Nov 4, 2015 - 11/4/2015. / Thomas Hörberg, Department of Linguistics. ⢠grammatical encoding conditioned on prominence (e.g. Silverstein 1976). Animacy:.
2015-11-04
Incremental syntactic prediction in the comprehension of Swedish Thomas Hörberg Department of Linguistics, Stockholm University
Outline • Grammatical relations in Swedish transitive sentences from a corpusdistributional and psycholinguistic perspective • Assumptions & background • Corpus-based model of incremental argument interpretation • Experimental test of model predictions
2015-11-04
/ Thomas Hörberg, Department of Linguistics
1
2015-11-04
Assumptions – Grammatical Relations •
express argument functions (actor/undergoer, topic/focus)
•
grammatical encoding conditioned on prominence Animacy:
human < animate < inanimate
Person:
first, second < third
(e.g. Silverstein 1976)
Referentiality:
pronoun < proper name < common noun
Definiteness:
definite < specific indefinite < unspecific indefinite prototypical
prototypical
subject
object
•
subject > object in prominence
•
Exceptions (i.e. object > subject) typologically marked and infrequent in discourse
11/4/2015
/ Thomas Hörberg, Department of Linguistics
Assumptions – Argument interpretation •
Assignment of argument functions
•
Highly incremental process that is probabilistic and frequency-driven, i.e.
(Actor / Undergoer)
draws upon statistical regularities in the input
/ determining W.O.
(e.g. Levy 2008; MacDonald &
Seidenberg 2006)
•
Morphosyntactic, prominence based and verb semantic information serve as Argument Interpretation Cues
The competition model (MacWhinney & Bates 1989)
11/4/2015
eADM (Bornkessel-Schlesewsky & Schlesewsky 2009; 2013)
/ Thomas Hörberg, Department of Linguistics
2
2015-11-04
Sentence properties • Potentially ambiguous SVO / OVS word order • Case marking on personal pronouns only [DeNP1.NOM] har [vapen som väl matchar alla övriga nationers i regionenNP2.AMB] ”[TheyNP1.NOM] have [weapons that match those of other nations in the regionNP2.AMB]” ”[Gallant FloweringNP1.AMB] varnade [viNP2.NOM] för kraftigt” ”[Gallant FloweringNP1.AMB] [weNP2.NOM] warned about heavily [En härligt örtkryddad soppaNP1.AMB] käkade [viNP2.NOM] bland annat ”[A wonderful soup spiced with herbsNP1.AMB] [weNP2.NOM] ate among other things”
2015-11-04
/ Thomas Hörberg, Department of Linguistics
NP1 + NP2 properties Animacy:
animate vs. inanimate
Givenness:
given vs. new
Definiteness:
definite vs. indefinite
Number:
singular vs. plural
Egophoricity:
1st / 2nd vs. 3rd person
Pronominality:
pronoun vs. noun
Case:
unmarked vs. nominative vs. accusative
Text deixis:
text deictic vs. non-deictic
2015-11-04
/ Thomas Hörberg, Department of Linguistics
3
2015-11-04
Verb semantic properties • verbs assign role semantic properties to NP arguments, i.e. “degree of Actorhood”
(e.g. Dowty 1991, Primus 2006)
• general verb semantic “entailments” and their interactions with prominence features therefore included Dowty (1991) Actor Undergoer Volitional Undergoes involvement change of state Sentience Cause event or Incremental theme change of state Causally affected Movement Stationary -
Category Volitionality Experiencer Causation Possession
2015-11-04
Primus (2006) Control – volitionality and intentionality Sentience Physical involvement Possession
/ Thomas Hörberg, Department of Linguistics
Corpus properties Svensk Trädbank: balanced written Swedish texts
Corpus
SUC
TB
Genre Press: reportage Press: Editorial Press: Reviews Skills, Trades and Hobbies Popular Lore Belles Letters, Biography, Memoirs Miscellaneous Learned and Scientific Writing General fiction Mysteries and Science fiction Light reading Humor Brochure texts Newspaper texts Educational texts Debate articles
N texts 44 17 27 58 48 26 70 83 82 19 20 6 25 28 14 18
N sentences 7278 2385 3961 8933 6525 3598 10847 9633 13028 4070 2908 1071 1733 1669 1624 1134
N words 106079 40887 66002 134947 109665 61297 163333 192827 191507 45321 46126 14428 23122 24125 25623 23476
N hits 1495 473 712 1840 1503 805 1540 1809 3110 826 749 248 390 361 374 316
SVO sentences: 15679 OVS sentences: 872 2015-11-04
/ Thomas Hörberg, Department of Linguistics
4
2015-11-04
Model of incremental argument interpretation •
Models the on-line change in the expectation of OVS (i.e. suprisal, Levy 2008)
•
given AICs provided by constituents over time
Suprisal of OVS modeled in terms of relative entropy between p(OVS | Ci) and p(OVS | Ci-1), i.e. the Kullback–Leibler divergence: DKL(p(OVS | Ci) || p(OVS | Ci-1)) = ∑jlog(p(OVS | Ci)j/ p(OVS | Ci-1)j) p(OVS | Ci)j
•
Based upon (penalized) logistic regression estimates of p(OVS | Ci) at constituents / time points Ci - Baseline model: p(OVS) - NP1 model: p(OVS | NP1) - NP1 + verb model: p(OVS | NP1, Verb) - Full model: p(OVS | NP1, Verb, NP2)
2015-11-04
/ Thomas Hörberg, Department of Linguistics
Model of incremental argument interpretation Probabilities
Suprisals
baseline model: p(OVS): ~0.05
Suprisal NP1
DKL(p(OVS | NP1) | | p(OVS))
NP1 model: p(OVS | NP1)
Suprisal verb
NP1 + verb model:
DKL(p(OVS | NP1 + verb) | | p(OVS | NP1))
p(OVS | NP1 + verb)
Suprisal NP2
full model:
DKL(p(OVS | NP1 + verb + NP2) | | p(OVS | NP1 + verb))
p(OVS | NP1 + verb + NP2)
2015-11-04
/ Thomas Hörberg, Department of Linguistics
5
2015-11-04
Suprisal of upcoming constituents DKL = 0.06
[De
NP1]
They
”[They
DKL = 0.00 DKL = 0.00
[harverb] [vapen som väl matchar alla övriga nationers i regionen weapons that match those of other nations in the region
have NP1]
[have
verb]
DKL = 0.02
[Gallant Flowering Gallant Flowering
”[Gallant Flowering
[weapons that match those of other nations in the region
NP1]
NP1]
DKL = 0.00
[varnade warned
[we
NP2]
DKL = 0.34
[En härligt örtkryddad soppa A wonderful soup spiced with herbs
[warned
NP1]
”[A wonderful soup spiced with herbs
2015-11-04
verb]
we
verb]
NP2]
för kraftigt
p=.88
about heavily
about heavily”
DKL = 0.72 ate
[we
NP2]”
DKL = 5.49
[vi
[käkade
NP1]
NP2] p.999
among other things
among other things”
/ Thomas Hörberg, Department of Linguistics
Suprisal at NP1 • Lexical NP1 • Some suprisal for inanimate / text deictic NP1
2015-11-04
/ Thomas Hörberg, Department of Linguistics
6
2015-11-04
Suprisal at verb • Lexical NP1 • Moderate suprisal for Inanimate & Volitional / Causative + Experiencer
2015-11-04
/ Thomas Hörberg, Department of Linguistics
Suprisal at NP2
(disambiguation towards OVS)
• Lexical NP1 • case marked NP2 that disambiguates sentence towards OVS
2015-11-04
/ Thomas Hörberg, Department of Linguistics
7
2015-11-04
Testing the model
•
Testing the strongest model predictions
•
Self-paced reading
•
Reading times assumed to reflect processing
2015-11-04
/ Thomas Hörberg, Department of Linguistics
Self paced reading
###### ####### ### #### ### #
2015-11-04
#######
/ Thomas Hörberg, Department of Linguistics
8
2015-11-04
Self paced reading
Bollen ####### ### #### ### #
#######
ball.the
2015-11-04
/ Thomas Hörberg, Department of Linguistics
Self paced reading
###### sparkar ### #### ### #
#######
kick
2015-11-04
/ Thomas Hörberg, Department of Linguistics
9
2015-11-04
Self paced reading
###### ####### han #### ### #
#######
he
2015-11-04
/ Thomas Hörberg, Department of Linguistics
Self paced reading
###### ####### ### mitt ### #
#######
middle
2015-11-04
/ Thomas Hörberg, Department of Linguistics
10
2015-11-04
Self paced reading
###### ####### ### #### upp #
#######
up
2015-11-04
/ Thomas Hörberg, Department of Linguistics
Self paced reading
###### ####### ### #### ### i
#######
in
2015-11-04
/ Thomas Hörberg, Department of Linguistics
11
2015-11-04
Self paced reading
###### ####### ### #### ### #
krysset top.corner.the
2015-11-04
/ Thomas Hörberg, Department of Linguistics
Self paced reading
Bollen sparkar han mitt upp i
krysset
ball.the
top.corner.the
kick
he
middle up
in
”The ball he kicks right up into the top corner” • Dependent variable: time latency between button presses • Analyses done on region RTs rather than word RTs • Task: Comprehension question following each sentence: e.g. ”Does he kick the ball right up into the top corner?”
2015-11-04
/ Thomas Hörberg, Department of Linguistics
12
2015-11-04
Predictions – sentence differences Word order
Verb class
Object animacy
Example
Inanimate
Bollen sparkar han mitt upp i krysset
Animate
Killen sparkar han mitt på smalbenet
Inanimate
Bollen glömmer han mitt på fotbollsplanen
Animate
Killen glömmer han sent på kvällen
Inanimate
Han sparkar bollen mitt upp i krysset
Animate
Han sparkar killen mitt på smalbenet
Inanimate
Han glömmer bollen mitt på fotbollsplanen
Animate
Han glömmer killen sent på kvällen
”The ball he kicks right up in the middle of the top corner”
Volitional ”The guy he kick on the middle of the shin”
OVS ”The ball he forgets in the middle of the football field”
Experiencer ”The guy he forgets late at night”
”He kicks the ball up in the middle of the top corner”
Volitional ”He kicks the guy on the middle of the shin”
SVO ”He forgets the ball in the middle of the football field”
Experiencer ”He forgets the guy late at night”
Regions • e.g. /Bollen sparkar
reg.1
/ jag mitt
reg.2
/ upp i
reg.3
/ krysset
FW
/
• Region 1 RTs correspond to verb suprisal, region 2 RTs to NP2 suprisal
2015-11-04
/ Thomas Hörberg, Department of Linguistics
Results – RT differences
** **
• Faster RTs in SVO sentences vs. OVS sentences • Faster RTs for animate vs. inanimate NP1 at region 1 in OVS experiencer verb sentences • Slower RTs for animate vs. inanimate NP1 at region 2 in OVS volitional verb sentences • No significant RT differences in SVO sentences 2015-11-04
/ Thomas Hörberg, Department of Linguistics
13
2015-11-04
Conclusions
• Animacy functions as a cue in argument interpretation that interacts with verb class • Distribution of prominence cues in discourse predicts reading times
2015-11-04
/ Thomas Hörberg, Department of Linguistics
References Bornkessel-Schlesewsky, I., & Schlesewsky, M. (2009). The role of prominence information in the real-time comprehension of transitive constructions: A cross-linguistic approach. Language and Linguistics Compass, 3(1), 19-58. Bornkessel-Schlesewsky, I., & Schlesewsky, M. (2013). Reconciling time, space and function: A new dorsal–ventral stream model of sentence comprehension. Brain and Language, 125, 60-76. Dowty, D. (1991). Thematic Protoroles-Roles and Argument Selection. Language, 67(3). Levy, R. (2008). Expectation-based syntactic comprehension. Cognition, 106(3), 1126-1177. MacDonald, M. C., & Seidenberg, M. S. (2006). Constraint satisfaction accounts of lexical andsentence comprehension in M. Traxler and M. A. Gernsbacher (Eds.) Handbook of Psycholinguistics (pp. 581-610). New York, NY: Academic Press). MacWhinney, B., & Bates, E. (1989). Functionalism and the Competition Model. In B. MacWhinney & E. Bates (Eds.), The Crosslinguistic Study of Sentence Processing (pp. 3-73). Cambridge: Cambridge University Press. Primus, B. (2006). Mismatches in semantic-role hierarchies and the dimensions of role semantics. In I. Bornkessel, M. Schlesewsky, B. Comrie, & A. D. Friederici (Eds.), Semantic Role Universals and Argument Linking:Theoretical, Typological and Psycholinguistic Perspectives (pp. 53-89). Berlin: Mouton de Gruyter. Silverstein, M. (1976). Hierarchy of features and ergativity. In R. M. W. Dixon (Ed.), Grammatical Categories in Australian Languages (pp. 112-171). New Jersey: Humanities Press.
2015-11-04
/ Thomas Hörberg, Department of Linguistics
14