Shravan Vasishth Universit at Potsdam, Germany vasishth ...vasishth/pdfs/Konstanz2016.pdfBruno Nicenboim, Pavel Loga cev, Carolina Gattei, and Shravan Vasishth. When high-capacity

Complexity and memory

Shravan VasishthUniversitat Potsdam, Germany

[email protected]

February 25, 2016

1 / 53

Acknowledgements

Several people have contributed to the work presented here today:

I Computational modeling using ACT-R: Felix Engelmann,Umesh Patil

I Modeling and reflexives research: Lena Jager

I Modeling and individual differences: Bruno Nicenboim

I Complex predicates in Hindi and Persian: Samar Husain,Farnoosh Safavi

I Modeling underspecification: Pavel Logacev

2 / 53

Introduction and backgroundThe ACT-R model of sentence comprehension

The original model (ACT-R 4.0) is described in:

1. Richard L. Lewis and Shravan Vasishth. An activation-basedmodel of sentence processing as skilled memory retrieval.Cognitive Science, 29:1-45, May 2005.Code: http://www.ling.uni-potsdam.de/∼vasishth/code/LewisVasishthModel05.tar.gz

2. Richard L. Lewis, Shravan Vasishth, and Julie Van Dyke.Computational principles of working memory in sentencecomprehension. Trends in Cognitive Sciences, 10(10):447-454,2006.

3 / 53

The model assumptionsThis is often called “the” cue based model, but there are manycue-based models (Van Dyke’s, McElree’s conceptions are differentfrom the LV05 model).

1. Grammatical knowledge and left-corner parsingalgorithm:Note that a parser can do nothing without a grammar. Soeven asking a question like “is it the grammar or the parser”is technically not meaningful.

I If-then production rules drive structure buildingI Rules are hand-crafted in toy models, but scaling up has been

done (Boston, Hale, Kliegl, Vasishth, Lang Cog Proc 2011).

2. Constraints on memory processes affecting retrieval:allows us to model individual differences in attention andworking memory capacity

Retrieval at any dependency completion point is a key (but notonly) determinant of processing difficulty or facilitation.

4 / 53

Introduction and backgroundThe memory constraints in the model

Code: https://github.com/felixengelmann/ACT-R-Parsing-Module

5 / 53

Introduction and backgroundThe memory constraints in the model: Similarity based interference

6 / 53

Introduction and backgroundThe memory constraints in the model: Partial Matching

The tough soldier who Kathy met killed himself. + c-commander

+ masculine + c-commander + masculine

The tough soldier who Bill met killed himself.- c-commander

+ masculine+ c-command

+ masculine + c-commander + masculine

* The tough girl who Kathy met killed himself.

+ c-commander + masculine

7 / 53

Introduction and backgroundEvidence for partial matching: Processing polarity

(1) a. Accessible NPI licensorKein Pirat, [der einen Braten gegessen hatte,] war jemalssparsam

b. Inaccessible NPI licensor*Ein Pirat, [der keinen Braten gegessen hatte,] war jemalssparsam

c. No NPI licensor*Ein Pirat, [der einen Braten gegessen hatte,] war jemalssparsam

Condition Data Model(1a) Accessible licensor 85 96(1b) Inaccessible licensor 70 61(1c) No licensor 83 86

[I will return to reflexives later]

8 / 53

Introduction and backgroundHow the memory constraints relate to fRM

There is a historically important idea in memory research:proactive (π) vs retroactive (ρ) interference.

(2) π Target ρ Retrieval

I Retroactive interference is much stronger in sentenceprocessing (See Van Dyke’s work)

I The hierarchical interference idea in fRM implements maps toretroactive interference:

(3) X [Z Y ]

except in Chinese: [[Z Y] X ]But 13 years after Hsiao and Gibson 2003, we still don’t knowwhat’s going on with Chinese RCs.

9 / 53


estimated coefficient (ms)

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

−300 −200 −100 −50 0 50 100 150 200 250 300

Gibson et al 13

Vas. et al 13, E3

Lin & Garn. 11, E1

Qiao et al 11, E1

Lin & Garn. 11, E2

Qiao et al 11, E2

Hsiao et al 03

Wu et al, 11

Wu 09

Jaeg. et al 15, E1

Chen et al 08

Jaeg. et al 15, E2

Vas. et al 13, E2

C Lin & Bev. 06

Vas. et al 13, E1

stud

y id

posterior

OR advantage SR advantage

10 / 53


So most of the discussion is about retroactive interferenceconfigurations.

Target Intruder Retrieval Cues LV05 fRM

AB C AB Baseline DisjunctionAB B AB SBI+misretrievals InclusionA B AB misretrievals ?

AB BC AB misretrieval Intersection

The predictions of the LV05 model will correspond to the fRMexcept in Inclusion, where speedups will occur due to misretrievalin some trials of the intruder.

11 / 53


I fRM’s feature ranking to explain non-criterial inclusion isimplemented as cue-weighting (see Kush PhD dissertationfrom Maryland)

I c-command has a special status in both fRM and our ACT-Rbased model.

In the rest of this talk, I will tell the story of intervention from myperspective as a psycholinguist.

12 / 53

Do grammatical constraints always play a role in sentencecomprehension?

I It is clear from the research on reflexives and so on thatsyntactic constraints come into play in parsing.

I However, one should not dismiss systematic departures fromthis case: local coherence effects, the ambiguity advantage,and good-enough parsing are examples.Local coherence:

(4) The coach [that] smiled at the player tossed afrisbee. . .

Ambiguity advantage:

(5) The son/car of the driver who had a moustache. . .

Good-enough parsing:

(6) While Mary dressed the baby slept. (Was Marydressing the baby? Was the baby sleeping? )

13 / 53

Intervention does not have a unitary effect on languagecomprehension

1. Intervention can lead to greater difficulty in completingdependencies due to

I backward looking processes: interference+decay constraintsI forward looking processes: storage cost (Gibson 2000),

predictions increasing entropy (Safavi et al 2016)

2. Intervention can also lead to reductions in difficulty :I through misretrievals of the intervener (NPIs, agreement

attraction)I through richer encoding of arguments (Hofmeister 2007)I by sharpening expectations (surprisal effects, Levy 2008).I by lowering entropy (Linzen and Jaeger 2015)

14 / 53

Some key factors in play when interveners are present

So, the effect of intervention seems to depend at least on:

1. what retrieval processes are triggered after interventionregion

2. what linguistic retrieval cues are deployed during retrieval(morphosyntactic cues play a role here)

3. individual differences in working memory capacity andcognitive control

4. systematic underspecification, which could lead to subjectnot completing a dependency

5. what is being predicted (could vary by individual-level abilityto maintain predictions)

6. what information is contained in the intervener (couldsharpen or weak predictions)

15 / 53

Intervention as a cause of complexity

Explanations for speedups and slowdowns in dependencycompletion have three major classes of explanation:

1. Inhibition and facilitation arising from constraints on retrievalThe computational model of Lewis and Vasishth 2005, and recent

extensions by Engelmann et al.

2. Inhibition and facilitation arising from capacity-inducedretrieval failuresThe computational model of Nicenboim et al. 2016

3. Inhibition and facilitation arising from predictive-processing(storage, surprisal, entropy)The probabilistic parsing computational models of Hale 2001 and

Levy 2008

I will discuss these three classes of explanation today.

16 / 53

Interference arising from constraints on retrievalExample 1a: Slowdowns due to proactive SBI in reflexives

Similarity-based interference (retrieval cues matching multiplenouns) leads to increased difficulty in completing noun-reflexivedependency:

Badecker and Straub 2002:

[Distractor Jane−masc−c-com/John+masc

−c-com] thought that [Target Bill+masc+c-com]

owed himself {mascc-com} another opportunity to solve the problem.

17 / 53

Interference arising from constraints on retrievalExample 1b: Slowdowns due to retroactive SBI in subject-verb dependencies

Van Dyke 2007: Another example of SBI leading to increaseddifficulty in completing argument-verb dependency:

The worker was surprised that the [Target resident+anim+locSubj] who was

living near the dangerous[Distractor warehouse−anim

−locSubj/neighbor+anim−locSubj]

was complaining{animlocSubj} about the investigation.

Cf: Rizzi 2013:

(7) a. ?[Which problem+Q+N] do you wonder how+Q

−N to solve GAP+Q+N

b. *How+Q−N do you wonder which problem+Q

+N to solve GAP+Q+N

18 / 53

Interference arising from constraints on retrievalExample 1b: Slowdowns due to retroactive SBI in subject-verb dependencies

Van Dyke 2007: Another example of SBI leading to increaseddifficulty in completing argument-verb dependency:

The worker was surprised that the [Target resident+anim+locSubj] who was

living near the dangerous[Distractor warehouse−anim

−locSubj/neighbor+anim−locSubj]

was complaining{animlocSubj} about the investigation.

Cf: Rizzi 2013:

(8) a. ?[Which problem+Q+N] do you wonder how+Q

−N to solve GAP+Q+N

b. *How+Q−N do you wonder which problem+Q

+N to solve GAP+Q+N

18 / 53

Interference arising from constraints on retrievalExample 2a: Speedups due to misretrievals in subject-verb dependencies (agreementattraction)

*The [Target key−plur+locSubj to

the [Distractor cabinet−plur−locSubj/cabinets+plur

−locSubj] were{plurlocSubj} rustyfrom many years of disuse.

This facilitation due to misretrieval is predicted for reflexives, butthere is only limited evidence for this at the moment (King et al2012):

The [Target mechanic−fem+c-com] who spoke to [Distractor

John−fem−c-com/Mary+fem

−c-com] sent a package to herself {femc-com} . . .

19 / 53

Inhibition and facilitation arising from SBI and misretrievals

The ACT-R model makes highly constrained predictions, but theempirical evidence has shown all possible effects!

Target Lewis & Vasishth ’05 Empirical findingsMechanism Predictions

Match SBI & misretrievals (A) Inhibition (A1) No effect(A2) Inhibition(A3) Facilitation

Mismatch Misretrievals (B) Facilitation (B1) No effect(B2) Facilitation(B3) Inhibition

20 / 53

Inhibition and facilitation arising from SBI and misretrievals

Target−match Target−mismatch

INHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITION

FACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATION

(Consistent with LV05)


INHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITIONINHIBITION

FACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATIONFACILITATION



−300

−100

−50

−20

−10

0

10

20

50

100

300

Reflexives Subj.−verb Num. agr. Reflexives Subj.−verb Num. agr.

Inte

rfe

ren

ce e

ffe

ct in

ms

For an explanation of this variability, see: Engelmann, Jager, &Vasishth, The determinants of retrieval interference in dependencyresolution: Review and computational modeling. Submitted, 2015.

21 / 53

Inhibition and facilitation due to capacity-induced retrievalfailure

I We usually assume that longer reading times index increasedprocessing difficulty and shorter reading times indexreduction in processing difficulty

I What if shorter RTs sometimes indexed increasedprocessing difficulty?

I We explored this question in:Bruno Nicenboim, Pavel Logacev, Carolina Gattei, and ShravanVasishth. When high-capacity readers slow down and low-capacityreaders speed up: Working memory differences in unboundeddependencies. 2016, Frontiers in Psychology, Special Issue onEncoding and Navigating Linguistic Representations in Memory.

22 / 53

Inhibition and facilitation due to capacity-induced retrievalfailureNicenboim et al 2016

We investigated short vs long wh-dependencies in Spanish andGerman (both SPR).

I short Someone asked what the man did last summer.

I long Someone asked what the man [INTERVENER] did lastsummer.

23 / 53


(9) a. short - unbounded dependency

LaThe

hermanayounger

menorsister

deof

SofaSofia

preguntasked

a quinwho.ACC

fuewas

quethat

MaraMara

haba saludadohad greeted

. . .

. . .

b. long - unbounded dependency

SofaSofia

preguntasked

a quinwho.ACC

fuewas

quethat

la hermana menor de Marathe younger sister of Mara


. . .

. . .

24 / 53


(10) a. short - baseline

LaThe

hermanayounger

menorsister

deof

SofaSofia

preguntasked

siif

MaraMaria


. . .

. . .

b. long - baseline

SofaSofia

preguntasked

siif

la hermana menor de Marathe younger sister of Maria


. . .

. . .

25 / 53


I One question was: does working memory capacity affectdependency resolution difficulty?

I The Lewis & Vasishth ACT-R model predicts strongerintervention effects for low capacities compared to highcapacities.

I This prediction turns out to be wrong.

26 / 53


27 / 53


28 / 53


I High capacity readers showed slowdowns due to intervention,but low capacities showed a speed-up (we see a gradedcrossover in the figures).

I Increase in processing difficulty may have different effects as afunction of working memory capacity.

I Intervention may lead to slowdowns in high-capacity readersand speedups in low-capacity ones.

I A computational implementation suggests one possibleexplanation:

I low capacities experience retrieval failure more frequentlyI retrieval failures end faster on average, because items in

memory with an activation below a certain threshold may showshorter latencies due to an early aborting of the retrievalprocess.

29 / 53

Inhibition and facilitation due to predictive processing(surprisal, entropy)Husain, Narayanan, Vasishth 2014

Surprisal claims that intervention sharpens expectations, leading tofaclitation (Levy 2008).But facilitation depends on something we call “expectationstrength”:

Definition of expectation strength

I Strong expectation: Prediction of exact verb (or verb class)

I Weak expectation: Prediction of some unknown verb

30 / 53


I 2 × 2 design.I Predicate type: Complex vs. SimpleI Distance between noun and verb: Long vs. Short

I A sentence completion study ensured that in CP cases, thenominal host predicts the ’exact’ verb with high probability:

I Mean percentage of exact verb predicted in Complex Predicateconditions: 86%.

I Mean percentage of exact verb predicted in Simple Predicateconditions: 17%.

31 / 53


Schematic of design:

(11) a. CP-Long : . . . noun-host 2-3-adjuncts verb . . .

b. CP-Short: . . . noun-host 1-2-adjuncts verb . . .

c. SP-Long : . . . noun.obj 2-3-adjuncts verb . . .

d. SP-Short: . . . noun-obj 1-2-adjuncts verb . . .

32 / 53


(12) a. Complex Predicate Conditions

maa nemother ERG

/ bachche kochild ACC

/ skUlaschool

/ CoRaadropped

/ Oraand

/

usseto her

/ kahaasaid

/ kithat

/ vahshe

/ apnaaher

/ khayaalcare

/

binaa kisi laaparvaahi kewithout any carelessness

/ achCe seproperly

/ rakhe,keep,

/ phirthen

/

vahshe

/ apneher

/ daftar ki oratowards office

/ chal paRiiproceeded

‘The mother dropped the child off at the school and askedher to take care of herself properly without any carelessness,she then proceeded towards her office.’

33 / 53


(13) a. Simple Predicate Conditions

maa nemother ERG

/ bachche kochild ACC

/ skUlaschool

/ CoRaadropped

/ Oraand

/

usseto her

/ kahaasaid

/ kithat

/ vahshe

/ apnaaher

/ gitaarguitar

/

binaa kisi laaparvaahi kewithout any carelessness

/ achCe seproperly

/ rakhe,keep,

/ phirthen

/

vahshe

/ apneher

/ daftar ki oratowards office

/ chal paRiiproceeded

‘The mother dropped the child off at the school and askedher to keep her guitar properly without any carelessness, shethen proceeded towards her office.’

34 / 53


Three alternative predictions:

I (Our version of) Expectation-based processing:I Reading time at the verb in the long conditions (SP-Long and

CP-Long) should be faster than the short conditions.I (New:) In the CP condition, a strong prediction of verb

identity is generated. So, the speed-up in the CP conditionshould be greater than in SP condition.

I Classical interference accounts (Gibson 2005, LV05):In both CP and SP, Long conditions should be slower thanShort.

I Interference and Expectation together:If strong prediction strength modulates the effects of locality,speed-up in CP-Long vs CP-Short, but slowdown in SP-Longvs SP-short.

35 / 53


36 / 53

Inhibition and facilitation due to predictive processingHusain, Narayanan, Vasishth 2014

The results point to a parsing process where strong predictionstrength can reduce or eliminate retrieval cost and consequentlyattenuate locality effects:

37 / 53

Inhibition and facilitation due to predictive processingSafavi, Husain, Vasishth 2016

We attempted to replicate the Hindi results using Persian, whichalso has complex predicates.Expt 1 (SPR), 3 (ET):

(14) a. CP-Short: . . . noun-host PP CP verb . . .

b. CP-Long : . . . noun-host RC+PP CP verb . . .

c. SP-Short: . . . noun-objPP Simple verb . . .

d. SP-Long : . . . noun.obj RC+PP Simple verb . . .

Expt 2 (SPR), 4 (ET):

(15) a. CP-Short: . . . noun-host Short PP CP verb . . .

b. CP-Long : . . . noun-host Long PP CP verb . . .

c. SP-Short: . . . noun-objShort PP Simple verb . . .

d. SP-Long : . . . noun.obj Long PP Simple verb . . .

38 / 53


(16) a. Strong predictability, short distance (PP)

AliAli

a:rezouyeewish-INDEF

bara:yefor

man1.S

karddo-PST

va. . .and. . .

‘Ali made a wish for me and. . . ’

b. Strong predictability, long distance (RC+PP)

AliAli

a:rezouyeewish-INDEF

kethat

besya:ra lot

doost-da:sht-amlike-1.S-PST

bara:yefor

man1.S

karddo-PST

va. . .and. . .

‘Ali made a wish that I liked a lot for me and. . . ’

39 / 53


(17) a. Weak predictability, short distance (PP)

AliAli

shokola:tichocolate-INDEF

bara:yefor

man1.S

xaridbuy-PST

vaand. . .

. . .

‘Ali bought a chocolate for me and . . . ’

b. Weak predictability, long distance (RC+PP)

AliAli

shokola:tichocolate-INDEF

kethat

besya:ra lot

doost-da:sht-amlike-1.S-PST

bara:yefor

man1.S

xaridbuy-PST

va. . .and. . .

‘Ali bought a chocolate that I liked a lot for meand. . . .’

40 / 53


−0.10−0.05

0.000.050.10

dist pred dist:predcomparison

estim

ate experiment

e1e2

Log RT

−0.15−0.10−0.05

0.000.050.10


estim

ate experiment

e3e4

Log FPRT

−0.10.00.1


estim

ate experiment

e3e4

Log RPD

41 / 53


I We also computed entropy using sentence completion data,and investigated whether entropy can explain the interventioneffects.

I Entropy is an information-theoretic measure that essentiallyrepresents how uncertain we are of an outcome (Shannon2001).

I In the present case, this would translate to our uncertaintyabout the upcoming verb.

I If there are n possible ways to continue a sentence, and eachof the possible ways has probability pi , where i = 1, . . . , n,then entropy is defined as −

∑i pi × log2(pi ).

42 / 53


Example:

I Two verbs predicted with equal probability:Entropy= −[0.5 log2(0.5) + 0.5 log2(0.5)] = 1

I Three verbs predicted with equal probability:Entropy=−[0.33 log2(0.33) + 0.33 log2(0.33) + 0.33 log2(0.33)] = 1.6

43 / 53


0

1

2

3

complex simplePredicate Type

Ent

ropy

Distance

short

long

Intervener RC+PP (Expts 1,3)

0

1

2

3

4

complex simplePredicate Type

Ent

ropy

Distance

short

long

Intervener PP (Expts 2, 4)

44 / 53


1. In Experiment 1, in addition to the effects of predictabilityand distance, we find an effect of entropy (coef.=0.05,SE=0.02, t=2.8), and an interaction between distance andentropy (coef.=0.04, SE= 0.02, t= 2.3), such that longdistance conditions lead to a greater effect of entropy.

2. None of the other experiments (Expts 2-4) showed any effectsof entropy.

Thus, we have at least some evidence that entropy could be anexplanation for intervention effects: the reader’s conditionalprobability of the target verb may be getting weaker (cf. Levy2008, Konieczny 2000).Replications of this effect using online sentence completionstudies are needed.

45 / 53


Why does entropy increase in longer-distance dependencies?

I A possible explanation suggests itself in terms of memoryoverload.

I In the long-distance complex predicate condition, participantsmight have forgotten that a noun-verb dependency exists .

I Prediction: participants would tend to produce moreungrammatical continuations in the long-distance conditionthan the short-distance condition.

I This is some evidence for this in experiment 1: the accuracyin the short condition was 0.97 and 0.92 (log odds -.29 [-0.73,0.09]).

The idea of forgetting-causing-entropy is worth investigating in theold locality studies involving English, German, Hindi.

46 / 53

Concluding remarks: My main point in this talk1. Inhibition and facilitation arising from constraints on

cue-based retrievalI Inhibition: Target match conditionsI Facilitation: Target mismatch conditions

The ACT-R model can explain most of these effects.2. Inhibition and facilitation arising from capacity-dependent

retrieval failuresI High capacities show inhibition (classical locality effects)I Low capacities show facilitation (due to fast retrieval failures)

Need special assumptions to explain these.3. Inhibition and facilitation arising from the end-products of

predictive processing (surprisal, entropy)I Inhibition: due to increased uncertainty (due to memory

overload causing forgetting?)I Facilitation: classical expectation-based effects (surprisal)

Information-theory based models cannot explain inhibitioneffects.

47 / 53

Concluding remarks

I No one theory can explain all the phenomena that are robustlyobserved

I That’s because multiple factors contribute to the observedeffects

I Computationally implemented models are crucial fornarrowing down the causal factors for intervention effects, andfor identifying the counter-examples

I Implementations force you to commit to details.

48 / 53

Concluding remarksEmpirical coverage

3. Shravan Vasishth and Richard L. Lewis. Argument-head distanceand processing complexity: Explaining both locality and antilocalityeffects. Language, 82(4):767-794, 2006.

4. Shravan Vasishth, Sven Bruessow, Richard L. Lewis, and HeinerDrenhaus. Processing polarity: How the ungrammatical intrudes onthe grammatical. Cognitive Science, 32(4), 2008.

5. Umesh Patil, Shravan Vasishth, and Richard L. Lewis. Retrievalinterference in syntactic processing: The case of reflexive binding inEnglish. Frontiers in Psychology, 2016.

6. Umesh Patil, Sandra Hanne, Frank Burchert, Ria De Bleser, andShravan Vasishth. A computational evaluation of sentencecomprehension deficits in aphasia. Cognitive Science, 2015.

7. Lena A. Jager, Felix Engelmann, and Shravan Vasishth. Retrievalinterference in reflexive processing: Experimental evidence fromMandarin, and computational modeling. Frontiers in Psychology,6(617), 2015.

49 / 53

Concluding remarksCombining eye-movement control and parsing

Felix Engelmann, Shravan Vasishth, Ralf Engbert, and Reinhold Kliegl. Aframework for modeling the interaction of syntactic processing and eyemovement control. Topics in Cognitive Science, 5(3):452-474, 2013.Code: https://github.com/felixengelmann/act-r-sentence-parser-em

50 / 53

Concluding remarksModel SR/OR asymmetry in English (eyetracking)

Engelmann PhD dissertation:

51 / 53

Concluding remarksModeling agrammatic Broca’s aphasic patients (visual world data)

Umesh Patil, Sandra Hanne, Frank Burchert, Ria De Bleser, and ShravanVasishth. A computational evaluation of sentence comprehension deficitsin aphasia. Cognitive Science, 2015.

Code: http://cogsci.uni-osnabrueck.de/ upatil/src/Patil-EtAl-2014-

AphasiaModels.zip

52 / 53

Flavia’s questions

I How could current syntactic theory and memory retrievalmechanisms be used to explain how sentence complexity isacquired and processed?

I What are the fundamental divergences and/or points ofconvergence between the two approaches?

I Is there a relation to complexity issues from other linguisticdomains, such as the lexicon, semantic, pragmatic and/orphonology?

53 / 53

Shravan Vasishth Universit at Potsdam, Germany vasishth ...vasishth/pdfs/Konstanz2016.pdfBruno Nicenboim, Pavel Loga cev, Carolina Gattei, and Shravan Vasishth. When high-capacity

Documents