A Memory-Based Approach to Semantic Role Labeling

1
A Memory-Based Approach to Semantic Role Labeling

• Beata Kouchnir
• Tübingen University

05/07/04
2
Introduction

• Applying Memory-Based Learning to the task of
  Semantic Role Labeling (using the TiMBL software)
• Data is processed by chunks, except for the
  target verb chunks, which can contain modals and
  negation
• Task is split into two modules
• Recognition identifies the arguments of a target
  verb
• Labeling assigns a semantic role to each argument

1
Beata Kouchnir
05/07/04
3
Memory-Based Learning

• Training instances are stored without abstraction
• Test instances are assigned the most frequent
  class within a set of k most similar examples
  (k-nearest neighbors)
• Similarity is computed based on a distance
  metric
• Overlap 1 if two values are the same, 0
  otherwise (for symbolic values)
• Modified value difference determines similarity
  based on co-occurrence of values with classes

2
Beata Kouchnir
05/07/04
4
Recognition Features

• Head word and POS of the focus element (head is
  last word of chunk)
• Chunk type one of the 12 chunks types
• Position in clause beginning, end or inside.
• Directionality with respect to target verb
  before, after, coincides
• Numerical distance (1 .. n) to the target verb.
• Adjacency to target chunk adjacent, not
  adjacent, inside the target chunk

3
Beata Kouchnir
05/07/04
5
Recognition Features (contd.)

• Target verb and voice passive if target verb is
  a past participle preceded by a form of to be
• Context the features head word, part of speech,
  chunk type and adjacency of the three chunks each
  to the left and right of the focus chunk

4
Beata Kouchnir
05/07/04
6
Labeling Features

• Word, POS and chunk sequence of the head words of
  all the chunks in the argument each sequence
  represents one value
• Clause information is argument a complete
  clause?
• Length of the argument in chunks
• Directionality, adjacency, target verb and voice
• Prop Bank roleset of target verb's first sense
  (86 of targets use first sense)

5
Beata Kouchnir
05/07/04
7
Evaluation

• Recognition module Prec. 53.21, Rec. 74.97, F
  62.25
• All features improved performance MVDM, k7 best
  parameter setting
• Labeling module Prec. 75.71, Rec. 74.60, F
  75.15
• POS-sequence and length worsen performance MVDM,
  k1 best parameter setting
• Overall development Prec. 44.93 , Rec. 63.12,
  F 52.50
• Overall test Prec. 56.86, Rec. 49.95, F 53.18

6
Beata Kouchnir
05/07/04
8
Conclusion and Future Work

• Recognizing arguments is more difficult than
  labeling
• Removing multiple A0-A5 arguments can increase
  precision
• IOB2 might not be the best representation
• Some chunkers can recognize recursive noun
  phrases
• Could improve results without adding too much
  comlexity
• Changing classifier's default parameters
  considerably improves performance

7
Beata Kouchnir
05/07/04