Transcript Semantics - Homepages | The University of Aberdeen

CS4025: Semantics
l
l
l
Representing meaning
Semantic interpretation
Word meaning
For more information: J&M, chap 14, 16 in
1st ed; 17, 19 in 2nd
Computing Science, University of Aberdeen
1
NL Understanding
Understanding written text
» Which books are bestsellers
» Who wrote them
For now, focus on “AI” approach
» explicit models of grammar, meaning, etc
Computing Science, University of Aberdeen
2
Stages
Morphology: analyse word inflection
Syntax: determine grammatical structure
Semantics: convert to a form that is meaningful to
a computer
» eg, SQL query
Pragmatics: influence of context
» eg, what them refers to
Computing Science, University of Aberdeen
3
Example
Original: Who wrote them
morph: who write/past them
Grammar: [verb=write, subject=who, object=them]
semantics: Select title, firstname, lastname from [X]
pragmatics:
» Select title, firstname, lastname from books
» Where salesthisyear >10000
Computing Science, University of Aberdeen
4
Definition
Semantic interpretation rewrites a parse tree into
a “meaning representation”
» Logic, SQL, knowledge base
Poorly understood compared to syntax
» apps that need complex semantics, like database front
ends or high-quality MT, have had limited success in the
past
Computing Science, University of Aberdeen
5
Meaning
How can we represent the meaning of an English
sentence?
Programming languages: “meaning” is the
equivalent machine code
a = b +c
means
load a
add b
store c
We could represent meaning as programs in some
language, in which case NLU would be a kind of
“compilation”
Computing Science, University of Aberdeen
6
Meaning Representation in NL
Many possibilities
–
–
–
–
executable programs
logical formulas
AI knowledge representation
nothing
No consensus on what is best - basic problem in
philosophy and psychology
Computing Science, University of Aberdeen
7
Criteria for an ideal MRL
Unambiguous
Able to express all necessary shades of meaning
for the application domain
Verifiability – system can tell whether a
statement is true according to a knowledge base
Canonical – different sentences with the same
meaning are mapped to the same representation
Support of inference
Computing Science, University of Aberdeen
8
Example: John passed CS1001
Different representations
» Program: C (or SQL) code to add an appropriate entry to
a student database
» Logic: pass(John, CS1001)
» AI Semantic Net
Agent
John
Computing Science, University of Aberdeen
Object
Pass
CS1001
9
Program as representation
Translate English into SQL (C, ...)
» MS English Query / AccessELF
– “List the bestsellers” translated into “Select titles from
books where sales>10000”
» Usually need a different translator for each application
– Good authoring environments for semantic rules are
essential
Computing Science, University of Aberdeen
10
Logic as a Representation
Translate into (first-order) logic
John is a man
John eats spinach
John sold all of his stocks
man(John)
eat(John,spinach)
(X)(stock(X) & own(John,X))  sell(John, X))
John sold Peter all of his stocks
(X)(stock(X) & own(John,X))  sell(John,X,Peter))
Computing Science, University of Aberdeen
11
Logic as Representation (2)
Good points
» Can represent any meaning (if you are inventive enough
about predicates etc.)
» Good support for compositionality, arbitrarily complex
statements
» Good support for quantifiers (all, some,...)
Bad points
» Doesn’t seem to really match the way people think.
– does  really mean some?
Computing Science, University of Aberdeen
12
Case Frames as a Representation
Form of (AI) semantic network
Assume verbs (and other words) are objects with
relations
AGENT - the person/thing acting
THEME - the person/thing acted upon
BENEFICARY - [of action]
AT-LOC - where action happened
Computing Science, University of Aberdeen
13
Example
John gave Peter the ball
John gave the ball to Peter
The ball was given to Peter by John
are all interpreted as
GIVE
agent = John
theme = ball
beneficiary = Peter
Computing Science, University of Aberdeen
14
Meaning Primitives
Meaning primitives are a fixed set of concepts/
roles etc. in terms of which any meaning can be
expressed
Makes reasoning, e.g. about whether two meanings
are the same, simpler.
Example: PURCHASE act
John bought the book from Sam
Sam sold the book to John
Difficult to define small set of primitives
» Conceptual Dependency was one serious attempt
Computing Science, University of Aberdeen
15
Conceptual Dependency
Primitives
»
»
»
»
»
»
»
ATRANS - abstract transfer
PTRANS - physical transfer
MTRANS - mental transfer
PROPEL - apply force to an object
INGEST - eat, drink, etc
CON - conceptualise
etc
Computing Science, University of Aberdeen
16
Example: "John bought a book from
Mary."
(BI-CAUSE
(SOURCE (ATRANS (ACTOR MARY)
(OBJECT BOOK)
(FROM MARY)
(TO JOHN)
(TIME PAST)))
(TARGET (ATRANS (ACTOR JOHN)
(OBJECT MONEY)
(FROM JOHN)
(TO MARY)
(TIME PAST))))
Computing Science, University of Aberdeen
17
Example: "Bob threw the ball to Bill."
(PTRANS (ACTOR BOB)
(OBJECT BALL)
(FROM BOB)
(TO BILL)
(TIME PAST)
(INSTRUMENT (PROPEL (ACTOR BOB)
(OBJECT BALL)
(FROM BOB)
(TO BILL)
(TIME PAST)))
Computing Science, University of Aberdeen
18
Knowledge Bases
Represent meaning using objects in a large AI
knowledge base
» CYC project - 15-year project to build a knowledge base
which holds the kind of general world knowledge that
people have
» Use Cyc primitives and KR language to represent
meaning?
Computing Science, University of Aberdeen
19
MRLs and Logic
Most existing meaning representation languages
(frames, semantic nets, case frames etc). can be
viewed as subsets of First Order Logic (where the
expressive power is restricted or the set of
predicates etc. is partially determined)
Main deficiencies of first order logic – inability to
express default inferences and inferences based
on partial information
Computing Science, University of Aberdeen
20
Choosing an MRL: What is the Task?
Why are we processing this sentence? This could
influence the kind of meaning representation
chosen
» database interface - perhaps use SQL rep?
» AI system which reasons about John’s problems perhaps use logic or AI KR?
» Information retrieval, speech dictation, grammar
checking - don’t build any meaning representation?
Computing Science, University of Aberdeen
21
Semantic Interpretation
Rewriting the parse tree into the target
representation
May be based on rewrite rules that insert a
semantic structure X if the parse tree contains
syntactic structure Y
For generality/coverage, needs to be
compositional, that is the meaning of the whole is
some fixed function of the meanings of the parts
More on this in the next lecture
Computing Science, University of Aberdeen
22
Ex: List the books
S: imperative
V: List
NP: X
mapped into
Select X.<name> from X
There are also cheaper/simpler approaches to
semantic interpretation in use…
Computing Science, University of Aberdeen
23
Semantic interpretation: Semantic
grammar
An attempt to reduce the “distance” between
syntactic and semantic representations
Grammar is defined in terms of semantic
categories
»
»
»
»
»
»
TIMEQ-> When does FLIGHT-NP FLIGHT-VP
FLIGHT-NP -> Flight NUMBER
FLIGHT-NP -> Flight to CITY
FLIGHT-NP -> TIME flight to CITY
FLIGHT-VP -> depart
FLIGHT-VP -> leave
Computing Science, University of Aberdeen
24
Semantic Interpretation:
Template spotting
Look for patterns (either in text or parse tree)
which identify meaning fragments
»
»
»
»
Example: How much is a ticket to London?
How much specifies cost query
a ticket specifies a single one-way ticket
to London specifies destination
Must be in limited domain
Patterns looked for can be informed by knowledge
about how words relate to underlying concepts and
what syntactic properties words have.
Computing Science, University of Aberdeen
25
Doctor-on-Board Problem
Simple rewriting may not be sufficient. Example:
– Is there a doctor within 200 miles of the Enterprise
» Database doesn’t have Doctor entities, instead it has
DoctorOnBoard attr for ships
» Need to rephrase this as
– Is there a ship within 200 miles of the Enterprise which has
a doctor on board?
» Restructure query from human’s data model to
database’s data model
Distance between syntactic and semantic
structure significant in this example
Computing Science, University of Aberdeen
26
Lexical (Word) Meaning
Logic (classical) model
» bachelor(X) = male(X) & adult(X) & ¬married(X)
– But: the pope? Divorcee? Muslim with 3 wives?
» Father(X) = male(X)&parent(X)
– Man who adopts a child?
– Sperm-bank donor?
– Unmarried partner to woman raising a child?
– Unmarried (gay) partner to man raising a child?
Prototype/exemplar models may be better when
words don’t have formal “definitions”
Computing Science, University of Aberdeen
27
Word meaning for time-series data
Weather reports
» Saturday will be yet another generally dull day with early
morning mist or fog and mainly cloudy skies being
prevalent. There will be the odd bright spell here and
there, but it will feel rather damp with patches of mainly
light rain to be found across many parts, especially the
west and south.
Ongoing research project in CS Dept
Computing Science, University of Aberdeen
28
Conclusion
Converting sentences to a “meaning
representation” is hard
» No agreement on best meaning-rep
» Word meaning is hard to pin down
Limited success in small domains, but we can’t
semantically interpret general text
» but we can parse general text
Computing Science, University of Aberdeen
29