Transcript Interpreting of dictionary definitions

Interpreting Dictionary
Definitions
Dan Tecuci
May 2002
Problem Description
• Interpretation = translating into one's own language
• Given:
• Dictionary definitions of actions
• KB = set of primives/components
•
•
•
•
Argument structure
Text generation
Axioms
Pre/post/during conditions
• User input
• Produce
• Representations of actions
Example
• Dictionary: carry = to move while supporting
• Challenges:
• Identify what primitive components are referred
• Move, Support
• Deal with missing arguments
• An agent moves an object while supporting it
• Resolve references between them
• Agent1 moves object2 while Agent1 supports object2
• Identify deep semantic relations among them
• There are two subevents of Carry, one in which Agent1
moves object2 and one in which Agent1 supports object2 so
that object2 cannot fall, and they happen in parallel
What Is the Goal?
• Acquire different kinds of knowledge:
• Taxonomic
• Semantics of actions
• Argument structure
• In order to:
•
•
•
•
Accelerate knowledge acquisition
Execute actions
Talk about them
Understand when someone talks about them
Knowledge of Argument Structure
• What arguments does the verb have and
where they surface (position)
• Multiple ways in which an argument can
surface. –
• E.G.
• V+o[+a] -> carry something [somewhere].
• The reverse of text-gen
Motivation
• Why this task
• Fast, automatic knowledge acquisition
• Language understanding and generation
• Available source of knowledge
• Why dictionaries
• Structured source of knowledge
• Taxonomic
• Argument structure
• Could be extended to full natural language
• Has been done before (manually)
Related Work - R. Amsler
• "The structure of MW dictionary" 80
• Analyses definitions based on “kernels”
(superclasses)
• Main goal - build a taxonomy of motion verbs
• Other
• Procedure to analyse the argument structure of
motion verbs (look at usage in other definitions and
use componential analysis)
• Manual WSD, manual kernel identification,
automatic taxonomy building
Related Work - C. Barrière
• “From a children dictionary to a LKB”
• Automatic translation of dictionary
definitions into a knowledge representation
formalism
• Specifics
• Uses an intermediary representation
• Only 1 sense of a word is analysed
• Children dictionary has usage examples
Related Work - C. Hastings
• Tries to acquire word (mainly verb) meaning from
context (sentence)
• Uses LINK parser, semantic knowledge, terrorism
domain
• KB has detailed info, fine-grained constraints
• Uses rules based on sentence structure to detect
case-role assignment
• Syntactic/semantic knowledge is expressed in the
same formalism (LINK)
• Algorithm:
• Identify slot fillers
• Based on this, identify matching components
Related Work - FrameNet
• Mainly the same goal, but bottom-up
• Not based on composing a set of primitives
• Manually annotate sentences, automatically
capture the organization of the annotation
results
• Frame –
• frame elements
• How FEs are realized in language
• Executable?
Dictionary Definitions
• Advantages
+ Good source of taxonomic knowledge
+ Follow a “genus-differentia” pattern
+ Some dictionaries tend to define everything in terms of a
basic vocabulary
• Disadvantages
- Definitions are elliptical
- Incomplete sentences (not easier to parse then NL)
- Leave blank argument positions that are nearly always filled in
usage
- One definition does not provide enough info
What We Need
• What kind of knowledge do we need for
such a task?
• Knowledge about primitive components
• Semantic - meaning
• Syntactic - argument structure
• How to determine when a concept is referenced?
• Knowledge about how to compose them and
how this is reflected in language
Complex Actions
• What are complex actions?
• How to discover them?
• Dictionaries
Example - Steps
•
•
Carry = to move while supporting
Steps:
1.
2.
3.
4.
5.
Identify referenced concepts/components
Identify their arguments
How are the components assembled
Resolve references
Get knowledge about argument structure
Identifying components
Task
i) Identify verbs
Method
POS tagging
- they constrain and interrelate the
entities mentioned in sentences.
ii) What sense(s) are
used?
iii) What components do
they match?
Example:
move #1, support#4
Move, Support
Word-sense
disambiguation
SME Dictionary
Identifying the arguments
• Arguments of Move, Support ?
• From
• definition
• Problem: definitions omit args that are usu. present
• KB
• Minimal number of required args
• Arguments of Carry
• Suggested by:
• syntax
• dictionary: transitive/intransitive
• definitional patterns - Carry ISA Move
• How are they related ?
Assembling the components
• Meaning of the whole = function of parts and the way they
are composed
• Discover deep semantic relations among components
referenced
• Prepositions – clues
• Example:
• “while” – co-temporality or detraction
• “by”- by-means-of, agent, time, location…
• Rules based on features of components
• How to test if a component is
correct/coherent? (test-cases?)
Resolving co-references
• What object are co-referential?
the agent of Move = the agent of Support
the object of Move = the agent of Support
• How to do it?
• Heuristic – assume everything maps unchanged
unless there is reason to believe otherwise
• Matching
• Machine learning
KB
• Move
• 16 senses in WN 1.6
• 8 represented in CompLib – tr. & intr.
• What about argument structure?
• Multiple argument structure can correspond to a sense
• Argument structures
• Subj + Move => agent ~ Subj
• Subj + Move + DObj => agent ~ Subj, object ~
object
• …
KB (cont.) – Arg Struct for Move
• Full arg structure for Move
“agent moves object over distance using
instrument along path from source to
destination”
• How to
• Acquire such knowledge?
• Express it?
• All possible trees
• Constraints (Subj always before Pred.)
Observations
• Verb definitions differ in level of generality
• More general - elliptical definitions
• “Carry = move while supporting”
• Inherit more from primitive components
• More specific - more complete definitions
• “Bioremediation = treating waste or pollutants by
the use of microorganisms (as bacteria) that can
break down the undesirable substances”
• Highly specialized versions of supers
What's next
•
•
•
•
Focus on a subproblem (e.g. WSD)
Get data (FrameNet?)
Design an experiment
Compare to existing methods
Research questions
• What dictionary to use? (WN?)
• How to represent arg struct knowledge?
• How would the special nature of the
knowledge we have might help in this task?
• actions can be executed
• their results can be tested
• Does compositionality help?