lecture13.pptx
LECTURE 13
Semantc Role Labelling
Arkaitz Zubiaga, 19
th
February, 2018
2
What is Semantc Role Labelling?
Thematc Roles:
FrameNet.
PropBank.
Approaches to Semantc Role Labelling.
Selectonal Restrictons.
LECTURE 13: CONTENTS
3
Semanti ole labelling (SRL): for a predicate or verb in a
sentence, task of identfying thematc roles, such as agent, goal,
or result.
SEMANTIC ROLE LABELLING
Theme LocationPredicateAgent
4
SRL can be useful for:
Queston anseering.
Machine translaton.
Informaton extracton.
To identfy sentences eith identcal meaning:
XYZ corporaton bought the stock.
They sold the stock to XYZ corporaton.
The stock eas bought by XYZ corporaton.
SEMANTIC ROLE LABELLING
5
Given a sentence:
Paul bought a car from Sarah for £7,000
Identfy predicaters) [i.e. the acton, generally the verb]:
Paul bought a car from Sarah for £7,000
Identfy thematc roles:
Paul bought a car from Sarah for £7,000
SEMANTIC ROLE LABELLING: EXAMPLE
buyer seller
goods
money
6
A typical set of thematc roles.
THEMATIC ROLES
7
Most f equently used themati oles:
AGENT: subject ractve sentence), by X rpassive sentence).
EXPERIENCER: animate subject in actve sentences.
THEME: object rtransitve verb), subject rnon-acton verb).
INSTRUMENT: eith X.
BENEFICIARY: for X.
THEMATIC ROLES
8
P oblem of themati oles:
Ha d to i eate standa d set of roles or formally defne them.
Alte native to thematc roles:
P opBank, feeer, more generalised semantc roles.
F ameNet, more roles, specifc to a group of predicates.
THEMATIC ROLES
9
Includes 12 domains rbody, cogniton, health, tme,…)
Conduited on f ame-by-f ame basis :
1. choose semantc frame re.g. Commerce_buy).
2. defne the frame and its frame elements re.g. BUYER,
GOODS, SELLER, MONEY).
3. list predicates that evoke it rbuy.v, purchase.v, purchase.n).
4. extract sentences for each predicate from Britsh Natonal
Corpus.
THE FRAMENET PROJECT
10
F ame Relaton: directed elaton beteeen teo f ames, i.e. a
Supe _F ame rless dependent, more abstract) and a Sub_F ame
rthe more dependent, less abstract).
e.g. the frame Commercial_transacton has teo sub-frames:
Commerce_goods_transfer
Commerce_money_transfer
FRAMENET: FRAME RELATIONS
11
One million eo d io pus annotated eith predicate argument
structures.
Developed at University of Pennsilvania.
P ediiates are lexicalised only by ive bs.
Adjunitive a guments : Locatve, Temporal, Manner, Cause, etc.
PROPOSITION BANK: PROPBANK
12
A g0 = agent
A g1 = direct object / theme / patent
A g2 = indirect object / benefactve / instrument / atribute / end
state
A g3 = start point / benefactve / instrument / atribute
A g4 = end point
PROPBANK: ARGUMENT NUMBERS
13
PROPBANK: ADJUNCTS
14
The company bought a eheel-loader from Dresser.
Arg0: The company
Rel: bought
Arg1: a eheel-loader
Arg2-from: Dresser
PROPBANK: EXAMPLE
15
TV statons bought “Cosby” reruns for record prices.
Arg0: TV statons
Rel: bought
Arg1: “Cosby” reruns
Arg3-for: record prices
PROPBANK: EXAMPLE
16
BUY SELL PAY
Arg0: buyer Arg0: seller Arg0: buyer
Arg1: commodity Arg1: commodity Arg1: price paid
Arg2: seller Arg2: buyer Arg2: seller
Arg3: price Arg3: price Arg3: commodity
Arg4: benefciary Arg4: benefciary Arg4: benefciary
PROPBANK: FLEXIBILITY
17
FRAMENET VS PROPBANK
18
Ea ly SRL systems r1970s-80s):
parser folloeed by hand-eriten rules for each verb.
dictonaries eith verb-specifc case frames.
Cu ent systems use maihine lea ning.
APPROACHES TO SEMANTIC ROLE LABELLING
19
A SIMPLE MODERN ALGORITHM
20
Uses parsed tree as input.
SEMANTIC ROLE LABELLING: EXAMPLE
21
1. P uning: use simple heuristcs to prune unlikely consttuents.
2. Identfiaton: a binary classifcaton of each node as an
argument to be labelled or a NONE.
3. Classifiaton: a 1-of-N classifcaton of all the consttuents that
eere labelled as arguments by the previous stage.
3-STEP VERSION OF SRL ALGORITHM
22
Imbalanie beteeen:
positve samples rconsttuents that are arguments of
predicate).
negatve samples rconsttuents that are not arguments of
predicate).
Imbalanied data ian be ha d for many classifers.
So ee p une the ive y unlikely ionsttuents f st , and then use a
classifer to get rid of the rest.
WHY PRUNING AND IDENTIFICATION STEPS?
23
Algo ithm so fa ilassifes eive ything loially – each decision
about a consttuent is independent of all others.
But this ian’t be ight: Lots of global or joint interactons
beteeen arguments.
A local system may incorrectly label teo overlapping
consttuents as arguments.
e.g. labelling one consttuent ARG0 should increase the
probability of another being ARG1
FINAL STAGE: JOINT INFERENCE
24
Re anking:
The frst stage SRL system produces multple possible labels
for each consttuent.
The second stage classifes the best global label for all
consttuents.
Ofen a classifer that takes all the inputs along eith other
features rsequences of labels).
HOW TO DO JOINT INFERENCE
25
SRL: NOT JUST ENGLISH
26
SEMANTIC ROLE LABELLING: EXAMPLE
27
Prune the predicate and all other elements that include it.
SRL EXAMPLE: PRUNING
28
Binary classifcaton of each element: argument or not.
SRL EXAMPLE: ARGUMENT IDENTIFICATION
29
SRL EXAMPLE: ARGUMENT IDENTIFICATION
30
Multclass classifcaton of argument type.
SRL EXAMPLE: ARGUMENT CLASSIFICATION
31
SRL EXAMPLE: ARGUMENT CLASSIFICATION
32
Features to classify arguments:
Governing predicate, e.g. issued.
Phrase type that it is, e.g. NP.
Path from argument to predicate.
Positon, i.e. before or afer predicate.
Voice, i.e. actve or passive.
Head eord of the consttuent.
SEMANTIC ROLE LABELLING: FEATURES
33
Content eord: lexicalised feature that selects an informatve
eord from the consttuent, other than the head.
POS tag of the content eord.
POS tag of the head eord.
Named Entty class of the content eord rlocaton, person,
organisaton,…)
Contains named enttes? Boolean feature.
SEMANTIC ROLE LABELLING: MORE FEATURES
34
First eord of the consttuent rand its POS tag).
Last eord of the consttuent rand its POS tag).
Lef and right sibling consttuent labels.
Lef and right sibling head eords rand their POS tags).
SEMANTIC ROLE LABELLING: MORE FEATURES
35
SRL: A level of shalloe semantcs for representng events and
their partcipants.
Teo common architectures, for various languages:
F ameNet: frame-specifc roles.
P opBank: Proto-roles.
Current systems extract by:
Parsing sentence.
Finding predicates, classifying associated arguments.
SEMANTIC ROLE LABELLING: SUMMARY
36
Consider the teo interpretatons of:
I eant to eat someehere nearby.
1. I eant to eat in a nearby place.
2. It is a place nearby the object that I eant to eat.
Difficult for an algorithm if ee expect the folloeing patern:
eat + [food], e.g. eat chicken
SELECTIONAL RESTRICTIONS
37
Selectonal restrictons are associated eith senses:
The restaurant se ives pizzas. [serves + food]
Britsh Aireays se ives London. [serves + place]
SELECTIONAL RESTRICTIONS
38
Common practce for selectonal restrictons is to use thesauri
such as WordNet.
e.g. The THEME of “eat” must be WordNet synset {food, nutrient}
SELECTIONAL RESTRICTIONS
39
RESOURCES
FrameNet:
http://framenet.icsi.berkeley.edu/
PropBank:
http://www.cs.rochester.edu/~gildea/PropBank/Sort/
NomBank:
http://nlp.cs.nyu.edu/meyers/NomBank.html
40
RESOURCES: SOFTWARE
SENNA:
https://ronan.collobert.com/senna/
SEMAFOR:
http://www.cs.cmu.edu/~ark/SEMAFOR/
MatePlus:
https://github.com/microth/mateplus
Open Sesame:
https://github.com/Noahs-ARK/open-sesame
41
ASSOCIATED READING
Jurafsky, Daniel, and James H. Martin. 2009. Speech and Language
Processing: An Introduction to Natural Language Processing, Speech
Recognition, and Computational Linguistics. 3rd edition. Chapter 22.