NearSynonymy and Lexical Choice

1
Near-Synonymy and Lexical Choice
Philip Edmonds Graeme Hirst (2002)
Summary by Michiel Bergmans
2
Introduction

• Representing fine-grained meanings of
  near-synonyms and their differences
• Clustered model of lexical knowledge
• Lexical choice
• Natural Language Generation
• Machine Translation
• Intelligent Thesaurus
• Automatic (post)editing of text

3
Synonymy

• Absolute Synonyms
• Substitutable in any context
• No change to truth value, communicative effect or
  meaning
• Quite rare (mostly dialectal variation and
  technical terms)
• Near-synonyms (plesionyms)
• Close in meaning

4
Synonymy - examples

• Absolute Synonyms
• Underwear (AmE) pants (BrE)
• Groundhog woodchuck
• Near-synonymy
• Woods forest
• Pissed drunk
• Skinny slim

5
Differences in near-synonyms

• Differ in
• Expression of concepts and ideas
• Manner
• Frequency
• Degree

6
Differences in near-synonyms

• Dimensions of variation
• Denotational variations
• Forest woods
• Stylistic variations
• Formal vs. informal (inebriated pissed)
• Forceful (ruin annihilate)
• Expressive variations
• Daddy father
• Structural variations
• Closed ajar

7
Problems to be solved

• Types of variation are qualitatively different,
  must be seperately modeled
• Difference in manner of conveying concepts,
  either in emphasis or indirectness
• Meanings and differences can be fuzzy
• Differences can be multidimensional
• Not just simple features, but concepts that
  relate roles and aspects of situation
• Differences depend on context

8
Problems with other computational models of the
lexicon

• Adequacy of coverage of near-synonyms
• Engineering (design, efficiency, robustness)
• Tractability of reasoning

9
Granularity of representation

• Level of detail used to represent meaning of
  words
• Coarse-grained crude
• Fine-grained subtle distinctions
• Granularity ? specificity (representation of
  concepts vs. concepts)
• Near-synonyms share coarse-grained, but differ in
  fine-grained concepts

10
Outline of model

• Meaning arises out of context-dependent
  combination of context-independent denotation and
  set of explicit differences to near-synonyms
• Each word and its near-synonyms form a cluster
• Core denotation
• Differentiation at subconceptual/stylistic level
  of semantics

11
Problems to be solved

• F

12
Example error nouns
13
Outline of model

• Core denotation inherent context-independent
  denotation shared by alle of its near-synonyms
• Peripheral concepts fine-grained denotational
  destinctions, used to represent non-necessary,
  indirect aspects of word meaning
• Represent differences explicitly for reasoning
  about them

14
Distictions

• Denotational distinctions
• w (frequency strength indirectness concept)
• blunder (usually medium implication Stupidity)
• Expressive distinctions
• w (frequency strength attitude entity)
• blunder (always medium pejorative V1)
• Stylistic distinctions
• w (degree dimension)
• blunder (high concreteness)

15
Problems with other methods to compute lexical
similarity

• Lack required precision
• Taxonomic hierarchies and semantic networks treat
  near-synonyms as absolute synonyms
• Standard dictionary definitions not fine-grained
  enough and circular
• Corpus-based methods not yet capable of
  uncovering more subtle differences

16
Comparing distinctions

• Seemingly incommensurate dimensions
• Usually interrelationships that can be modeled
  and exploited
• Forest wildness, size, distance to civilization
• Simplifying assumptionIndependent dimensions
  that can be reduced to true numeric dimensions

17
(No Transcript)
18
Lexical choice

• Denotational preferences
• Expressive preferences
• Stylistic preferences

19
Lexical choice

• Approximate matching algorithm to satisfy as many
  preferences as possible
• Two-tiered lexical-choice algorithm
• Chose cluster
• Chose near-synonym from cluster

20
I-Saurus

• Extends Stedes MOOSE (1999)
• To find globally preferred sentence plan, make
  the most preferred local choices
• Efficient only a few near-synonyms per cluster
• Robust Right coarse-grained meaning lexacilized,
  even if a poor near-synonym is chosen

21
Example
22
Conclusion

• Clustered model keeps advantages of conventional
  model
• Efficient paraphrasing
• Lexical choice at coarse grain
• Mechanisms for reasoning
• Overcomes shortcomings concerning near-synonymy
• Subconceptual/stylistic level more expressive
  than top level
• Yet tractable and efficient processing because it
  partitions expressiveness in small clusters

23
Open problems

• Recovering nuances from text
• Evaluating effectiveness of similarity measures
• Determining similarity of conceptual structures
• Understanding complex interaction of lexical and
  structural decisions during lexical choice
• Exploring requirements for logical inference
• Modeling other aspects of fine-grained meaning
  (e.g. emphasis)
• Understanding context-dependent nature of lexical
  differences and lexical knowledge

24
Comment

• Many open questions
• I-Saurus is just a prototype, no proof of
  performance
• Similarity functions?

25
Questions?