MaxMargin Matching for Semantic Role Labeling

1
Max-Margin Matching for Semantic Role Labeling

• David Vickrey
• James Connor
• Daphne Koller
• Stanford University

2
Overview

• We consider two complementary models for the same
  task
• In our case, one discriminative, one generative
• Combine the models by feeding output predictions
  into a discriminative classifier
• Try two different classifiers for combining
  multi-class SVM and max-margin matching

3
Semantic Role Labeling

• Label the arguments of a verb in context

I gave the dog a bone.
Giver
Recipient
Gift

• PropBank 1m labeled words (Wall Street Journal)

4
Syntactic Model

• Most useful features
• Argument word/part of speech
• Path from argument to verb in parse tree

S
VP
NP
NP
NP
VP
I
gave
dog
bone

• Use a standard classifier, e.g. SVM
• One vs. All classifier for each possible argument
  type
• Trained across all verbs at once

5
Semantic Model

• Data set words occurring as Eater for verb eat
• dog, cat, he,
• Usually, will either be a person or an animal
• We want to generalize to unseen animals or people
• Google Sets problem
• Our idea use a word hierarchy to find categories
  of words
• Used WordNet as the hierarchy
• Selected categories using Bayesian score
• Will be presented tomorrow
• Bayesian Methods for Natural Language Processing
• On its own, improves log-likelihood of test sets
  on PropBank
• Train one model for each argument for each verb

6
Combining Models

• For each word w in a sentence with verb v
• For each possible argument (Giver, Gift,
  Recipient, etc.)
• Margin of One vs. All classifier using syntactic
  features
• log Pv(Arg w) using semantic model trained on
  Arg and verb v
• Use these as inputs to multi-class SVM
• One weight for each argument for each model (not
  specific to v)

I wanted my dog to eat the pickle.
Syntax
Semantic
Dog?
Eater 1.0a-0.29b
Food 1.2c-1.39d
7
Results

• Tested on first 500 frames ( ½ of the data)

8
Max-Margin Matching

• Each argument should be assigned to only one word
• Complete bipartite graph
• Weight of edge from a word w to argument a
• Syntax only Margin of classifier for a applied
  to w
• Syntax and Semantic (weighted) sum of
  confidences of each
• Same set of weights as in Multi-class SVM
• Apply max-margin matching learning for these
  weights

I
1.0
1.0a-0.29b
Eater
dog
Food
pickle
1.2c-1.39d
1.2
9
Results

• Can do matching for any type of training of
  weights

10
Results Summary

• Classifying using confidences of One vs. All can
  help
• Improving One vs. All classifier may remove this
  benefit
• Combining the models in a classifier worked
• Able to improve using WordNet
• However, only worked with both matching and
  training!
• Matching helped, but training the max-margin
  matching did not
• Why?
• Not that much data (shared weights across all
  verbs)
• Not that many parameters

11
Future Directions

• Previous work used a Markov random field over
  classification decisions
• Cant do exact inference
• Can include potentials besides one word per
  argument
• We could try to extend max-margin matching
• Unlabeled data
• Bootstrap between different classifiers
• When to include an example?
• High confidence under a single classifier?
• Or, high confidence in combined classifier?

Toutanova, Haghighi, Manning, ACL 2005