CS626-449: NLP, Speech and Web-Topics-in-AI
description
Transcript of CS626-449: NLP, Speech and Web-Topics-in-AI
CS626-449: NLP, Speech and Web-Topics-in-AI
Pushpak BhattacharyyaCSE Dept., IIT Bombay
Lecture 37: Semantic Role Extraction (obtaining Dependency Parse)
Vaquious Triangle
2
Analy
sis
Generation
Transfer Based(do deep semantic processBefore entering the target language)
Direct(enter the target Language immediatelyThrough a dictionary)
Interlingua based (do deep semantic processBefore entering the target language)
Vaquious: an eminentFrench Machine Translation Researcher-Originally a Physicist
3
Universal Networking Language Universal Words (UWs) Relations Attributes Knowledge Base
4
UNL Graph
objagt
@ entry @ past
minister(icl>person)
forward(icl>send)
mail(icl>collection)
He(icl>person)
@def
@def
gol
He forwarded the mail to the minister.
5
AGT / AOJ / OBJ AGT (Agent)
Definition: Agt defines a thing which initiates an action
AOJ (Thing with attribute)Definition: Aoj defines a thing which is in a state or has an attribute
OBJ (Affected thing)Definition: Obj defines a thing in focus which is directly affected by an event or state
6
Examples John broke the window.
agt ( break.@entry.@past, John)
This flower is beautiful.aoj ( beautiful.@entry, flower)
He blamed John for the accident.obj ( blame.@entry.@past, John)
7
BEN BEN (Beneficiary)
Definition: Ben defines a not directly related beneficiary or victim of an event or state
Can I do anything for you?ben ( do.@entry.@interrogation.@politeness, you )obj ( do.@entry.@interrogation.@politeness,
anything )agt (do.@entry.@interrogation.@politeness, I )
8
PUR PUR (Purpose or objective)
Definition: Pur defines the purpose or objectives of the agent of an event or the purpose of a thing exist
This budget is for food.pur ( food.@entry, budget )mod ( budget, this )
9
RSN
RSN (Reason)Definition: Rsn defines a reason why an event or a state happens
They selected him for his honesty.agt(select(icl>choose).@entry, they)obj(select(icl>choose) .@entry, he)rsn (select(icl>choose).@entry, honesty)
10
TIM TIM (Time)
Definition: Tim defines the time an event occurs or a state is true
I wake up at noon.agt ( wake up.@entry, I )tim ( wake up.@entry, noon(icl>time))
11
TMF TMF (Initial time)
Definition: Tmf defines a time an event starts
The meeting started from morning.obj ( start.@entry.@past, meeting.@def )tmf ( start.@entry.@past, morning(icl>time) )
12
TMT TMT (Final time)
Definition: Tmt defines a time an event ends
The meeting continued till evening.obj ( continue.@entry.@past, meeting.@def )tmt ( continue.@entry.@past,evening(icl>time) )
13
PLC PLC (Place)
Definition: Plc defines the place an event occurs or a state is true or a thing exists
He is very famous in India.aoj ( famous.@entry, he )man ( famous.@entry, very)plc ( famous.@entry, India)
14
PLF PLF (Initial place)
Definition: Plf defines the place an event begins or a state becomes true
Participants come from the whole world.
agt ( come.@entry, participant.@pl )plf ( come.@entry, world )mod ( world, whole)
15
PLT PLT (Final place)
Definition: Plt defines the place an event ends or a state becomes false
We will go to Delhi.agt ( go.@entry.@future, we )plt ( go.@entry.@future, Delhi)
16
INS INS (Instrument)
Definition: Ins defines the instrument to carry out an event
I solved it with computeragt ( solve.@entry.@past, I )ins ( solve.@entry.@past, computer )obj ( solve.@entry.@past, it )
17
Attributes Constitute syntax of UNL Play the role of bridging the conceptual world
and the real world in the UNL expressions Show how and when the speaker views what is
said and with what intention, feeling, and so on Seven types:
Time with respect to the speaker Aspects Speaker’s view of reference Speaker’s emphasis, focus, topic, etc. Convention Speaker’s attitudes Speaker’s feelings and viewpoints
18
Tense: @past
The past tense is normally expressed by @past
{unl}agt(go.@entry.@past, he)…{/unl}
He went there yesterday
19
Aspects: @progress
{unl}man
( rain.@entry.@present.@progress, hard )
{/unl}
It’s raining hard.
20
Speaker’s view of reference
@def (Specific concept (already referred))
The house on the corner is for sale. @indef (Non-specific class)
There is a book on the desk @not is always attached to the UW
which is negated.He didn’t come.
agt ( come.@entry.@past.@not, he )
21
Speaker’s emphasis @emphasis
John his name is.mod ( name, he )aoj ( John.@emphasis.@entry, name )
@entry denotes the entry point or main UW of an UNL expression
22
Subcategorization Frames Specify the categorial class of the lexical
item. Specify the environment. Examples:
kick: [V; _ NP]cry: [V; _ ] rely: [V; _PP] put: [V; _ NP PP]think: : [V; _ S` ]
23
Subcategorization Rules
V y /_NP]_ ]_PP]_NP PP]_S`]
Subcategorization Rule:
24
Subcategorization Rules1. S NP VP
2. VP V (NP) (PP) (S`)…3. NP Det N4. V rely / _PP]5. P on / _NP]6. Det the7. N boy, friend
The boy relied on the friend.
25
Semantically Odd Constructions Can we exclude these two ill-
formed structures ? *The boy frightened sincerity. *Sincerity kicked the boy.
Selectional Restrictions
26
Selectional Restrictions Inherent Properties of Nouns:
[+/- ABSTRACT], [+/- ANIMATE]
E.g., Sincerity [+ ABSTRACT]Boy [+ANIMATE]
27
Selectional Rules A selectional rule specifies certain selectional
restrictions associated with a verb.
V y /[+/-ABSTARCT][+/-
ANIMATE]
V frighten
/ [+/-ABSTARCT][+ANIMATE
]
____
____
28
Subcategorization FrameforwardV__ NP PP
invitationN__ PP
accessibleA__ PP
e.g., An invitation to the party
e.g., A program making science is more accessible to young people
e.g., We will be forwarding our new catalogue to you
29
Thematic RolesThe man forwarded the mail to the minister.
forward
V__ NP PP
Event FORWARD [Thing THE MAN], [Thing THE MAIL],
[Path TO THE MINISTER]
()
30
How to define the UWs in UNL Knowledge-Base?
Nominal concept Abstract Concrete
Verbal concept Do Occur Be
Adjective concept Adverbial concept
31
Nominal Concept: Abstract thing
abstract thing{(icl>thing)}culture(icl>abstract thing)civilization(icl>culture{>abstract thing})direction(icl>abstract thing)east(icl>direction{>abstract thing})duty(icl>abstract thing)mission(icl>duty{>abstract thing})responsibility(icl>duty{>abstract thing})accountability{(icl>responsibility>duty)}event(icl>abstract thing{,icl>time>abstract thing}) meeting(icl>event{>abstract thing,icl>group>abstract thing})conference(icl>meeting{>event}) TV conference{(icl>conference>meeting)}
32
Nominal Concept: Concrete thing
concrete thing{(icl>thing,icl>place>thing)}building(icl>concrete thing)factory(icl>building{>concrete thing})house(icl>building{>concrete thing})substance(icl>concrete thing)cloth(icl>substance{>concrete thing})cotton(icl>cloth{>substance})fiber(icl>substance{>concrete thing})synthetic fiber{(icl>fiber>substance)}
textile fiber{(icl>fiber>substance)}liquid(icl>substance{>concrete thing})
beverage(icl>food,icl>liquid>substance}) coffee(icl>beverage{>food}) liquor(icl>beverage{>food})
beer(icl>liquor{>beverage})
33
Verbal concept: do
do({icl>do,}agt>thing,gol>thing,obj>thing)express({icl>do(}agt>thing,gol>thing,obj>thing{)})
state(icl>express(agt>thing,gol>thing,obj>thing))explain(icl>state(agt>thing,gol>thing,obj>thing))
add({icl>do(}agt>thing,gol>thing,obj>thing{)})change({icl>do(}agt>thing,gol>thing,obj>thing{)})convert(icl>change(agt>thing,gol>thing,obj>thing)classify({icl>do(}agt>thing,gol>thing,obj>thing{)})divide(icl>classify(agt>thing,gol>thing,obj>thing))
34
Verbal concept: occur and be occur({icl>occur,}gol>thing,obj>thing)
melt({icl>occur(}gol>thing,obj>thing{)})divide({icl>occur(}gol>thing,obj>thing{)})arrive({icl>occur(}obj>thing{)})
be({icl>be,}aoj>thing{,^obj>thing}) exist({icl>be(}aoj>thing{)})born({icl>be(}aoj>thing{)})
35
How to define the UWs in UNL Knowledge Base?
In order to distinguish among the verb classes headed by 'do', 'occur' and 'be', the following features are used:
UW[ need an agent ]
[ need an object ]
English
'do' + + "to kill"'occur' - + "to fall"
'be' - - "to know"
36
The verbal UWs (do, occur, be) also take some pre-defined semantic cases, as follows:
How to define the UWs in UNL Knowledge-Base?
UW PRE-DEFINED CASES
English
'do' takes necessarily agt>thing
"to kill"
'occur' takes necessarily obj>thing
"to fall"
'be' takes necessarily aoj>thing
"to know"
37
Complex sentenceI want to watch this movie.
movie(icl>)
want (icl>)@entry.@pa
stob
j
@def
:01
I (iof>person)
watch (icl>do)@entry.@inf
objag
tag
tI (iof>person)
38
Approach to UNL Generation
Problem Definition Generate UNL expressions for English
sentences in a robust and scalable manner, using syntactic analysis and lexical
resources extensively. This needs
detecting semantically relatable entities and solving attachment problems
Semantically Relatable Sequences (SRS)Definition: A semantically relatable
Sequence (SRS) of a sentence is a group of words in the sentence (not necessarily consecutive) that appear in the semantic graph of the sentence as linked nodes or nodes with speech act labels
(This is motivated by UNL representation)
SRS as an intermediary to and intermediary
SourceLanguageSentence
TargetLanguageSentenceSRS UNL
Example to illustrate SRS
“The man bought a
new car in June” in: modifiera: indefinite
the: definite
man
past tense
agent
bought
objecttime
car
new
June
modifier
Sequences from “the man bought a new car in June”
a. {man, bought}b. {bought, car}c. {bought, in, June}d. {new, car}e. {the, man}f. {a, car}
Basic questions
Which words can form semantic constituents, which we call Semantically Relatable Sequences (SRS)?
What after all are the SRSs of the given sentence?
What semantic relations can link the words in an SRS and the SRSs themselves?
Postulate
A sentence needs to be broken into Sequences of at most three forms {CW, CW} {CW, FW, CW} {FW, CW}
where CW refers to content word or a clause and FW to function word
SRS and Language Phenomena
Movement: Preposition Stranding John, we laughed at.
(we , laughed.@entry)---------(CW, CW)
(laughed.@entry,at, John)---(CW, FW, CW)
Movement: Topicalization The problem, we solved.
(we , solved.@entry)------------(CW, CW)
(solved.@entry , problem)-----(CW,CW)
(the, problem)--------------------(CW,CW)
Movement: Relative Clauses John told a joke which we had already heard.
(John, told.@entry) -------------------(CW, CW) (told.@entry, :01) ---------------------(CW,CW) SCOPE01(we,had,heard.@entry)-------(CW,
FW,CW) SCOPE01(already,heard.@entry)-------(CW,CW) SCOPE01(heard@entry,which,joke)----
(CW,FW,CW) SCOPE01(a, joke)--------------------------(FW,CW)
Movement: Interrogatives Who did you refer her to?
(did , refer.@entry.@interrogative)-------(FW,CW)
(you, refer.@entry.@interrogative)--------(CW,CW)
(refer.@entry.@interrogative , her)--------(CW,CW)
(refer.@entry.@interrogative , to,who)----
(CW,FW,CW)
Empty Pronominals: to-infinitivals Bill was wise to sell the piano.
(wise.@entry , SCOPE01)---------------(CW,CW) SCOPE01(sell.@entry , piano)---------(CW,CW) (Bill, was, wise.@entry) -----------------(CW,
FW,CW) SCOPE01(Bill, to, sell.@entry)---------(CW,
FW,CW) SCOPE01(the, piano) --------------------(FW,CW)
Empty pronominal: Gerundial The cat leapt down spotting a thrush on the lawn. (The, cat) -------------------------------(FW, CW) (cat, leapt.@entry) --------------------(CW, CW) (leapt.@entry , down) ----------------(CW, CW) (leapt.@entry , SCOPE01) -----------------(CW, CW) SCOPE01(spotting.@entry,thrush)--------(CW,CW) SCOPE01(spotting.@entry,on,lawn)---(CW,FW,CW)
PP Attachment John cracked the glass with a stone.
(John, cracked.@entry)--------------(CW,CW) (cracked.@entry, glass)-------------(CW,CW) (cracked.@entry, with, stone)----(CW,FW,CW) (a, stone)------------------------------(FW,CW) (the,glass)-------------------------(FW,CW)
SRS and PP attachment (Mohanty, Almeida, Bhattacharyya, 04)
Conditions Sub-conditions Attachment Point
[PP] is subcategorized by the verb [V]
[NP2] is licensed by a preposition [P]
[NP2] is attached to the verb [V] (e.g., He forwarded the mail to the minister)
[PP] is subcategorized by the noun in [NP1]
[NP2] is licensed by a preposition [P]
[NP2] is attached to the noun in [NP1](e.g., John published six articles on machine translation )
[PP] is neither subcategorized by the verb [V] nor by the noun in [NP1]
[NP2] refers to [PLACE] / [TIME] feature
[NP2] is attached to the verb [V](e.g., I saw Mary in her office; The girls met the teacher on different days)
Linguistic Study to Computation
Syntactic constituents to Semantic constituents
A probabilistic parser (Charniak, 04) is used.
Other resources: Wordnet and Oxford Advanced Learner’s Dictionary
In a parse tree, tags give indications of CW and FW: NP, VP, ADJP and ADVP CW PP (prepositional phrase), IN
(preposition) and DT (determiner) FW
Observation: Headwords of sibling nodes form SRSs
“John has bought
a car.”
SRS:{has, bought}, {a, car}, {bought, car} a
(C) VP bought
(F) AUX has (C) VP bought
(C) VBD bought (C) NP car
(F) DT a (C) NN car
bought
car
has
Need: Resilience to wrong PP attachment
“John has published an
article on linguistics” Use PP attachment heuristics Get
{article, on, linguistics}
on linguistics
(C)VP published
(F) PP on(C)VBD published (C)NP article
published
(F)DT an
an
(C)NNarticle
(F)IN on
article
(C)NNS linguistics
(C)NPlinguistics
to-infinitival“I forced him to watch this movie” Clause boundary is the VP node, labeled with SCOPE
Tag is modified to TO, a FW tag, indicating that it heads a to-infinitival clause,
The duplication and insertion of the NP node with head him (depicted by shaded nodes) as a sibling of the VBD node with head forced is done to bring
out the existence of a semantic relation between force and
him.
(C)VP watch
(C)VBD forced (C)NP him(C) S SCOPE
(F)TO toto
(C)VP forced
to
forced
(C)VP
(C)PRP him
him
(C)NP him
him
(C)PRP him
Linking of clauses: “John said that he was reading a novel” Head of S node marked as Scope SRS: {said, that, SCOPE}.
Adverbial clauses have similar parse tree structures except that the subordinating conjunctions are different from that.
(C)VBD said (F) SBAR that
(C) VP said
(F) IN that(C) S SCOPE
said that
Implementation Block Diagram of the system
Parse Tree
Charniak Parser
Scope Handler
Attachment Resolver
WordNet 2.0
Sub-categorization Database
Input Sentence
Parse Tree modification and augmentation with head and scope
information
AugmentedParse Tree
Semantically Related Sequences
Noun classification
Semantically Relatable Sequences Generator
THAT clause as Subcat property
Preposition as Subcat property
Time and Place features
Head determination Uses a bottom-up strategy to determine the
headword for every node in the parse tree. Crucial in obtaining the SRSs, since wrong
head information may end up getting propagated all the way up the tree
Processes the children of every node starting from the rightmost child and checks the head information already specified against the node’s tag to determine the head of the node
Some special cases are: SBAR node A VP node with PRO insertion, copula, Phrasal verbs
etc. NP nodes with of-PP cases and conjunctions under
them, which lead to scope creation.
Scope handler Performs modification on the parse
trees by insertion of nodes in to-infinitival cases
Adjusts of the tag and head information in case of SBAR nodes
Attachment resolver
Takes a (CW1, FW, CW2) as input and checks the time and place features of CW2, the noun class of CW1 and the subcategorization information for the CW1 and
FW pair to decide the attachment. If none of these yield any deterministic
results, take the attachment indicated by the parser
SRS generator Performs a breadth-first search on the
parse tree and performs detailed processing at every node N1 of the tree.
S nodes which dominate entire clauses (main or embedded) are treated as CWs.
SBAR and TO nodes are treated as FWs.
AlgorithmAlgorithmIf the node N1 is a CW (new/JJ,
published/VBD, fact/NN, boy/NN, John/NNP) perform the following checks:
If the sibling N2 of N 1 is a CW (car/NN, article/NN, SCOPE/S)
Then create {CW,CW} ({new, car}, {published, article}, {boy, SCOPE})
If the sibling N2 is a FW (in/PP, that/SBAR, and/CC)
Then, check if N2 has a child FW, N3 (in/IN, that/IN) and a child CW, N4 (June/NN, SCOPE/S)
If yes,Then use attachment resolver to decide
the CW to which N3 and N4 attach.Create{CW,FW,CW} ({published, in,
June}, {fact, that, SCOPE})If no,
Then check if next sibling N5 of N 1 is a CW (Mary/NN)
If yes,Create {CW,FW,CW} ({John, and, Mary})If the node N1 is a FW (the/DT, is/AUX,
to/TO), perform the following checks: If the parent node is a CW (boy/NP,
famous/VP)Check if sibling is an adjective.i. If yes, (famous/JJ)Then, create {CW,FW,CW} ({She, is,
famous})ii. If no, (boy/NN)Then, create {FW,CW} ({the, boy}, {has,
bought})If the parent node N6 is a FW (to/TO) and
the sibling node N7 is a CW (learn/VB)Use attachment resolver to decide on the
preceding CW to which N6 and N7 can attach.
Create {CW,FW,CW} ({exciting, to, learn})
Evaluation FrameNet corpus [Baker et. al., 1998], a
semantically annotated corpus, as the testdata.
92310 sentences (call this the gold standard) Created automatically from the FrameNet
corpus taking verbs, nouns and adjectives as the targets Verbs as the target- 37,984 (i.e., semantic frames
of verbs) Nouns as the target-37,240 Adjectives as the target-17,086
Score for high frequency verbsVerb Frequency ScoreSwim 280 0.709Depend 215 0.804Look 187 0.835Roll 173 0.7Rush 172 0.775Phone 162 0.695Reproduce 159 0.797Step 159 0.795Urge 157 0.765Avoid 152 0.789
Scores of 10 verb groups of high frequency in the Gold Standard
Scores of 10 noun groups of high frequency in the Gold Standard
An actual sentence A. Sentence : A form of asbestos
once used to make Kent cigarette filters has caused a high percentage of cancer deaths among a group of workers exposed to it more than 30 years ago, researchers reported.
Relative performance on SRS constructs
0 20 40 60 80 100
Total SRSs
(FW,CW)
(CW,FW,CW)
(CW,CW)
Para
met
ers
mat
ched
Recall/Precision
Recall
Precision
Results on sentence constructs
0 20 40 60 80 100
To-infinitival clause resolution
Complement-clause resolution
Clause linkings
PP Resolution
Para
met
er
Recall/Precision
Recall
Precision
Rajat Mohanty, Anupama Dutta and Pushpak Bhattacharyya, Semantically Relatable Sets: Building Blocks for Repesenting Semantics, 10th Machine Translation Summit ( MT Summit 05), Phuket, September, 2005.
Statistical Approach
Use SRL marked corpora Daniel Gildea and Daniel Jurafsky. 2002. Automatic labeling of
semantic roles. Computational Linguistics, 28(3):245–288. PropBank corpus
Role annotated WSJ part of Penn Treebank [10] PropBank role-set [2,4]
Core roles: ARG0 (Proto-agent), ARG1 (Proto-patient) to ARG5 Adjunctive roles:
ARGM-LOC (for locatives), ARGM-TMP (for temporals), etc.
SRL marked corpora contd… PropBank roles: an example
[ARG0 It] operates] [ARG1 stores] [ARGM−LOC mostly in Iowa and Nebraska]
Preprocessing systems [2] Part of speech tagger Base Chunker Full syntactic parser Named entities recognizer
Fig.4: Parse tree output, Source: [5]
Probabilistic estimation [1] Empirical probability estimation over candidate roles for each
constituent based upon extracted features
here,t is the target wordr is a candidate role,h , pt, gov, voice are features
Linear interpolation, with condition
• Geometric mean, with condition
),,,,,(#),,,,,,(#),,,,,|(
tvoicepositiongovpthtvoicepositiongovpthrtvoicepositiongovpthrP
),,|()|(),,|(),|()|()|( 54321 tpthrPhrPtgovptrPtptrPtrPtconstituenrP
)},,|()|(),,|(),|()|(exp{1)|( 54321 tpthrPhrPtgovptrPtptrPtrPz
tconstituenrP
1)|( r tconstituenrP
1i i
A state-of-art SRL system: ASSERT [4]
Main points [3,4] Use of Support Vector Machine [13] as classifier Similar to FrameNet “domains”, “Predicate Clusters” are introduced Named Entities [14] is used as a new feature
Experiment I (Parser dependency testing) Use of PropBank bracketed corpus Use of Charniak parser trained on Penn Treebank corpusParse Task Precision (%) Recall (%) F-score (%) Accuracy (%)
TreebankId. 97.5 96.1 96.8 -
Class. - - - 93.0
Id. + Class. 91.8 90.5 91.2 -
CharniakId. 87.8 84.1 85.9 -
Class. - - - 92.0
Id. + Class. 81.7 78.4 80.0 -
Table 1: Performance of ASSERT for Treebank and Charniak parser outputs.Id. Stands for identification task and Class. stands for classification task. Data source: [4]
Experiments and Results Experiment II (Cross genre testing)
1. Training on PropBanked WSJ data and testing on Brown Corpus2. Charniak parser trained on first PropBank then Brown
Table 2: Performance of ASSERT for various experimental combinations Date source: [4]