Sparse Forward-Backward for Fast Training of Conditional Random Fields

Abstract

Complex tasks in speech and language processing often include random variables with large state spaces, both in speech tasks that involve predicting words and phonemes, and in joint processing of pipelined systems in which the state space can be the labeling of an entire sequence. In large state spaces, however, discriminative training can be expensive, because it often requires many calls to forward-backward. Beam search is a standard heuristic for controlling complexity during Viterbi decoding, but during forward-backward, standard beam heuristics can be dangerous, as they can make training unstable. The authors introduce sparse forward-backward, a variational perspective on beam methods that uses an approximating mixture of Kronecker delta functions. This motivates a novel minimum-divergence beam criterion based on minimizing Kullback-Leibler (KL) divergence between the respective marginal distributions. This beam selection approach is not only more efficient for Viterbi decoding, but also more stable within sparse forward-backward training. For a standard text-to-speech problem, they reduce CRF training time fourfold -- from over a day to 6 hours -- with no loss in accuracy.

Open PDF

Document Details

Document Type
Technical Report
Publication Date
Jan 01, 2006
Accession Number
ADA443633

Entities

People

  • Andrew McCallum
  • Charles Sutton
  • Chris Pal

Organizations

  • University of Massachusetts Amherst

Tags

DTIC Thesaurus Topics

  • Abstracts
  • Accuracy
  • Computations
  • Computer Science
  • Computer Vision
  • Data Sets
  • Decoding
  • Delta Functions
  • Hidden Markov Models
  • Markov Models
  • Models
  • National Security
  • Probability
  • Random Variables
  • Sequences
  • Standards
  • Training

Fields of Study

  • Engineering

Readers

  • Adaptive Control and Estimation with Uncertainty in Dynamic Systems.
  • Neural Network Machine Learning.
  • Speech Processing/Speech Recognition.

Technology Areas

  • AI & ML
  • AI & ML - Bayesian Inference
  • AI & ML - Machine Learning Algorithms
  • AI & ML - Machine Translation
  • Space
  • Space - Spacecraft Maneuvers