Interlingua-based MT
Interlingua-based Machine Translation
• Syntactic transfer-based MT
–
Couples the syntax of the two
languages
• What if we abstract away the
syntax
All that remains is meaning
– Meaning is the same across
languages
– Simplicity: Only N components
needed to translate among N
languages
Interlingua
Semantic
Interpretation
Syntactic
Structure
–
• Two “small” problems:
What is meaning?
– How do we represent meaning?
Semantic
Generation
Syntactic
Structure
Transfer-based
MT
Parsing
Source
Syntactic
Generation
Target
Direct MT
English
analyzer
Spanish
analyzer
Japanese
analyzer
–
Interlingual representation
English
generator
Spanish
Generator
Japanese
Generator
Example of Interlingua Machine Translation
Interlingua representation
need
I
Need( I , e1); e1  Make( I , e2); collect _ call(e2)
make
to
必要があります
(need)
call
a collect
私は
(I)
 Event: Need

Tense : present



 Agent: I



Event
:
Make




Tense : Infinitive







Agent: I

Them
e
:



call









Them e: attributes: collect   





Definitene
ss
:
indef


  


コールを
(call)
コレクト
(collect)
かける
(make)
Ingredients of a semantic representation
• language neutral
–
•
•
•
•
Syntactic variations should result is the same semantics
sense of a word
deep semantic role labels
scope of quantifiers, adverbials, adjectives
polarity information
Distinguish between
surface structure (syntactic structure) and
deep structure (semantic structure) of sentences.
Different forms of semantic representation:
logic formalisms
ontology / semantic representation languages
•
•
•
•
Case Frame Structures (Filmore)
Conceptual Dependy Theory (Schank)
Description Logic (DL) and similar KR languages
Ontologies
Text Meaning Representation
• Lexicon has two components
Syntactic part
– Semantic constraints part
–
• Given a sentence, the syntactic part analyzes the input
syntactically and the semantic constraints create semantic
expressions that can be evaluated.
• Ontology specifies the type hierarchy
Used for checking selectional restrictions
– Selectional restrictions used for word-sense disambiguation
–
• e.g. accident is an event; organization has humans
Constructing a Semantic Representation
General approach:
Start with surface structure derived from parser.
Map surface structure to semantic structure:



Use phrases as sub-structures.
Find concepts and representations for central phrases (e.g. VP,
NP, then PP)
Assign phrases to appropriate roles around central concepts
(e.g. bind PP into VP representation).
Semantic Representation
Semantic Representations are based on some form of (formal) Representation
Language.
•
Semantics Networks
•
Conceptual Dependency Graphs
•
Case Frames
•
Ontologies
•
DL and similar KR languages
Important note: Difference between relations between text strings and
referents in the world.
Ontology (Interlingua) approach
Ontology: a language-independent classification of objects, events, relations
A Semantic Lexicon, which connects lexical items to nodes (concepts) in
the ontology
An analyzer that constructs Interlingua representations and selects an
appropriate one
Semantic Lexicon
Provides a syntactic context for the appearance of the lexical item
Provides a mapping for the lexical item to a node in the ontology (or more
complex associations)
Provides connections from the syntactic context to semantic roles and
constraints on these roles
Constructing an InterLingua Representation
For each syntactic analysis:
Access all semantic mappings and contexts for each lexical item.
Create all possible semantic representations.
Test them for coherency of structure and content.
Basic Semantic Dependency - Example
Input: John makes tools
Syntactic Analysis:
cat
subject
object
verb
root
tense
make
present
root
cat
john
noun-proper
root
cat
number
tool
noun
plural
Lexicon Entries for John and tool
John-n1
syn-struc
root
cat
sem-struc
human
tool-n1
syn-struc
root
cat
sem-struc
tool
john
noun-proper
name
john
gender male
tool
n
Ontological Representation - Example
Relevant extract from the specification of the
ontological concept used to describe the
appropriate meaning of make:
manufacturing-activity
...
agent
human
theme
artifact
…
who
what
Semantic Dependency Component
The basic semantic dependency component of the “Text
Meaning Representation” (TMR) for:
John makes tools
manufacturing-activity-7
agent
human-3
theme
set-1
element
cardinality
…
tool
>1
semantic representation of “try-v3”
try-v3
syn-struc
root
cat
subj
xcomp
sem-struc
set-1
try
v
root
cat
root
cat
form
$var1
n
$var2
v
OR infinitive gerund
element-type refsem-1
cardinality >=1
refsem-1
sem
event
Means “non finished
agent ^$var1
action; outcome unclear”
effect refsem-2
modality
modality-type
epiteuctic
modality-scope
refsem-2
modality-value
< 1
refsem-2
value ^$var2
sem
event
REQUEST-INFO-130
THEME
TEXT-POINTER
INSTANCE-OF
DEVELOP-2601.PURPOSE DEVELOP-2601.REASON
why
REQUEST-INFO
DEVELOP-2601
THEME
AGENT
PHASE
TIME
INSTANCE-OF
TEXT-POINTER
SET-2555
NATION-97
CONTINUOUS
FIND-ANCHOR-TIME
DEVELOP
developing
NATION-97
HAS-NAME
INSTANCE-OF
TEXT-POINTER
Iraq
NATION
Iraq
SET-2555
ELEMENT-TYPE
CARDINALITY
INSTRUMENT-OF
THEME-OF
INSTANCE-OF
TEXT-POINTER
WEAPON
> 1
KILL-1864
DEVELOP-2601
WEAPON
weapons
KILL-1864
THEME
INSTRUMENT
INSTANCE-OF
TEXT-POINTER
SET-2556
SET-2555
KILL
destruction
SET-2556
THEME-OF
ELEMENT-TYPE
CARDINALITY
INSTANCE-OF
TEXT-POINTER
KILL-1225
HUMAN
> 100
HUMAN
mass
“Why is Iraq developing weapons
of mass destruction?”
Word sense Disambiguation
Methods

Constraint checking



make sure the constraints imposed on context are met
Graph traversal

is-a links are inexpensive

other links are more expensive

the “cheapest” structure is the most coherent one
Hunter-gatherer processing

find (hunt) and eliminate (kill) unlikely interpretations

collect (gather) remaining interpretations
Ontological Semantics: An example
semantic representation language
slides from S. Nirenberg
Ontological semantics is a computationally tractable
theory of meaning in natural language as well as a
suite (OntoSem) of implemented NLP programs and a set of
static knowledge resources that support these programs.
Ontological semantics deals directly with extraction,
representation and manipulation of text meaning.
Ontosem text analyzers produce interpreted knowledge ready
to be used in reasoning-heavy applications that include
question answering, cross-document and cross-lingual text
summarization, question answering, machine translation and
others.
Support of intelligent human-computer interaction in
domain- and task-oriented environments is squarely
within the purview of ontological semantics.
Ontological semantics concentrates on content
of representations and is adaptable to a number of
different representation formats.
Ontological semantics is both a producer and a
consumer of knowledge: deriving text meaning is
itself a knowledge-intensive task
OntoSem
• is devoted to processing naturally occurring texts
• strives for high-quality results first followed by concern
for broad coverage
• expects “unexpected” inputs
• seeks quality heuristics of any provenance (knowledgebased or probabilistic, cooccurrence-based)
• does not grant syntax a privileged position among
the providers of heuristics for semantic processing
• does not make a strong distinction between
semantics and pragmatics
• is applicable to any natural language
Ontological-semantic analyzers take natural
language texts as inputs and generate machinetractable text meaning representations (TMRs)
that form the basis of various reasoning processes.
Sample Input Sentence:
Iran, Iraq and North Korea on Wednesday rejected an
accusation by President Bush that they are developing
weapons of mass destruction.
The TMR (presented graphically) for the above is
as follows:
Output: A Text Meaning Representation (TMR)
This presentation is simplified; the system, in fact, derives much more from text;
event instances are shown in ellipses; object instances, in rectangles; only case
role and set membership relations are shown (as labels on links); numerical constraints
can be fuzzy, as in the cardinality of SET-1226.
Semantic Dependencies
(fillers of ontological
properties mentioned in
text; not simply relations
among textual strings)
DENY-1224
AGENT
THEME
TIME
INSTANCE-OF
TEXT-POINTER
;; speech act
SET-1224
DEVELOP-1224
< FIND-ANCHOR-TIME WEDNESDAY-1224
DENY
reject
ACCUSE-1224
AGENT
BENEFICIARY
THEME
INSTANCE-OF
TEXT-POINTER
;; President BushÕs accusation
HUMAN-15691
Instances of Ontological
Concepts
SET-1224
DEVELOP-1224
Many additional properties
ACCUSE
stored with concepts
accusation
underlying instances
Word Sense Disambiguation
DEVELOP-1224
;; developing weapons
THEME
SET-1225
THEME-OF
DENY-1224 ACCUSE-1224
AGENT
SET-1224
Triggers for further contextPHASE
CONTINUOUS
dependent processing
PURPOSE
WARN-1224
TIME
FIND-ANCHOR-TIME
A pretty-printed fragment of
INSTANCE-OF DEVELOP
the actual TMR representation
TEXT-POINTER
developing
for sample input
Ontological-semantic systems centrally rely on the following
static knowledge resources:



a language-independent ontology that includes
knowledge about types of entities in the world,
e.g., ATHLETE, WELD or SPEED;
ontology-oriented lexicons (and onomasticons,
or lexicons of proper names) for each natural
language in the system; and
a fact repository containing instances of
ontological concepts, e.g., Andre Agassi
(ATHLETE-3176) or the Apollo 13 mission
(SPACEFLIGHT-142)
A Sample Screen of the Ontology/Lexicon/Fact Repository Browsing and Editing Environment
(diagnosis
(diagnosis-n1
(cat n)
(anno
(def "")
(ex "The diagnosis (of cancer) (by the specialist) was made
quickly")
(comments ""))
(syn-struc
((root $var0) (cat n)
(pp-adjunct
((root of) (root $var1) (cat prep) (opt +)
(obj ((root $var2) (cat n)))))
(pp-adjunct
((root by) (root $var3) (cat prep) (opt +)
(obj ((root $var4) (cat n)))))))
(sem-struc
(DIAGNOSE
(agent (value ^$var4))
(theme (value ^$var2)))
(^$var1 (null-sem +))
(^$var3 (null-sem +))))
)
; diagnosis
; of
; disease
; by
; someone
; the ontological mapping
; the case roles
; blocks compositional analysis of preps
(cancer
(cancer-n1
(cat n)
(anno
(def "a disease")
(ex "")
(comments "")
)
(syn-struc
((n ((root $var1) (cat n) (opt +)))
; animal part as modifier
(root $var0) (cat n)
; cancer
))
(sem-struc
(CANCER
(location (value ^$var1) (sem animal-part)))
)
)
(cancer-n2
(cat n)
(anno
(def "a sign of the zodiac")
(ex "")
(comments "")
)
(syn-struc
((root $var0) (cat n)
))
(sem-struc
(CANCER-ZODIAC)
)
)
)
Currently Available Static Knowledge Sources for English:
• Ontology of about 6,500 concepts
(about 95,000 property-value pairs)
• English lexicon of about 40,000 entries
• Fact repository of about 20,000 facts (outside medical domain)
• English Onomasticon of about 350,000 entries
• Tokenization knowledge, morphological and syntactic grammars
for a number of languages
The analyzer’s conceptual architecture
Input
Text
(in reality, not strictly pipelined)
Preprocessor
Syntactic
Analyzer
TMR
Semantic
Analyzer
Processing Modules
Grammar:
Ecology
Morphology
Syntax
Static Knowledge Resources
Lexicon and
Onomasticon
Ontology and
Fact Repository
The basic (“who did what to whom”) semantic
dependency is derived, in the general case, on the
basis of
a) lexical-semantic expectations (selectional
restrictions) recorded in the ontology and the
lexicon and
b) syntactic dependency derived from the results of
syntactic analysis.
The beginnings of system evaluation
Sentences
Words
Senses
Words in / not in lexicon
Syntactic ambiguity count
Overall ambiguity count
WS disambiguation I
Semantic dependencies I
WS disambiguation II
Semantic dependencies II
WS disambiguation III
Semantic dependencies III
1
28
79
28/0
192
>1.7M
52%
67%
96%
69%
96%
85%
2
33
86
32/1
32
>149M
48%
33%
68%
50%
100%
100%
3
8
29
5/3
16
64
50%
17%
67%
63%
67%
63%
4
24
150
24/0
19
>199M
46%
40%
83%
33%
88%
90%
5
33
96
31/2
63
>418M
30%
33%
88%
69%
90%
100%
Run I: “raw”
Run II: preprocessor output correct;
Run III: preprocessor and syntactic analysis output correct
6
26
76
24/2
47
>268K
50%
29%
54%
29%
100%
86%
Average
25.33
86
24/1.33
61.5
>120M
48.0%
36.5%
76.0%
52.2%
90.2%
87.3%
In addition to the basic semantic dependency, TMRs also
include parameterized information provided by the microtheories
of aspect, modality (including speaker attitudes), time, style
and others.
Most of these microtheories have been implemented. All would
benefit from further work. We are also actively looking into
possibilities of borrowing some microtheories -- either in toto or
partially.
FrameNet: Another example of semantic representation
Frame Semantics
(Fillmore 1976, 1977, ..)
•
Frame: a conceptual structure or prototypical situation
•
Frame elements (roles)
Identify participants of the situation
– Are local to their frame
–
•
Frame evoking elements (verbs, nouns, adjectives) introduce frames
•
E.g. VERDICT:
[The jury]Judge convicted [him]Defentant [on the counts of theft]Charges.
On Thursday [a jury]Judge found [the youth]Defendant [guilty of wounding Mr Lay]
Finding
Berkeley FrameNet Project
•
Database of frames for core lexicon of English
•
Current release: 610 frames, about 9000 lexical units
Types of Relations
FrameNet Relations
•
Frame hierarchy: inherits
•
Subframes
Contextual Relations between instantiated frames and roles
•
Syntactic and/or semantic embedding
•
Discourse relations
•
Anaphoric relations
Inferences
•
On the basis of both
A Case Study
In the first trial in the world in connection with the terrorist attacks of 11
September 2001, the Higher Regional Court of Hamburg has passed down the
maximum sentence. Mounir al Motassadeq will spend 15 years in prison. The
28-year-old Moroccan was found guilty as an accessory to murder in more
than 3000 cases.
FrameNet „as a Net“
– Frame-to-Frame Relations –
Subframe relation
•
Super frame represents complex event
•
Subframes represent sub-events
•
Subframes usually inherit some roles of the super frame
Defendant
Court
Defense
Judge
Jury
Offense
Criminal
process
Charge
Prosecution
Defendant
Charge
Arraignment
...
...
Arrest
...
Sentencing
...
Trial
Local Roles
In the first trial in the world in connection with [the [terrorist]Assailant
attacks of [11 September 2001]Time]Case, [the Higher Regional Court of
Hamburg]Court has passed down the [maximum]Type sentence.
Local Roles
[Mounir al Motassadeq]Inmates will spend [15 years]Duration in prison.
Local Roles
[The 28-year-old Moroccan]Defendant was found [guilty]Finding as [an
accessory to [murder]FocalEntity [in more than 3000 cases]Victim ]Charge.
Unfilled Roles
Target
Frame
Frame roles
trial
TRIAL
CASE
attacks
sentence
(10)
prison
terrorist attacks
accessory to murder
(2)
COURT
Higher Regional Court
(3)
DEFENDANT ...
28-year-old Moroccan
(4)
ASSAILANT terrorist
VICTIM ...
(7)
SENTENCING
CONVICT
Higher Regional Court
found VERDICT
CASE
year-old Moroccan
CHARGE
(15)
murder
ASSISTANCE
KILLING
(5)
(6)
PRISON
INMATES ...
DURATION (exth.) 15 years
accessory
(1)
CHARGE
ATTACK
September 2001
Filler (given vs. Induced)
(9)
11
Mounir al Motassadeq
TYPE ...
(8)
COURT
maximum sentence
Mounir al Motassadeq
(12)
(11)
terrorist attacks
(13)
accessory to murder
FINDING ...
guilty
CO-AGENT
(14)
(17)
FOCAL_ENTITY
murder
HELPER ...
28-year-old Moroccan
KILLER
VICTIM ...
TIME (exth.)
(18)
(19)
(20)
m.t. 3000 cases
(21)
(16)
DEFENDANT 28-
Target
Frame
Frame roles
Filler (given vs. Induced)
trial
TRIAL
CASE
terrorist attacks
(1)
CHARGE
accessory to murder
(2)
COURT
Higher Regional Court
(3)
DEFENDANT ...
28-year-old Moroccan
(4)
ASSAILANT
terrorist
attacks
ATTACK
VICTIM ...
(6)
TIME (exth.)
sentence
SENTENCING
CONVICT
(5)
11 September 2001
Mounir al Motassadeq
(7)
(8)
COURT
Higher Regional Court
(9)
TYPE ...
maximum sentence
(10)
prison
PRISON
INMATES ...
Mounir al Motassadeq
DURATION (exth.) 15 years
(11)
(12)
Found
VERDICT
CASE
(13)
accessory
murder
ASSISTANCE
KILLING
terrorist attacks
CHARGE
accessory to murder
(14)
DEFENDANT
28-year-old Moroccan
(15)
FINDING ...
guilty
(16)
CO-AGENT
(17)
FOCAL_ENTITY
murder
(18)
HELPER ...
28-year-old Moroccan
(19)
KILLER
VICTIM ...
(20)
m.t. 3000 cases
(21)
Target
Frame
Frame roles
Filler (given vs. Induced)
trial
TRIAL
CASE
terrorist attacks
(1)
CHARGE
accessory to murder
(2)
COURT
Higher Regional Court
(3)
DEFENDANT ...
28-year-old Moroccan
(4)
ASSAILANT
terrorist
(5)
attacks
ATTACK
VICTIM ...
sentence
prison
found
accessory
murder
SENTENCING
PRISON
VERDICT
ASSISTANCE
KILLING
(6)
TIME (exth.)
11 September 2001
(7)
CONVICT
Mounir al Motassadeq
(8)
COURT
Higher Regional Court
(9)
TYPE ...
maximum sentence
(10)
INMATES ...
Mounir al Motassadeq
(11)
DURATION (exth.) 15 years
(12)
CASE
terrorist attacks
(13)
CHARGE
accessory to murder
(14)
DEFENDANT
28-year-old Moroccan
(15)
FINDING ...
guilty
(16)
CO-AGENT
(17)
FOCAL_ENTITY
murder
(18)
HELPER ...
28-year-old Moroccan
(19)
KILLER
VICTIM ...
(20)
m.t. 3000 cases
(21)
Linking Frames and Roles in Context
At the instance level
•
given frame instances f1:F1 and f2:F2, where
f1 and f2 stand in a contextual relation (syn, sem, discourse)
– frame types F1 and F2 stand in some frame relation
–
=> identify role instances (referents) of f1 and f2 (r1 (= r0) =
r2)
frame relation
context-related instances
inferred relation
Linking Frames and Roles in Context
In the first trial in the world in connection with the terrorist attacks of 11
September 2001, the Higher Regional Court of Hamburg has passed
down the maximum sentence.
Criminal
Process
Court
Sentencing
Trial
frame relation
Linking Frames and Roles in Context
In the first trial (f1) in the world in connection with the terrorist attacks of
11 September 2001, [the Higher Regional Court of Hamburg] (r2) has passed
down the maximum sentence (f2).
Criminal
Process
Court
Sentencing
Functional
Embedding
Trial
The Higher Regional Court of Hamburg
frame relation
context-related instances
Linking Frames and Roles in Context
In the first trial (f1) in the world in connection with the terrorist attacks of
11 September 2001, [the Higher Regional Court of Hamburg] (r2=r0= r1) has
passed down the maximum sentence (f2).
Criminal
Process
Court
Sentencing
Functional
Embedding
Trial
The Higher Regional Court of Hamburg
frame relation
context-related instances
inferred relation
Linking Frames and Roles in Context
At the type level (more involved)
•
If instances of frame roles f1:F1 and f2:F2 are often found coreferent within particular contextual relations
=> Hypothesize a frame relation between F1 and F2
(no) frame relation context-related instances
inferred relation
Linking Frames and Roles in Context
… the Higher Regional Court of Hamburg has passed down the
Maximum sentence. [Mounir al Motassadeq] will spend 15 years in prison.
Prison
• New Frame Relation
Sentencing
Inmates
Discourse
Relation
• (Role Binding: Convict=Inmates)
Convict
(Co-reference)
(no) frame relation context-related instances
inferred relations
Frame, Contextual, and Inferred Relations
CRIMINAL PROCESS
PRISON (2)
Inmates
Duration
SENTENCING (1)
Convict
Type
TRIAL (1)
Court Court Defendant
Charge Case
VERDICT (3)
Defendant Finding
(1)
sentence number
Subframe/FE
Contextual Relation
Inferred Relation
KILLING (3)
Victim
Killer
Charge Case
ASSISTANCE (3)
Helper Focal_entity
Co_agent
CRIMINAL PROCESS
PRISON
Inmates
(Motus.)
Duration
(15Y)
SENTENCING
TRIAL
Convict Duration
Court
(maximum) (Hmbg.)
Defendant
Charge Case
(9/11)
VERDICT
Defendant
(the Moroccan)
Charge
Case
(accessory)
ASSISTANCE
Hierarchy/Subframe/FE
Contextual Relations
Inference
KILLING
Victim
(3000)
Goal
(murder)
Helper
Co_agent
Killer
In the first trial .. the higher Regional Court .. has passed down the maximum sentence.
Mounir al Motussadeq will spend 15 years in prison.
The 28-year-old Moroccan was found guilty as an accessory to murder in .. 3000 cases.
Statistical Semantic Role Labeling
References
Jurafsky, D. & J. H. Martin, Speech and Language Processing, Prentice-Hall,
2000. (Chapters 9 and 10)
Helmreich, S., From Syntax to Semantics, Presentation in the 74.419 Course,
November 2003.
Nirenburg, S. & V. Raskin, Ontological Semantics, MIT Press, 2004.
Wordnet, http://wordnet.princeton.edu/
Suggested Upper Merged Ontology (SUMO),
http://www.ontologyportal.org/
Descargar

Interlingua