Transcript Guiding Inference with Conceptual Graphs

Processing Metonymy and Metaphor
Dan Fass, as summarized/(mis-)interpreted
by Peter Clark
Metonymy and Metaphor
• Really part of the bigger problem of “non-literal
language”
• What exactly is “non-literal”?
– Departs from truth conditions
– Violates “standard” use of language
Metaphor
“Application of a descriptive term to an object or action to
which it is not literally applicable.” (Oxford Dictionary)
•
•
•
•
“My car drinks gasoline.”
“The computer died”
“The virus attacks the cell”
“The polymerase slides along the DNA” (?)
- whether something is a metaphor depends on what you/the
computer understands by that word, I.e. metaphor is relative
to the underlying representation.
4 Views of how to Process Metaphor
• Comparison view:
– Compare & match features between base and target
Car  person
Use  drink
Gasoline  water
But: any two things are similar in some respect; doesn’t account
for what is important about the metaphor
• Interaction view:
– Transfer (part of) a system of axioms from base to
target
4 Views of how to Process Metaphor
• Selection Restrictions Violations view:
– Metaphor = violation of semantic restrictions
– But:
• “All men are animals” (no violations, interpretation is
context dependent)
• Conventional Metaphor view:
– There are conventional metaphors, which can be
catalogued
• Time as a substance
• Argument as war
• More/happy is up
Metonymy
“Substitution for the thing meant of something
closely associated with it.”
• “The ham sandwich is waiting for his check.”
– NB more ambiguity here than meets the eye
•
•
•
•
“The kettle is boiling.”
“I’m just going to change the washing machine.”
“It’s your turn to clean out the rabbit.”
(NY times example)
Types of Metonymy
• Popular to catalog different metonymy types
• E.g., Lakoff and Johnson’s list of eight:
– PART for WHOLE (“Get your butt over here”)
– FACE for PERSON (“We need some new faces around
here”)
– PRODUCER for PRODUCT (“I’ll have a Lowenbrau”)
– CONTROLLER for CONTROLLED (“A Mercedes
rear-ended me”)
– INSTITUTION for PEOPLE RESPONSIBLE (“Exon
has raised its prices again”)
– PLACE for INSTITUTION (“The White House isn’t
saying anything”)
– PLACE for EVENT (“Remember the Alamo”)
• Not all metonymys fit these rules (“novel metonymys”)
Metonymy and Language Processing
• Metonymic relationships can link sentences
– “I found an old car on the road. The steering
wheel was broken”
• Metonymy and anaphora closely related
– Both allow one entity to refer to another
• “The ham sandwich is waiting for his check”
• “He is waiting for his check”
Metaphor vs. Metonymy
• Metaphor is type of Metonymy?
• Metonymy is type of metaphor?
• Completely different?
• Metaphor founded on similarity, metonymy on
contiguity.
• Metaphor is primarily is about understanding
(conceiving of one thing in terms of another)
• Metonymy is primarily about reference (one
entity stands for another)
“America believes in democracy” – can be interpreted both ways
Fass’s Approach
• Aspects of Wilks’ “preference semantics” in it.
• Given a pair of word senses, each word sense
suggests/implies properties about the other
– “suggests” = preferences/expectations (soft constraints)
– “implies” = assertions (hard constraints)
• Can categorize the nature of the match (the “semantic
relation”) between suggested/implied & actual properties
– “Collation” = this matching process
– “Collative Semantics” = his overall approach
Types of Match
• 4 preference-based semantic relations:
– Between suggested and actual properties
• Literal (“the man drank beer”)
• Metonymic (“the man drank the glasses”)
• Metaphorical (“my car drank gasoline”)
• Anomalous (“The idea drank the heart”)
• 3 assertion-based semantic relations:
– Between implied and actual properties
• Redundant (“female girl”)
• Inconsistent (“female man”)
• Novel (“tall man”)
Identifying Preference-Based Relation:
GIVEN: two word senses
FIND: the appropriate preference-based semantic relation
Preferences satisfied?
(i.e., preferences of each word
sense are compatible)
Literal
Do inference
Metonymic inference possible?
Metonymic
Relevant metaphor?
Metaphorical
Anomalous
Details: Metonymic Inferences
• 5 (ordered) rules:
– PART for WHOLE
– PROPERTY for WHOLE
– CONTAINER for CONTENTS
– CO-AGENT for ACTIVITY
– ARTIST for ART FORM
• Apply rules in turn:
– “Arthur Ashe is black”  “Arthur Ashe’s skin is black”
Details: Search for Metaphor
•
•
Match “relevant” fact from base with some fact in target.
E.g. “My car drinks gasoline”
– “drink” prefers an animal as agent, so:
a) Find fact about animals drinking: “animals drink drinks”
b) Find a matching fact about cars, where “match” means the
participants are siblings in the taxonomy: Here, “cars use
gasoline”
expend
drink(v.)
isa
use(v.)
liquid
isa
drink(n.)
c) If good enough match, it’s a metaphor
gasoline
Representation
• How to represent preferences/expectations?
• Three types of “sense frame” representations:
– Verbs, nouns, and adjective/adverbs (ie verb senses etc)
• Verbs and adj/adv prefer certain types of object,
specified by either:
– Concept name (if one exists), e.g. “drink” prefers
“animal” as agent (Concept name is “macro” for properties)
– Concept properties, e.g. “yellow” prefers a
bounded, physical, non-living entity.
• Nouns have properties, and thus can meet/not meet
these preferences
Concept (“noun”) Properties
• 7 Dimensions (Jackendoff-style)
– Boundedness
– extent (dimensionality)
– Composition
– behavior (state)
– Animacy
– biological category
– sex
Representation:
VERBS:
“isa” hierarchy
sf(eat1,
[[arcs,[[supertype,[ingest1,expend1]]]]
[node2,
Preferences
[agent,[preference,animal1]]
[object,[preference,food1]]])
node2 means
it’s a verb
Representation:
ADJECTIVES AND ADVERBS:
“isa” hierarchy
sf(yellow1,
[[arcs, [[superproperty,coloured1],
[property,yellow1]]]
Preferences follow…
[node1,
[[preference,
7 Dimensions:
[[bounds1,bounded1],
node1 means
boundedness
[composition1,physical1],
it’s an adj/adv
extent (dimension[extent1, [not1,zero_dimensional1]]
ality)
[animacy1,nonliving1]]]]]
composition
[assertion,
behavior (state)
[[color1,yellow1]]]]).
animacy
biological category
sex
Representation:
NOUNS:
sf(animal1,
[[arcs, [[supertype,organism1]]],
[node0,
[[biology1,animal1],
Properties (along 7 dimensions)
[composition1,flesh1],
[it1,drink1,drink1],
facts (triples)
[it1,eat1,food1]]]]).
sf(crook1,
[[arcs, [[supertype,criminal1]]],
[node0,
[[it1,steal1,valuable1]]]]).
node0 means
it’s a noun
“isa” hierarchy
facts
More on Representation
• Inheritance
– Inheritance with overrides
– Need to properly match facts from superclass
with facts from subclass during inheritance
• Primitives
– No semantic primitives!
– Everything defined in terms of everything else
– Bounded computation to avoid infinite loops
The Semantic Vector
- a data structure recording the matches between a
preference (e.g. “animal”) and an actual object
(e.g. “car”)
1) subsumption relation (“network path”)
Does A subsume B, B subsume A, or neither?
2) matching facts (“cell match”)
How many properties of A subsume/are
subsumed by/neither properties of B?
- Use heuristic scoring function to find “best match”
Result
• For a word pair, search the M*N possible word
senses. Find the best combination according to the
preceding algorithm.
• Just dealing with three-element sentences, e.g.
– “John baked the potatoes”
Related Work
• Katz, Wilks, Schank
• Pustejovsky
– “newspaper” has different aspects
– wants single definition + rules of semantic
composition
– Sure seems like noun + rules of metonymy
– Another example:
• “John baked the potatoes”
• “Mary baked the cake”
• Dolan