Semisupervised Learning Approaches

Posted in Science on September 12, 2008


Semisupervised Learning Approaches

Lecture slides:

  • Semi-Supervised Learning over Text
  • Statistical learning methods
  • Outline
  • Many text learning tasks
  • Semi-supervised Document classification
  • Document Classification: Bag of Words Approach
  • Supervised: Naïve Bayes Learner
  • Twenty NewsGroups
  • What if we have labels for only some documents?
  • Nigam et al.
  • E-step, M-step
  • Using one labeled example per class
  • 20 Newgroups
  • Downweight the influence of unlabeled examples by factor lambda
  • Why/When will this work?
  • EM for Semi-Supervised Doc Classification
  • Using Redundantly Predictive Features
  • Redundantly Predictive Features
  • Co-Training - 1
  • CoTraining Algorithm #1
  • CoTraining: Experimental Results
  • Co-Training for Named Entity Extraction
  • CoTraining setting
  • Co-Training Rote Learner
  • Co-Training - 2
  • Expected Rote CoTraining error given m examples
  • How many unlabeled examples suffice?
  • PAC Generalization Bounds on CoTraining
  • Co-Training Theory
  • What if CoTraining Assumption Not Perfectly Satistfied? - 1
  • Example 2: Learning to extract named entities
  • Co-Training for Named Entity Extraction
  • Bootstrap learning to extract named entities
  • Co-EM
  • CoEM applied to Named Entity Recognition
  • Bootsrapping Results
  • Some nodes are more important than others
  • Power-Law Distribution
  • What if CoTraining Assumption Not Perfectly Satistfied? - 2
  • What Objective Function?
  • What Function Approximators?
  • Gradient CoTraining
  • Example 3: Word sense disambiguation
  • Example 4: Bootstrap learning for IE from HTML structure
  • Example Bootstrap learning algorithms
  • What to Know
  • Further Reading

Author: Tom Mitchell, Carnegie Mellon University

Watch Video

Tags: Science, Lectures, Computer Science, Machine Learning, VideoLectures.Net