Semantic Roles

Semantic Roles
How the arguments of a predicate map to functional elements of the event the predicate
is about

I The idea goes all the way back to Panini (Pān.ini circa 350BC)
I Donald Davidson’s event representation for logical form
I Postulate an event, e
I Assert the type of e via a unary predicate
I crossing(e)
I Assert e’s attribute values via binary predicate named after the
attribute with its second argument being the value
I agent(e, John), patient(e, EnglishChannel)
I Thematic roles 6= semantic roles
I Express important arguments of a predicate
I As a potential terminological confusion, theme is just one of
many thematic roles

Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 243

 Semantic Roles

Major Thematic Roles in the Literature

Not a fixed set
Th. Role Definition Sample Words
Agent Volitional causer (includes acci- Kick
dents)
Experiencer One who experiences it Has (a feeling)
Force Nonvolitional causer Tsunami
Theme One (most) directly affected Shut (the door)
Result Outcome Wrote (a book)
Content Proposition of a propositional Asked
event
Instrument With a screwdriver
Beneficiary For his son
Source Origin of the object in a transfer Shipped
event
Goal Destination of the object in a Delivered
transfer event

Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 244

 Semantic Roles

Thematic Roles Exercise

For each thematic role, state an example sentence that illustrates it

Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 245

 Semantic Roles

Thematic Grid or Case Frame or θ -Grid of a Verb

The set of thematic roles that the verb takes on

I Constraints on how a verb’s thematic roles are presented and ordered

John broke the window
agent theme
John broke the window with a rock
agent theme instrument
The rock broke the window
instrument theme
The window broke
theme
The window was broken by John
theme agent

Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 246

 Semantic Roles

Diathesis Alternation or Verb Alternation

Multiple alternative mappings from arguments to syntactic positions

I For break (previous page)

Subject Object Preposition (With)
Phrase
agent theme
agent theme instrument
instrument theme
theme
I For give, dative alternation
Doris gave the book to Edward
agent theme goal
Doris gave Edward the book
agent goal theme

Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 247

 Semantic Roles

VerbNet
Gathers knowledge about verbs

I Class hierarchy of verbs that maps out what alternations each verb
participates in
I Verbs that support the dative alternation
I Verbs of future having: advance, allocate, offer, owe
I Verbs of sending: forward, hand, mail
I Verbs of throwing: kick, pass, throw
I Levin’s classification
I 47 high-level classes
I 193 low-level classes
I 3,100 verbs

Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 248

 Semantic Roles

Problems with Thematic Roles

Despite their intuitive appeal, . . .

I Difficult to standardize on set of thematic roles

I Difficult to formally specify
I Frequent need to refine (fragment) the roles
I Example: instrument seems to be two subroles
I This alternation works for intermediate instrument
The cook opened the jar with the new gadget
The new gadget opened the jar
I But not for enabling instrument
The cook ate noodles with a fork
*A fork ate the noodles
I How about this?
The cook whisked the eggs with a fork
A fork whisked the eggs

Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 249

 Semantic Roles

PropBank: Proposition Bank

Labels of (English and Chinese) sentences in Penn Treebank with semantic roles

I Semantic roles are defined specific to verb senses, not universally

I Not given meaningful names (helps avoid unnecessary controversy, I
assume)
I Some generalizations
I Arg0: proto-agent
I Arg1: proto-patient
I Arg2: often benefactive, instrument, attribute, or end
state
I Arg3: often benefactive, instrument, attribute, or
starting point
I Arg4: often end point
I Helps recover shallow semantic information from arguments of verbs

Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 250

 Semantic Roles

PropBank Frame File Example: Agree.01

I Arg0: Agreer ≈ Agent

I Arg1: Proposition being agreed to ≈ Content
I Arg2: With whom (if any) ≈ Beneficiary
[Arg0 The group] agreed [Arg1 it wouldn’t make an offer]
[ArgM-TMP Usually] [Arg0 John] agrees [Arg2 with Mary] [Arg1 on everything]

Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 251

 Semantic Roles

PropBank Frame File Example: Fall.01

I Arg0: Not defined since the normal subject of fall is proto-patient

I Arg1: Thing falling, which is the logical subject and patient
I Arg2: Extent, amount fallen
I Arg3: Start point
I Arg4: End point, end state of Arg1
[Arg1 Sales] fell [Arg4 to $25 million] [Arg3 from $27 million]
[Arg1 The average junk bond] fell [Arg2 by 4.2%]

Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 252

 Semantic Roles

PropBank Frame File Example: Increase.01

Extracting shallow semantic information from verb arguments

I Arg0: Causer of increase

I Arg1: Thing increasing
I Arg2: Amount increased by; or, manner
I Arg3: Start point
I Arg4: End point
Below, Dole is the agent and the price of Bananas is the theme
[Arg0 Dole] increased [Arg1 the price of Bananas]
[Arg1 The price of Bananas] was increased by [Arg0 Dole]
[Arg1 The price of Bananas] increased [Arg2 5%]

Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 253

 Semantic Roles

PropBank Modifiers and Adjuncts, Named ArgM-X

https://verbs.colorado.edu/∼mpalmer/projects/ace/PBguidelines.pdf
Name Definition Example
DIR: Directional To or from He smiled at her
LOC: Locative Where He added an amount to the penalty
MNR: Manner How She sang happily
TMP: Temporal When Now, recently
EXT: Extent How much AA raised fares as much as UA did
REC: Reciprocal Reflexives themselves, each other
PRD: Secondary Resultative, ate the fish raw
predication depictive
PNC: Purpose Because I left early to catch my flight
CAU: Causative Why, because Delayed because of weather
DIS: Discourse However, and And, that’s how it ends
(at beginning)
ADV: Adverbial On sentence Happily, she sang (cf. above)
MOD: Modal
NEG: Negation n’t, no longer
 Semantic Roles

NomBank
Project for annotations on nouns

I When different parties have distinct views of the concept referenced

in the noun
I Example: Apple’s agreement with IBM
I Arg0: Apple
I Arg2: IBM

Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 255

 Semantic Roles

FrameNet
Semantic role labeling based on commonsense (background) knowledge

I Distinct sentences, with different verbs and nouns, may map to the
same meaning
I The price of oil increased 7%
I Oil went up 7%
I We saw an escalation of 7% in the price of oil
I The idea is to represent the meaning of a sentence in a normalized
form
I Frame ≈ model ≈ script
I Representation of background knowledge that lends meaning to
language
I Each word produces one or more frames
I Frame elements: frame-specific semantic roles
I Frame predicates: those applicable to these roles

Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 256

 Semantic Roles

Example Frame: Change Position on a Scale

FrameNet labelers guide

Core Roles
item The entity that has a position on the scale
attribute A scalar property of the item whose value is changing
difference The displacement of the item on the scale
initial value Position on the scale from which the item moves
initial state item’s state before change: independent predication
final value Position on the scale where the item ends up
final state item’s state after change: independent predication
value range Part of the scale over which the attribute varies
Selected Noncore Roles
duration Over which the change takes place
speed The rate of change of the attribute’s value
group The group in which an item changes the value of an
attribute in a specified way

Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 257

 Semantic Roles

Exercise: Identify the Roles in Each Sentence

I Oil prices have risen by 7%

I The price of oil has gone up by $2 since last Thursday

Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 258

 Semantic Roles

Words in the Example Frame

The complete frame

Verbs
advance climb decline decrease diminish dip
double drop dwindle edge explode fall
fluctuate gain grow increase jump move
mushroom plummet reach rise rocket shift
skyrocket slide soar swell swing triple
tumble
Nouns
decline decrease escalation explosion fall fluctuation
gain growth hike increase rise shift
tumble
Adverbs
increasingly

Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 259

 Semantic Roles

Frames Build on Other Frames

I Cause Change of Position on a Scale: composes

I Change of Position on a Scale
I Cause relation
I agent role

Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 260

 Semantic Roles

Selectional Restrictions
Constraints on a word’s arguments that reflects its meaning
Consider I ate tofu today I ate nearby today
I Typical reading of eat in the “nearby” sentence
I Intransitive
I Nearby indicates location of the eating event
I Funny reading of eat in the “nearby” sentence
I Transitive
I Nearby indicates its direct object
I Selectional restriction: theme of eat is (usually) edible
I Associated with a word sense, not an entire lexeme
I Two senses for serves, whose , respectively
Emirates serves breakfast and lunch theme is food
Emirates serves Dubai and Mumbai theme is location
I Adjectives can have arguments too: odorless applies naturally to
objects that can have an odor
Silence of the Lambs: I am having an old friend for dinner
Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 261
 Semantic Roles

Representing Selectional Restrictions

I Words vary in the restrictiveness of their restrictions

I Imagine’s theme can be any entity
I Lift’s theme can be any physical entity
I Diagonalize’s theme must be a matrix
I Represent as unary predicate capturing the restriction on the specified
argument
I Accurate but cumbersome
I Identify which WordNet synset is acceptable
I A filler is acceptable if it is a hyponym of this synset
I That is, one sense of the filler word satisfies the restriction

Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 262

 Semantic Roles

Selectional Preferences
Generalizing beyond hard restrictions
I Strict restrictions are often violated in language
I With negation: can’t eat gold
I With anomalous or surprising occurrences: eat glass
I Selectional preference strength—how selective a verb is
I Eat is informative about its direct objects
I Be is not too informative about its direct objects
I Compare probability distributions of object class c with object class c
given verb v
I P(c|v ): actual distribution of c given v
I P(c): approximation of above not knowing v
I KL divergence from P(c|v ) to P(c): How much information verb
v carries about its arguments
P(c|v )
S(v ) = ∑ P(c|v ) log
c P(c)

Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 263

 Semantic Roles

Selectional Association with WordNet Classes

I Selectional association of a WordNet class c and v

1 P(c|v )
A(c, v ) = P(c|v ) log
S(v ) P(c)
I Positive when v prefers c; negative when v repels c
I For n, let cmax = arg max A(c, v ), from the Brown corpus
n belongs to c
I S(v ) is summed over all classes (previous page)
I P(c|v ) and P(c) refer to this class

Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 264

 Semantic Roles

Examples of Selectional Association with WordNet Classes

Shown for direct objects

I A(cmax , v ) figures are scaled up 100×

I Plausibility below of object-verb pairs: judged by humans in a prior
study
Plausible direct objects Implausible direct objects
Verb v Object Class cmax A(cmax , v ) Object Class cmax A(cmax , v )
Read article writing 6.80 fashion activity −0.20
Write letter writing 7.26 market commerce 0.00
See friend entity 5.79 method method −0.01
Judge contest contest 1.30 climate state 0.28
Answer request speech act 4.49 tragedy communication 3.88

I Some mistakes, apparently due to word sense confusions

I For Answer, tragedy is treated as a play (hence, placed in
communication)

Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 265

 Semantic Roles

Simplified Selectional Preferences

Avoids use of WordNet

I Simply calculate co-occurrence for specific pairs of words

I For verb v , noun n, relation r , estimate one of these from the counts
I P(n|v , r )
I P(v |n, r )
I log count(v , n, r )

Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 266

 Semantic Roles

Evaluating Selectional Preferences

How well a verb matches a noun in a role

I Human judgments
I About plausibility of verb-argument pairs
I Use as basis for correlation with a model
I Pseudowords: for each verb
I Take a legitimate argument
I Generate a confounder as the nearest neighbor in the sense of
having a frequency closest to but greater than the legitimate
argument
I Evaluate how often a model chooses the legitimate word or the
confounder
I Variations on how to generate the confounders

Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 267

 Semantic Roles

Decomposition into Predicates

Conceptual dependencies: relate to frames and scripts
Seek to capture the core meaning of a sentence, i.e., a verb
Primitive Definition
atrans The abstract transfer of possession or control from one
entity to another
ptrans The physical transfer of an object from one location to
another
mtrans The transfer of mental concepts between entities or within
an entity
mbuild The creation of new information within an entity
propel The application of physical force to move an object
move The integral movement of a body part by an animal
ingest The taking in of a substance by an animal
expel The expulsion of something from an animal
speak The action of producing a sound
attend The action of focusing a sense organ

Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 268

 Semantic Roles

Conceptual Dependency Example

Maps each primitive to a fixed set of roles

The waiter brought Mary the check

I Physical transfer of the check: ptrans, p
I Actor of p: the waiter
I Object of p: the check
I Destination of p: Mary
I Abstract transfer of the check: atrans, a
I Actor of a: the waiter
I Object of a: the check
I Destination of a: Mary

Munindar P. Singh (NCSU) Natural Language Processing Fall 2020 269