a , b , c and Zhong Ming a , b , c liujixiong@email.szu.edu.cn, { panweike, mingz } @szu.edu.cn a

855 views • 40 slides
An Integrated Framework for Margin-based Sequential Discriminative Training over Lattices using

An Integrated Framework for Margin-based Sequential Discriminative Training over Lattices using

Overview Background Unification of Margin-modified MPE and MMI An Integrated Framework for Margin-based Sequential Discriminative Training over Lattices using differenced Maximum Mutual Information (dMMI) Erik McDermott - Google Inc.

621 views • 31 slides
Generative vs. discriminative Generative Discriminative Belief network A is more More

Generative vs. discriminative Generative Discriminative Belief network A is more More

Generative vs. discriminative Generative Discriminative Belief network A is more More robust modular Dont need precise model Class-conditional densities are specification, so long as it is likely to be local, characteristic

1.09k views • 15 slides
A Comparison of Structural Correspondence Learning and Self-training for Discriminative Parse

A Comparison of Structural Correspondence Learning and Self-training for Discriminative Parse

A Comparison of Structural Correspondence Learning and Self-training for Discriminative Parse Selection Barbara Plank b.plank@rug.nl University of Groningen (RUG) The Netherlands NAACL HLT 2009 Workshop on Semi-supervised Learning for Natural

634 views • 27 slides
Multi-Task Learning for Improved Discriminative Training in SMT Patrick Simianer and Stefan

Multi-Task Learning for Improved Discriminative Training in SMT Patrick Simianer and Stefan

Introduction Related Work Algorithms Experiments Conclusion Multi-Task Learning for Improved Discriminative Training in SMT Patrick Simianer and Stefan Riezler Department of Computational Linguistics, Heidelberg University, Germany 1 / 22

487 views • 37 slides
Lecture 12 Discriminative Training, ROVER, and Consensus Michael Picheny, Bhuvana Ramabhadran,

Lecture 12 Discriminative Training, ROVER, and Consensus Michael Picheny, Bhuvana Ramabhadran,

Lecture 12 Discriminative Training, ROVER, and Consensus Michael Picheny, Bhuvana Ramabhadran, Stanley F . Chen, Markus Nussbaum-Thom Watson Group IBM T.J. Watson Research Center Yorktown Heights, New York, USA

1.29k views • 85 slides
Lecture 10 Discriminative Training, ROVER, and Consensus Michael Picheny, Bhuvana Ramabhadran,

Lecture 10 Discriminative Training, ROVER, and Consensus Michael Picheny, Bhuvana Ramabhadran,

Lecture 10 Discriminative Training, ROVER, and Consensus Michael Picheny, Bhuvana Ramabhadran, Stanley F . Chen IBM T.J. Watson Research Center Yorktown Heights, New York, USA {picheny,bhuvana,stanchen}@us.ibm.com 10 December 2012 General

1.16k views • 90 slides
Automatic Speech Recognition (CS753) Automatic Speech Recognition (CS753) Lecture 20:

Automatic Speech Recognition (CS753) Automatic Speech Recognition (CS753) Lecture 20:

Automatic Speech Recognition (CS753) Automatic Speech Recognition (CS753) Lecture 20: Discriminative Training for HMMs Instructor: Preethi Jyothi Mar 30, 2017 Discriminative Training Recall: MLE for HMMs Maximum likelihood estimation (MLE)

234 views • 22 slides
Winter School Day 5: Discriminative Training and Factored Translation Models MT Marathon 30

Winter School Day 5: Discriminative Training and Factored Translation Models MT Marathon 30

Winter School Day 5: Discriminative Training and Factored Translation Models MT Marathon 30 January 2009 MT Marathon Winter School, Lecture 5 30 January 2009 1 The birth of SMT: generative models The definition of translation probability

1.04k views • 75 slides
Generative and discriminative classification techniques Machine Learning and Category

Generative and discriminative classification techniques Machine Learning and Category

Generative and discriminative classification techniques Machine Learning and Category Representation 2014-2015 Jakob Verbeek, November 28, 2014 Course website: http://lear.inrialpes.fr/~verbeek/MLCR.14.15 Classification Given training data

295 views • 28 slides
Discriminative Models Joakim Nivre Uppsala University Department of Linguistics and Philology

Discriminative Models Joakim Nivre Uppsala University Department of Linguistics and Philology

Discriminative Models Joakim Nivre Uppsala University Department of Linguistics and Philology joakim.nivre@lingfil.uu.se Discriminative Models 1(11) 1. Generative and Discriminative Models 2. Log-Linear Models 3. Local Discriminative Models

278 views • 11 slides
Generative and discriminative classification techniques Machine Learning and Object Recognition

Generative and discriminative classification techniques Machine Learning and Object Recognition

Generative and discriminative classification techniques Machine Learning and Object Recognition 2015-2016 Jakob Verbeek, December 11, 2015 Course website: http://lear.inrialpes.fr/~verbeek/MLOR.15.16 Classification Given training data

862 views • 61 slides
Discriminative vs. Generative Learning CS 760@UW-Madison Goals for the lecture you should

Discriminative vs. Generative Learning CS 760@UW-Madison Goals for the lecture you should

Discriminative vs. Generative Learning CS 760@UW-Madison Goals for the lecture you should understand the following concepts the relationship between logistic regression and Nave Bayes the relationship between discriminative and

289 views • 27 slides
Joint Optimisation of Tandem Systems using Gaussian Mixture Density Neural Network Discriminative

Joint Optimisation of Tandem Systems using Gaussian Mixture Density Neural Network Discriminative

Joint Optimisation of Tandem Systems using Gaussian Mixture Density Neural Network Discriminative Sequence Training Chao Zhang and Phil Woodland March 8, 2017 Cambridge University Engineering Department Introduction Tandem Systems as Mixture

926 views • 17 slides
Machine Translation at Edinburgh Factored Translation Models and Discriminative Training Philipp

Machine Translation at Edinburgh Factored Translation Models and Discriminative Training Philipp

Machine Translation at Edinburgh Factored Translation Models and Discriminative Training Philipp Koehn, University of Edinburgh 9 July 2007 Philipp Koehn, University of Edinburgh 9 July 2007 EuroMatrix 1 Overview Intro: Machine

870 views • 76 slides
Staying Safe around Dogs & Reducing the Risk of Dog Related Injuries Canine Behaviour &

Staying Safe around Dogs & Reducing the Risk of Dog Related Injuries Canine Behaviour &

International Training Programme Staying Safe around Dogs & Reducing the Risk of Dog Related Injuries Canine Behaviour & Research Department Objectives & Learning Outcomes Discuss why and which types of dogs are likely to bite

768 views • 58 slides
Introduction to Deep Processing Techniques for NLP Deep Processing Techniques for NLP Ling 571

Introduction to Deep Processing Techniques for NLP Deep Processing Techniques for NLP Ling 571

Introduction to Deep Processing Techniques for NLP Deep Processing Techniques for NLP Ling 571 January 4, 2017 Gina-Anne Levow Roadmap Motivation: Applications Language and Thought Knowledge of Language

794 views • 40 slides
Schedule and Cost Stewart Loken Physics Division Lawrence Berkeley National Laboratory PDG

Schedule and Cost Stewart Loken Physics Division Lawrence Berkeley National Laboratory PDG

Schedule and Cost Stewart Loken Physics Division Lawrence Berkeley National Laboratory PDG Computing Review, September 17, 2010 Stewart Loken (LBNL), Page 1 Funding History Proposal submitted to NSF and DOE in 2007 NSF funded in 2007

477 views • 12 slides
Small Modular Reactor Design and Deployment IREA 10/14/2018 www.inl.gov INL SMR Activities

Small Modular Reactor Design and Deployment IREA 10/14/2018 www.inl.gov INL SMR Activities

INL/MIS-15-34247 Small Modular Reactor Design and Deployment IREA 10/14/2018 www.inl.gov INL SMR Activities INL works with all vendors to provide fair access to the laboratory benefits INL works with industry on SMR technology and

634 views • 24 slides
Communication and Language Chapter 22 Chapter 22 1 Outline Communication Grammar

Communication and Language Chapter 22 Chapter 22 1 Outline Communication Grammar

Communication and Language Chapter 22 Chapter 22 1 Outline Communication Grammar Syntactic analysis Problems Chapter 22 2 Communication Classical view (pre-1953): language consists of sentences that are true/false (cf.

714 views • 57 slides
Grammar and word order Grammar and word order Grammar Grammar Includes morphology and syntax

Grammar and word order Grammar and word order Grammar Grammar Includes morphology and syntax

LIGN171: Child Language Acquisition http://ling.ucsd.edu/courses/lign171 http://ling.ucsd.edu/courses/lign171 LIGN171: Child Language Acquisition Grammar and word order Grammar and word order Grammar Grammar Includes morphology

565 views • 34 slides
For Monday Read chapter 23, sections 1-2 FOIL exercise due Program 4 Any questions?

For Monday Read chapter 23, sections 1-2 FOIL exercise due Program 4 Any questions?

For Monday Read chapter 23, sections 1-2 FOIL exercise due Program 4 Any questions? Natural Language Processing Whats the goal? Communication Communication for the speaker: Intention: Decided why, when, and what

476 views • 30 slides
Autonomous Intelligent Robotics Instructor: Shiqi Zhang

Autonomous Intelligent Robotics Instructor: Shiqi Zhang

Spring 2018 CIS 693, EEC 693, EEC 793: Autonomous Intelligent Robotics Instructor: Shiqi Zhang http://eecs.csuohio.edu/~szhang/teaching/18spring/ Natural language processing Slides adapted from Ray Mooney Natural Language Processing NLP

429 views • 21 slides

More recommend