speech & nlp (fall 2014): conceptual dependency & linguistic relativity
TRANSCRIPT
Speech & NLP
Conceptual Dependency Theory
&
Linguistic Relativity
Vladimir Kulyukin
Outline
Background
Conceptual Dependency
Linguistic Relativity
Semantic Primitives of Conceptual Dependency
Conceptual Analysis of Natural Language
Mapping NL Inputs to Conceptual Dependency
Graphs
Background
Content Analysis
Language is a medium whose primary purpose is
communication
Primary focus of linguistics is to determine what
kinds of things can be communicated
Primary emphasis is on content, not form (this is
in direct opposition to formal language theory)
Emphasis on content is common in AI; emphasis
on form is more common in computational
linguistics
Is Pro-Content Anti-Syntax?
Primary focus on content does not mean that
the focus on anti-syntax
Syntax is very important but its role should be
secondary to the study of knowledge and
meaning
There should be no independent syntactic pass
over the input: syntactic & semantic processing
should go hand in hand
Conceptual Dependency Theory
Basic Axioms of CD Theory
For any two sentences that are identical in
meaning, regardless of language, there
should be only one representation
Any information in a sentence that is
implicit must be made explicit in the
representation of the meaning of that
sentence
Basic Definitions of CD Theory
The meaning propositions underlying language are called
conceptualizations
A conceptualization can be active or stative
An active conceptualization has the form:
<ACTOR, ACTION, OBJECT, DIRECTION, INSTRUMENT>
A stative conceptualization has the form:
IS_IN(OBJECT, STATE, VALUE) – OBJECT is in STATE whose
value is equal to VALUE
Event Representation
Every EVENT has:
ACTOR
ACTION performed by ACTOR
OBJECT that ACTION is performed on
DIRECTION in which ACTION is oriented
INSTRUMENT with which ACTOR does
ACTION
ACTOR & OBJECT
ACTOR is a concrete object (aka PICTURE
PRODUCER or PP)
ACTOR can decide to apply ACTION to another PP
called OBJECT
A rock is a PP but cannot be an ACTOR because it
cannot decide to apply ACTION to any other object
Honesty, justice, truth and other mass nouns are not
PPs
ACTION represent a physical action or a mental
action
Semantic Action & State Primitives
CD constructs the meaning of NL input from a
finite set of semantic primitives
Semantic primitives can be considered as an
interlingual vocabulary in terms of which one
can, in principle, represent the meaning of
every word in every language
To date, there is no universally accepted set of
semantic action & state primitives but the
search for this set continues
Linguistic Relativity
Sapir-Whorf Hypothesis: Alternative to Universal Semantic
Primitives
Edward Sapir, 1884 - 1939 Benjamin Lee Whorf, 1897 - 1941
Linguistic Relativity
Benjamin Lee Whorf (1897 – 1941) was an American linguist
& anthropologist whose mentor was Edward Sapir (1884 –
1939), another American linguist & anthropologist
Whorf formulated the Principle of Linguistic Relativity (aka
the Sapir-Whorf Hypothesis) that states that speakers of
different languages conceptualize and experience the world
differently due to linguistic differences in grammar and
usage
Linguistic Relativity Principle states that it is impossible to
find the universal set of semantic action and state primitives
Semantic Primitives
of
Conceptual Dependency
Verb Representation in CD
A verb is represented as a particular
combination of primitive actions (acts) and
states none of which are unique to that verb
but whose combination is entirely unique
R. Schank, R. Abelson. “Scripts, Plans, Goals, & Understanding:
An Inquiry into Human Knowledge Structures”
ATRANS
ATRANS – transfer of an abstract relationship (e.g., possession,
ownership, control)
Examples:
1) GIVE is an ATRANS of something to someone else
2) TAKE is an ATRANS of something to oneself
3) BUY is an ATRANS of something to oneself and another
ATRANS of money from oneself to the owner of something
The robot gave John a cup of coffee.
The robot took a cup of coffee from the coffee machine.
John bought a new car.
PTRANS
PTRANS – transfer of the physical location of an
object
Examples:
1) GO is an PTRANS of oneself to a place
2) PUT is an PTRANS of an object to a place
The robot went to the lab.
The robot put the block on the table.
PROPEL
PROPEL – application of a physical force to an object;
this primitive applies whenever any force is applied
Examples:
PUSH, PULL, KICK, THROW have the PROPEL primitive
The robot pushed the chair to the wall.
This is an instance of PROPEL by the robot to the chair
that caused a PTRANS of the chair from its current
location to the wall.
MOVE
MOVE – the movement of a body part of an
agent/animal by that agent/animal
Examples:
KICK, HAND have the MOVE primitive
The boy kicked the ball.
This is an instance of MOVE by the boy of his foot to
the ball that causes a PTRANS of the ball from its
current location to some unknown location.
GRASP
GRASP – the grasping of an object by an actor
Examples:
HOLD, GRAB have the GRASP primitive
The robot picked up the ball from the floor.
This is an instance of GRASP by the robot of the ball to
the ball that causes a PTRANS of the ball from the
floor into the robot’s gripper. This is also an instance of
MOVE by the robot of its gripper to the ball.
INGEST
INGEST – the taking of an object by an animal/agent to the inside of that
animal agent
Examples:
EAT, DRINK, SMOKE, BREATHE have the INGEST primitive
The robot charged.
John ate an apple.
These are instances of INGEST. The first sentence is an INGEST by the
robot of electricity inside the robot’s batter. The second sentence is an
instance of INGEST by John of the apple to John’s stomach.
EXPEL
EXPEL – the expulsion of an object from the body of an animal/agent to
the outside of the body
Examples:
SWEAT, CRY have the EXPEL primitive
Mary cried.
John spat on the floor.
Both sentences are instances of EXPEL. The object of the first instance of
EXPEL is tears. The object of the second instance of EXPEL is saliva.
MTRANS
MTRANS – the transfer of mental information within one animal/agent or between/among
animals/agents.
CD Theory partitions the agent’s memory into two components: CP (conscious processor
where current mental manipulation occurs) and LTM (long-term memory where things are
stored)
Examples:
TELL, INFORM, SEE, FORGET have the MTRANS primitive
Mary told the robot how to get to the lab.
The robot told Mary which rooms it had cleaned.
Both sentences are instances of MTRANS. Mary does an MTRANS of a route from some
location to the lab. The robot does an MTRANS of the rooms it had cleaned to Mary.
MBUILD
MBUILD – the construction of an agent/animal of new
information from old information.
Examples:
DECIDE, CONCLUDE, REMEMBER have the MBUILD primitive
The robot concluded that it is lost.
John remembered that he had promised Mary to take her to the
movies.
SPEAK
SPEAK – the production of sounds by an animal/agent.
Examples:
SHOUT, PURR, BEEP have the SPEAK primitive
The robotic car beeped twice.
Mary yelled at John.
ATTEND
ATTEND – the focusing of a sense organ by an animal/agent
toward a stimulus.
Examples:
ATTEND(EAR) – LISTEN
ATTEND(EYE) – SEE
ATTEND(NOSE) – SMELL
ATTEND(SKIN) – TOUCH
The robot detected a door.
John saw an exit.
Categorization of Primitive CD ACTs
Physical ACTs: 1) PROPEL, 2) MOVE, 3)
INGEST, 4) EXPEL, 5) GRASP
ACTs that cause state changes: 1)
PTRANS, 2) ATRANS
Instrumental ACTs: 1) SPEAK, 2)
ATTEND
Mental ACTs: 1) MTRANS, 2) MBUILD
CD Representation of States
States are presented as attribute-value pairs
The values come from arbitrary ranges
constructed by the knowledge engineer
For example, an agent’s health can be
represented on a scale from -10 to +10
CD has never formulated a coherent set of
state primitives comparable to its primitive acts
and adhered to by all its proponents
Conceptual Analysis
of
Natural Language
Conceptual Analyzer
Conceptual Analysis: CD Parsing
Natural Language Input
CD Graphs (aka CDs) Inference Engine
Modified and/or New CDs
LTM
CD Representation Rules
PP ACT
Picture Producer PP can perform some act ACT
CD Representation Rules
ACT PP
Some act ACT has some PP as its object
o
CD Representation Rules
ACT
Some act ACT is directed from PP2 to PP1
PP2
PP1 D
CD Representation Rules
ACT
Some act ACT receives something from PP2 and
gives it to PP1
PP2
PP1 R
CD Representation Rules
ACT1
Some act ACT1 is accomplished (instrumented) by another
act ACT2 done by some picture producer PP2
PP2
I
ACT2
CD Graph Examples
CD Graph Example 01
The robot went to the kitchen.
CD Graph Example 01
Robot
PTRANS O
Robot
Unknown
D
Kitchen
CD Graph Example 02
The robot went from the lab to the
kitchen.
CD Graph Example 02
Robot
PTRANS O
Robot
Lab
D
Kitchen
CD Graph Example 03
JOHN ATE AN APPLE.
D
I
D
CD Graph Example 03
John INGEST Apple
Unknown Mouth
MOVE John Hand O
Unknown Mouth
O
CD Graph Example 04
The robot saw a door.
D
I
D
CD Graph Example 04
Robot MTRANS Door
Camera CP
ATTEND Robot Camera O
Unknown Door
O
CD Graph Example 04
John saw an exit.
D
I
D
CD Graph Example 05
John MTRANS Exit
Eyes CP
ATTEND John Eyes O
Unknown Exit
O
CD Graph Example 06
The robot read a street sign.
D
I
D
CD Graph Example 06
Robot MTRANS Sign
Camera CP
ATTEND Robot Camera O
Unknown Sign
O
CD Graph Example 07
John promised to give Mary a book.
D
O
D
CD Graph Example 07
John MTRANS
LTM Unknown
ATRANS John Book O
John Mary
Mapping NL Inputs to CD Graphs
Syntactic Parsing vs. Conceptual Analysis
The objective of syntactic parsing is to
construct a parse tree (or multiple
parse trees) of the input
The objective of conceptual analysis
(CA) is to construct a conceptual
dependency graph representing the
meaning of the input
Expectations
Conceptual analysis (CA) is an
expectation-driven process
Syntactic parsers (e.g., Early parser)
also use expectations
In a syntactic parser, expectations
come from a grammar
in a conceptual analyzer, expectations
come from a library of CD structures
CD Database
Suppose that we have compiled a database of PP and CD
structures
For example, our database may have CD structures like
<INGEST :ACTOR NULL
:OBJECT NULL
:FROM NULL
:TO NULL
:INSTRUMENT NULL
:TIME NULL>
<PP :CLASS APPLE :REF NULL NUMBER: NULL>
<PP :CLASS BOOK :REF NULL NUMBER: NULL>
<PP :CLASS HUMAN :NAME JOHN :GENDER MALE>
CD Database
CD structures have slots and fillers
Consider this CD:
<INGEST :ACTOR <PP :CLASS HUMAN :NAME JOHN :GENDER MALE>
:OBJECT NULL
:FROM NULL
:TO NULL
:INSTRUMENT NULL
:TIME NULL>
In this CD, the slot :ACTOR has the filler <PP :CLASS HUMAN
:NAME JOHN :GENER MALE> while the slot :OBJECT has the
filler
Insight into CA Algorithm
The conceptual analyzer has access to the CD database, the
input, and a concept list (the list models a short-term
memory)
The basic operation of the conceptual analyzer is to read the
input, retrieve an CD (or a set of CDs), and fill the retrieved
CDs on the basis of the read input, and find a CD (or a set of
CDs) on the concept list that expect the newly constructed CD
as a filler of a slot
If such a CD is found on the concept list, the newly
constructed CD fills one of its slots
If no such CD is found, the new CD is placed on the concept
list
References & Reading Suggestions
R. Schank, C. Riesbeck W. A. (1981) Inside Computer
Understanding. Lawrence Erlbaum & Associates.
R. Schank, R. Abelson. (1977). Scripts, Plans, Goals,
and Understanding: An Inquiry Into Human Knowledge
Structures (Artificial Intelligence Series). Lawrence
Erlbaum & Associates.