Transcript Verb classes and event classes: From grammar to processing

Verb classes and event classes:
From grammar to processing
Jean-Pierre Koenig
University at Buffalo
Collaborators
•
•
•
•
•
•
Breton Bienvenue
Gail Mauner
Karin Michelson
Shaakti Poornima
Doug Roland
Hong-Oak Yun
Verb classes vs. Event classes I
• Lots of way of classifying event-types,
– Some of them are well-established in memory;
– Some less, cf. Barsalou’s (1983) ad hoc categories :
• Ways to escape being killed by the Mafia;
• Linguists focused on event classes that matter
for morphosyntax = verb classes:
– If rule of grammar targets a class of verbs, then
this class is real (verb class);
Outline
1. Two examples of “true” verb classes that
reference relatively rare semantic properties;
2. Two examples of the use of semantically
coherent classes of verbs to answer
foundational semantic questions about
what’s in a verb meaning
3. Verb classes and classes of verbs associated
with an event class differ in
ontological/epistemological status
Outline
1. Two examples of “true” verb classes that
reference relatively rare semantic properties;
What are the boundaries of the semantic properties
relevant to morphosyntactic processes?
Iroquoian kin “verbs”
• Most stems denoting kin relations in Iroquoian are partly
verbal and partly nominal (Koenig and Michelson, In Press);
• All verbal stems realize both arguments of kin terms;
(1)waʔ-shako-hnutla-neʔ
FACTUAL.MODE-3MASC.SG>3-catch.up.toPUNCTUAL.ASPECT
‘he caught up to her’
• (Synchronically) a single pronominal prefix encodes
properties of both arguments (written AGT>PAT)
Generational age in Oneida (Iroquoian)
(1)
(2)
lo-nulhá·
3ZOIC.SG>3MASC.SG-mother
‘his mother’
luwa-yʌha
3FEM.SG>3MASC.SG-child
‘her son‘
• Subject assignment rule 1 (refers to generation):
The argument that corresponds to the older
generation maps onto the “Agent,” while the
argument that corresponds to the younger
generation maps onto the “Patient.”
• We must add Generational (and Absolute) age
of the arguments in our list of event
properties relevant to linking;
• Rule 1 is not unique to Oneida, true of several
“kin verbs” languages (e.g., Ilgar, Australia);
The rule that targets a verb class
• If a verbal stem denotes a kin relation, then the
Agent>Patient prenominal prefix encodes
properties of the generationally older>younger
relata;
• Aside from the content of the rule, nothing
special about Oneida kin terms:
– The content of rule 1 is different from other eventproperties that affect linking to subject, but its form is
similar:
• for all X and Y, if MOTHER(X,Y), then X is generationally older
than Y
Conditional counter-expectations in
Hindi ergative case marking
• Hindi uses ergative case in sentences
containing both transitive and intransitive
verbs;
• Rule 1: If the verb is transitive and perfective,
the subject is assigned ergative case.
• The situation is more complex for intransitive
verbs;
(1) mein bahut log moujuud th-ee phir bhii kiisii par
court in many people present be-Past.3.Pl still
bhii kuttee=ne bhauunk-aa tak nahii
any on also dog=Erg bark-M.Sg even neg
‘Many people were present in court but still the dog did not
even bark at anyone.’
(2) Tansen=ne bas gungunaa-yaa aur barish shuru ho gay-ii
Tansen=Erg just hum-M.Sg and rain start be go-Perfv.F.S
‘Tansen (famous 15th century singer) just hummed and it
started raining.’
Conditional counter-expectations
• Ergative case marks that it is surprising that the
dogs didn’t bark and that Tansen only hummed
and it rained (we would have expected singing to
be required);
• We can model the semantic contribution of Hindi
intransitive ergative case-marking as follows:
– It introduces a function that selects possible worlds in
which Tansen acts: W → WTansen=Agent
– A Kratzer-style analysis of unexpectedness applies on
the resulting worlds:
• WTansen=Agent → WTansen=Agent+expected
• p not in WTansen=Agent+expected
Rule for Hindi ergative case-marking
on intransitive verbs
• If the verb is intransitive and perfective, denotes
a bodily function, and the action is unexpected
given the actor, then the subject is assigned
ergative case (≠Butt, de Hoop and Narasimhan).
khaas ‘cough’, chiikh ‘sneeze’, bhauk ‘bark’, ciik
‘scream’, cillaa ‘yell’, muut ‘urinate’, and thuuk
‘spit’
• Why only verbs denoting bodily functions are
targeted is unclear.
Conclusions
• The rules for Oneida and Hindi lend “reality”
to verbal kin stems and verbs denoting bodily
functions, respectively:
– These verbs behave as a class for a linguistic
process;
– The morphosyntactic processes is what make
these classes useful
Event classes
• Talk of verb classes is often simply a shorthand for the event classes associated with
various sets of verbs;
• Selecting a class of verbs on the basis of
shared event features can be a very useful
discovery procedure or useful tool for
purposes of experimental manipulations…
• …but there is no guarantee that the resulting
classes of verbs are “real” (are verb classes)?
Outline
2. Two examples of the use of semantically
coherent classes of verbs to answer
foundational questions about what’s in a
verb meaning
a. (Semantic) arguments and adjuncts
(Koenig et al., 2003)
b. What kind of idiosyncratic information
information verbs include?
Syntactic optionality
1. John was chased by someone.
2. John ate pizza.
3. The refugees emigrated to Canada last year.
4. The library provides web access to students.
5. John borrowed a book from Mary.
6a. The knight beheaded the king with a sword.
6b. The knight killed the king with a sword.
7. Mark hid the picture in the closet
8. Kim ate lunch in the park.
9. John practiced piano yesterday.
10. The swimmer won the race with ease.
11. Sue baked a cake for the PTA.
Agent
Patient
Goal
Recipient
Source Loc
Obl. Instrument
Opt. Instrument
Participant Loc
Event Loc
Event Time
Manner
Beneficiary
Kenny’s problem
• Given rampant syntactic optionality of postverbal
dependants, how do we know how many argument
positions a verb’s denotation has?
(1)
(2)
John ate → John ate something
John ate → John ate somewhere…
• How do we know that a participant role is part of the
meaning of a verb ≈ strongly activated upon
recognition of a verb?
Frequency of expression won’t do
(≠ McDonald et al., 1994)
• Obligatory instruments: 8% (Brown)
• Optional instruments: 10% (Brown)
• Source locations: 20% (BNC; range: 1.4%50.4%)
• Participant locations: 30% (BNC)
• Event locations: 7.5-8.8 (BNC; range: .15%93%)
Category Utility (Corter and
Gluck)/Mutual Information (Church
and Hanks)
• Literature on categorization might help here
• “Utility” of a category depends on how many
predictable features it has that not many
categories have (that are distinctive):
– Inversely proportional to the conditional
probability of the feature given the category: How
frequently tokens of an event category include
feature (1/p(f) or-p(f));
– Proportional to p(f|c)
Categories and feature activation
• Features that are more distinctive of a
category are more activated and activated
more quickly than features shared with many
other categories (Cree et al.; Sparck-Jones);
– BANANA: HAS A PEEL >> HAS A SKIN
– BUTTERFLY: HAS A COCOON >> HAS ANTENNAE
Event categories and feature activation
(1) Cordelia kissed Xander in the library. (Event location)
(2) Willow hid the amulet in her pocket. (Participant
location)
(3) Buffy expelled Spike from the club. (Source location)
• Event-features = participant roles:
– EXPEL: INVOLVES A SOURCE LOCATION >> OCCURS
SOMEWHERE;
– HIDE: INVOLVES A PARTICIPANT LOCATION >> OCCURS
SOMEWHERE
• The more distinctive a participant role is, the more
quickly and strongly it should get activated;
– Activation is proportional to 1/p(f) or to –p(f)
Measuring distinctiveness
• Two raters judged for the around 4,000 verbs
they knew that:
– 98% of verbs required an event location;
– 14% of verbs required a source location;
– 7% of verbs required a participant location;
• Event locations should be weakly activated
(semantic adjuncts); source/participant locations
should be strongly activated (semantic
arguments);
Testing distinctiveness
• The integration of WH-fillers into a sentence
representation is sensitive to the lexical
properties of a verb (Stowe; Boland);
• This is true of PATIENT/THEME and RECIPIENT;
• …But, we predict, also of the less frequently
expressed semantic arguments like SOURCE
and PARTICIPANT LOCATION
Materials and methods
1. (In/From) which office | was the incompetent
employee | reprimanded /dismiss| (in/from) [gap] by
the manager | yesterday?
2. (In) which bush | were the squirrel’s acorns |
eaten/hoarded | (in) [gap] by the chipmunk | last fall?
•
•
Non-accumulating moving window with stopsmaking-sense judgment;
Materials normed extensively for
grammaticality/plausibility/plausibility of fillers as
instruments/implausibility of fillers as patient
Logic and predictions
• Integration of WH-fillers into sentence
representations should be easier when the
relevant participant role is more quickly and
more strongly activated;
• Readers should take longer to read verb or
post-verbal regions when location role feature
is weakly activated by event category than
when it is strongly activated by event
category;
Results (only source/event location
contrast
is
shown,
Conklin
et
al.,
2004)
500
Residual Reading Times (ms)
NP-Filler "eject"
400
NP-Filler "beaten"
300
PP-Filler "eject"
200
PP-Filler "beaten"
100
0
-100
-200
-300
p at ien t
-400
v er b
ag e n t
Conclusions
• Participant role distinctiveness affects the
activation of a participant role (both for
source and participant locations);
• Our conditions grouped together verbs in
terms of the location role they include: These
verbs behave as a class.
• Does that provide evidence of the “reality” of
the class of participant and source location
verbs?
Co-occurrence of participant role and
tokens of event category
• Strength of association between category and
features depends on how frequently tokens of
an event category include participant role;
– Activation proportional to p(f|c)
• We concentrate on the end of the continuum:
+/-Obligatory (which has a special linguistic
status)
Obligatory/Optional instruments
Semantically Obligatory Instrument Verb
The barbarian hacked someone with a sword
during the attack
Semantic Optional Instrument Verb
The barbarian injured someone with a sword
during the attack
•
Two raters judged for all the verbs they knew
(about 4,000) whether their meaning required
(12%) or merely allowed an instrument (35%).
Participant role information is used to generate
expectations about who or what is going to be
mentioned next
• Altmann & Kamide (1999)
‘The boy will move the cake’ or ‘The boy will eat the cake
Visual Materials (Bienvenue, forthcoming)

Visual display depicted
 Agent, Instrument, two scene relevant distractors, NO Patient
 Position counterbalanced across trials
Norming
 Foils equally atypical
as instruments for both
sentences
 On screen prior to, during
and after sentence heard

Predictions
• More anticipatory looks when instruments
are obligatory than when they are optional
(because instrument role feature is more
activated in first case)
• Anticipatory looks will emerge at the verb
(because of early presentation of visual
scene);
More trials with looks to instruments in instrument argument than
instrument adjunct sentences
Differences emerge at verb
Proportion of Trials with Saccades
0.7
0.6
Sword (Hack)
Sword (Injure)
0.5
0.4
0.3
0.2
0.1
0
hacked
someone
with
Region
a sword
during...
What’s in a verb meaning?
• Category utility/mutual information (low type
frequency and high token frequency) provides
a good solution to Kenny’s problem and a
good model for the distinction between
semantic arguments and adjuncts
Classes of verbs or verb classes?
• Our experiments (and other similar ones)
show that as a group (1) verbs that require an
instrument differ in a processing relevant way
from verbs that allow an instrument and (2)
verbs that have distinctive location roles differ
in a processing relevant way from verbs that
have non-distinctive location roles;
…Just classes of verbs
• These experiments do not demonstrate that +/obligatory instrument or +/-distinctive location
are an organizing principle of the verbal lexicon;
– The similarity reduces to a similarity of event
categories ;
– The effect did not depend on substantive semantic
similarities of verbs within each condition (they
involve instruments/locations…), but on formal
similarities (P(role) was low or P(role|category) was
high);
What’s in a meaning of a verb?
A comprehensive look at a corner
of semantic space
Two distinct parts to the meaning of
verbs
• Carter/Levin and Rappaport:
– Structural vs. idiosyncratic aspects of verb meaning:
• Kill: CAUSE(X, BECOME(DEAD(Y)))
• Questions:
(1) What kind of information does idiosyncratic aspects
of verb meaning encode?
(2) Is the maximal complexity of structural meaning truly
a single cause-effect pair?
cause(s1, s2) vs. cause (s1, s2) and cause (s2, s3)
A comprehensive look at a corner of
semantic space (Koenig et al., 2008)
• We examined the list of verbs that
semantically require (≈500) or allow an
instrument (≈1,300);
• Classify them in terms of:
– Subsituations: s1 (Agent and possibly instrument);
s2 (Instrument and possibly patient); s3 (patient
and possibly instrument) (s2 was not necessarily
present);
– s1 precedes s2; s2 precedes s3.
Examples
CUT.
cause(s1, s2) ∧ act(s1, A, I) ∧ contact(s2, I, P) ∧
cause(s2, s3) ∧ incised(s3, P)
1. Incise : carve (a piece of wood), notch, plow, scratch,
etch ;
2. Pierce : puncture, harpoon, knife, prick, lance ;
3. Sever : amputate, bone, core, eviscerate, castrate,
gore, hack, prune, mow ;
4. Shred : shred, it includes grind, dice, cube, scallop, and
mince ;
DRUG: drug, gas, anesthesize, immunize, vaccinate, dope ;
flavor season.
cause(s1, s2) ∧ act(s1, A, I)∧ in (s2, I, P) ∧ cause(s2, s3) ∧
change-of-state(s3, P)
FILL.
cause(s1, s2) ∧ act(s1, A, I) ∧ in(s2, I, P) ∧ cause(s2, s3) ∧
change-of-configuration(s3, P)
(1) Jim loaded the truck with boxes with a forklift
SKI. Canoe, bicycle, skate, drive, toboggan.
cause(s1, s2) ∧ act(s1, A, I)∧ pred2(s2, I, A) ∧ and ∧ partcause+(s2, s3) ∧ movemanner(s3, A)
SCOOP. Spoon, pump, milk, sponge, ladle, shovel, siphon.
cause(s1, s2) ∧ act(s1, A, I) ∧ in(s2, P, I) ∧ enable(s2, s3) ∧
go-to(s3, P, Z)
(1) The plug’s coming loose let the water flow from the
tank.
EAT.
• Very large class with little semantic coherence:
a. Jean doesn’t know how to eat with chopsticks.
b. Jill drank her soda with a straw.
c. Ryan watched the bird with his new binoculars.
d. Alicia lectures with overheads rather than with handouts.
e. Joan hunted the turkey with a bow and arrows.
f. He plays volleyball with gloves.
g. Susan always practices the piano with a metronome.
h. Max repaired the faulty switch with a screwdriver.
help+(s1, s3) ∧ pred2(s1, A, I) ∧ pred1(s3, P)
What do you get for classifying 1,800
verbs?
1. Expansion of maximum bound on structural
semantic complexity is needed, but limited:
use of tools;
2. Idiosyncratic information specifies more
instrument activity and change of state in
patient than agent activity;
– Another example of goal bias (voir Lakusta and
Landau (2005)) and lexical reification of discourse
distribution (Slobin (2004)) ?
3. There is variation in causal relation between
subsituations, suggesting that root meanings might
sometimes be molecular
a. John watches birds all day with his binoculars.
b. Bill cooks his steaks with butter.
c. Floyd baked the cake with yeast.
d. Bill entered Joan’s room with a duplicate key.
e. Joe scooped the ice-cream with a wooden spoon.
f. Connie skied down the slope with her new skis.
g. Alisa walks her cat with a leash.
• s2 can be the true cause of the final change of
state s3: cut
• s2 can be the cause of a precondition of the
change of state s3: open
• s2 can be one of a joint set of causes of the
change of state s3: ski
• s2 can enable a change of location s3: scoop
• s2 can cause the event/action to lead to a better
resulting state or to be performed better: cook
with butter
An intensional analysis of “helping”
• Definition 1 An eventuality e1 helps the
occurrence of token e2 of the event category
C iff (i) there is an ordering of tokens of C
along a pragmatically defined scale (ease of
performance, how good the resulting state is,
fewer unwelcome “side-effects”); (ii) e1
caused the token e2 of C to be higher on that
ordering than it would otherwise have been.
Intensional causality
• Our analysis of helping seems to make causality
dependent on how events are described (e..g, cooked
better);
(1) Jeri’s new shoes made her run fast (#made her run);
(2) Marc’s numbness made him drive above the speed
limit (#made him drive);
(3) Roberto’s painkillers made him paint less realistically
(#made him paint);
• This view accords well with view that events are
categorized processes (Link)
4. Obligatory instrument verbs constrain more
instrument properties than optional verbs do
(cf. behead vs. kill);
(There is some fun computational modeling
work Roland, Yun, Koenig, and Mauner have
done that explores this difference!)
Instrument subclasses
• Classification isolated subclasses of event
categories that require or allow instruments ;
• No “reality” for these verb classes or event
classes can come from syntax (no alternation
differences);
• Is there any evidence that this classification
organizes event classes in semantic memory?
• Semantic priming reported for verb-specific thematic
role features (McRae, et al., 1997):
– ARREST/COP, CUT/KNIFE
• Will semantic features shared among members of
instrument subclasses lead to semantic priming
• Will membership in the obligatory instrument
category lead to priming when there is no featural
overlap?
Materials
Close Semantic Neighbor Prime (Shared Features/Category Membership)
Which knife | did the waitress | slice | the pie with | at the restaurant?
Distant Semantic Neighbor Prime (Joint Category Membership)
Which wine | did the chef | flavor | the duck with | for the inauguration?
Semantically Unrelated Prime (No Shared Features/Category Membership)
Which fork | did the customer | eat | the salad with | at the restaurant?
Target
Which chainsaw | did the technician | lop off| the tree limbs with | to
create room for the power lines?
• Participant-paced region-by-region reading task
• Continuous Priming: prime-target relationship not salient
Processing Assumptions and
Predictions
Close
Neighbor
Close
Neighbor
Prime Verb
Role
Features
Role
RoleFeatures
Features
Target Verb
Target
Verb
Role
Features
Close
Neighbor
Rolefeatures
Features
Role
Verb
Recognition
Close Semantic Neighbor Prime
• Target
Final
region
Closesentence
Neighbor
Primefollowing
Recognition
Somefeatures
target
role with
features
primed
some
inhibited
activeverb
unshared
features
frombut
prime
compete
with target
•• Highly
Role
shared
Close
Neighbors
are
activated
• At•direct
object,
should see
wh-filler integration
verb
features
Unshared
features
arefaster
inhibited
• Due
to more
highly
active
roleof
features
• Competition
lowers
availability
target
verb features needed to
• Some
features
of Target
will
be inhibited
distinguish
verb diminishes
meaning from
verb
• Featuretarget
activation
onlyprime
via interference
(Waugh & Norman,
• 1965)
Lowered availability of features slows processing in final region
950
900
850
Reading Times (ms)
800
DISTANT NEIGHBOR
CLOSE NEIGHBOR
UNRELATED
750
700
650
600
550
500
450
400
Verb
Direct Object
Target Sentence
Final
Evidence for instrument verb
subclasses?
• No!
– Hypothesized priming mechanism is strictly semantic
in nature, so if anything it’s evidence for instrument
event subclasses not verb subclasses;
• The isolated event classes have no privilege
status:
– Priming in verbs is like priming in nouns , driven by
featural overlap, not category membership (McRae &
Boisvert, 1998)
– Many shared features among event classes may lead
to priming, not necessarily the subclasses we isolated
Methodological conclusions
• The grammar of natural languages involves, in
a crucial way, verb classes:
– Nothing new, but there can be some exotic bird
out there (e.g., Oneida, Hindi), at least from our
shores.
• The linguist’s use of semantically homogenous
classes of verbs is critical in investigating the
meaning of verbs and the organization of
semantic/conceptual event space
– …but these are not verb classes!
Substantive conclusions I
• Oneida and Hindi linking rules involve
relatively rare semantic properties, but,
formally, they are rather typical of
semantically-sensitive linking rules;
– Why are languages so conversative when it comes
to the semantic underpinning of linking?
• Event types (the denotation of verbs) behave
like categories
– We can give a motivated answer to the question
of how many argument positions a verb has;
• Event categories that describe the use of a
tool to perform an action are the most
complex class of event-types;
• Idiosyncratic verb meaning information can be
shared across classes of verbs;
– We can use priming to study the clustering of
verbs in semantic space share features induce;
• Talks of causality are event-description
dependent