Ben Hutchinson: Acquiring the meaning of discourse markers

1
Ben HutchinsonAcquiring the meaning of
discourse markers

• Paper presentation
• Tomas Singliar
• CS3730 Feb 23, 2005

2
Outline

• Structural connective classification
• Corpus extraction
• Features
• Experiments
• Distance measures for 1NN
• Performance measures
• Results and implications
• Conclusions Discussion

3
Structural connective classification

• DM that behave like conjunctions
• Polarity
• POS (because), NEG (although)
• Veridicality
• VERIDICAL (while), NON-VERIDICAL (unless)
• Type
• ADDITIVE (but), TEMPORAL (before), CAUSAL
  (provided that)
• Gold-standard classes defined (ambiguity!)

4
Corpus extraction

• Automatically extracted from BNC
• Parsed witch Charniak parser
• (Syntax-based-rule)-based extraction
• Close to S probably a connective
• 12 noisy examples

5
Features

• Lexical
• Lemmatised high-freq words
• POS Tags (coarser-grained Penn TreeBank)
• Occurs in super- or sub-ordinate clause
• BE, DO, VP-ELLIPSIS ( then I will!)
• PRONOUNS1S, PRONOUNS3P

6
Features - syntactic

• POSITION (relative to clauses)
• Embedding depth in syntax tree
• Negation NEG-SUBJ, NEG-VERB, NPI-AND-NEG (not
  give any), NPI-WO-NEGATION
• 2 Dimensional both clauses represented
• MODALITY, MOOD
• PERFECT, PROGRESSIVE, TENSE

7
Features (cont)

• STRUCTURAL-SKELETON
• WORDS, NPS, VPS
• Embedded CLAUSES
• TEMPEX temporal expression (since!)

8
Classifiers

• 1-Nearest Neighbor
• Euclidian
• Smoothed Kullback-Leibner
• Jaccard coefficient
• Naïve Bayes
• 10-fold cross validation
• Implicit train/test split (done by Weka)

9
Performance 1NN
10
Feature selection

• Select most informative features
• Weka uses MI with class
• Co-occurring markers, example
• X, but Y DM Z indicates POS-POL between Y,Z
• Use only these for learning remove noise

11
Discussion

• Achievement of this paper?
• Structural connectives well classified because
  classification is learnable
• How does it follow from the results that there
  should be a separate TEMPORAL class?