goldberg feb ‘07 georgetown linguistics colloquiumfaculty.georgetown.edu/tyleran/adele...
TRANSCRIPT
Goldberg Feb ‘07Georgetown Linguistics Colloquium
1
1
Adele GoldbergPrinceton University
Georgetown LinguisticsColloquium
February 27, 2007
2
“Splitters see very small, highly differentiated units–their critics say that if they can tell two animals apart,they place them in different genera, and if they cannottell them apart, they place them in different species….
Lumpers, on the other hand, see only large units - theircritics say that if a carnivore is neither a dog nor abear….they call it a cat” (Simpson, 1945).
3
We clear store a massive amount of quitespecific, idiosyncratic information
4
Adjective tests:can appear premominallyafter verbs like be & seem:
The red car. The car seemed red.
5
But then what are mere, aghast?
The mere child. ??The child seems mere.
??The aghast man. The man seemed aghast.
6
Rather selects (non-locally) for would I’d rather have Thai or Indian. *I rather have Thai or Indian. *I will rather have Thai or Indian.(cf. with normal verb like prefer)
Many words have idiosyncratic properties: e.g., letalone, notwithstanding, respectively, each,occasional, need, dare, over, in, break open, ago,once, neither-nor, no matter, whether…
(e.g., Culicover 1999)
Goldberg Feb ‘07Georgetown Linguistics Colloquium
2
7
Construction Label example
Mad Magazine construction Him, a doctor?!
N P N construction house by house; day after day
Time away construction Twistin the night away
GoVPing Don’t go sticking your nose…
[V] this! Construction Gate-check this!
Enough Already Construction Enough with the examples!
Fillmore, Kay and O’Connor 1988; Culicover 1999; Jackendoff 2003;Williams 1994;2006; Zwicky 1974, Lambrecht 1994, Goldberg 2006… 8
And yet it would be a mistake to concludethat we do not form generalizations.
9
Constructionist approaches
Knowledge of language consists of learned, form-function correspondences at varying levels ofabstraction.
E.g., Fillmore, Kay, O’Connor 1988; Langacker 1987; 1991; Gleitman1994; Lambrecht 1994; Goldberg 1995; 2003; 2006; Culicover 1999;Kay 2001; Croft 2001; Ninio 2001; Ruppenhofer and Michaelis 2001;Jackendoff 2002; Booij 2002; Culicover and Jackendoff 2005;Tomasello 2003; Pinker and Jackendoff 2005; Saffran and Glenberg2005; Bender and Sag 2004….
10
She daxed him something.
What does dax mean?
11
Argument structure CONSTRUCTIONS
Subj V Obj Obj2She daxed him something.
X causes Y to receive Z= “give”
Subj V Obj RPHe made her crazy.
X causes Y to become Z
Subj V Obj PPShe put the ball in the box.
X causes Y to move Z
Subj V PPShe went down the street.
X moves (to) Y
FormExample
Meaning
12
Argument structure CONSTRUCTIONS
Subj V Obj Obj2She shinned him something.
X causes Y to receive Z
Subj V Obj RPHe drank himself silly.
X causes Y to become Z
Subj V Obj PPShe sneezed her tooth across town.
X causes Y to move Z
Subj V PPThe ice cream truck jingled downthe street.
X moves (to) Y
FormExample
Meaning
Goldberg Feb ‘07Georgetown Linguistics Colloquium
3
13
argument structure constructions≈“linking rules”
14
…argument structure constructions ≈ “linking rules”
Notice that many linking rules are construction-specific
theme -> direct object (in the transitive construction) She kicked the ball.
theme -> subject (in intransitive construction) The ball rolled.
theme -> second object (in ditransitive) She kicked him the ball.
theme --> prepositional phrase (in “locative” construction) She loaded the truck with the balls.
15
…argument structure constructions ≈ “linking rules”
… many linking rules are construction-specific
recipient -> subject (in the transitive construction) She received it.
recipient -> first object (in ditransitive) He gave her a letter.
recipient --> prepositional phrase (in “dative” construction) She gave the letter to her.
16
Is it possible to learn new constructions withoutexplicit training or feedback?
17
Rules linking, e.g., agent to subject are “near-universal in their essential aspects and thereforemay not be learned at all” (Pinker 1989: 248)
“there is sufficient cross-linguistic similarity in theselinking rules to get the learning procedurestarted...” (Naigles et al. 1993)
Preferential looking results have indicated thatlinking rules might in place as young as 20-30months, the implication being that they are notlearned (Fernandez et al. 2006; Gertner and Fisher 2006)
18
Learning a novel construction:
Experiments designed to test whether anovel construction can be generalizedwithout explicit instruction.
(Casenhiser and Goldberg, Developmental Science. 2005)
Goldberg Feb ‘07Georgetown Linguistics Colloquium
4
19
Form: Subj Obj V-o Meaning: theme APPEARS in location
Example:
“The frog the chair moopo-ed.” Video: the frog appears on the chair.
20
Training condition: witnessed 16 instancesof novel construction with 5 novel verbs
(4-4-4-2-2)
Control condition: watched same 16 videoclips without sound
21
Test: forced choice
Determine which scene a sentencecorresponds to:
Scene #1: scene of appearanceScene #2: matched foil scene
22
Test1 Test2Novel instances of the new construction
(involve NEW novel verbs; NEW scenes)
23
Comparison of two conditions (mean age 6;4, n=34)(Casenhiser & Goldberg 2005, Dev. Sci.)
*
Children were able to learn novel construction with 3minutes of training
24
Learning is not likely an effect ofpreexisting knowledge:
Novel construction violates a proposeduniversal Locative phrase -> oblique (PP in English)
Construction relating “appearance” meaningto unusual verb position is unusual.
Goldberg Feb ‘07Georgetown Linguistics Colloquium
5
25
Morphology is not necessary (mean age 6;3; N=48)(Casenhiser & Goldberg 2005)
*
26
skewed frequency training condition: 8-2-2-2-2(Balanced) training condition: 4-4-4-2-2Control: watched video without sound
Controlled for overall token/type frequency:Total # of scenes: 16Type frequency (number of novel verbs): 5
All three conditions watched exactly the same video
27
Actual input children receive…
28
Verb tokens in mother’s utterances
Percentage of tokensof various verbs inthe caused motionconstruction:
why
put(40%)do
have
take
get
read
see
pick
set
throwSlice11Slice12Slice13Slice14Slice15Slice16Slice17Slice18Slice19Slice20Slice21Slice22Slice23Slice24Slice25Slice26Slice27Slice28Slice29Slice30Slice31Slice32Slice33Slice34Slice35Slice36Slice37Slice38Slice39Slice40
Subj V Obj PP(She Ved it on the table)
29
One verb often does account for the lion’s share of tokens
8 verbs40% think(Kidd et al. to appear)
Subj V Scomp
> 13 verbs44% give (226/517)Switchboard (Bresnan and Nikitinato appear)
Subj V Obj Obj2Recall dax
39 verbs39% go (136/353)(Bates et al. 1988 corpus)
Subj V Oblique
Total # of verbtypes
Corpus dataConstruction
30
skewed frequency training condition: 8-2-2-2-2(Balanced) training condition: 4-4-4-2-2Control: watched video without sound
Controlled for overall token/type frequency:Total # of scenes: 16Type frequency (number of novel verbs): 5
All three conditions watched exactly the same video
Goldberg Feb ‘07Georgetown Linguistics Colloquium
6
31
Results for all three conditions (mean age 6;4, n=51)
**
*
The existence of a single high frequency exemplar facilitatesthe learning of the novel construction
32
Similar results for adults (n = 81)(Goldberg, Casenhiser and Sethuraman 2004, Cog. Ling.)
Comparison with kids
****
33
Idiomatic constructions also often have high-frequencyexemplars:
The more the merrierThe Xer the Yer
Screw that![V] this!
Dancin’ the night awayTime away
house by house; day after dayN P N construction
Idiomatic expressionConstruction
34
Possible concerns…
No-sound control too weak? Do subjects attend to the word order of
novel construction?
Younger than 6?
35
We know language can focus attention andencourage generalization…
36
Additional control condition:
2 nouns heard during training phase: sun…sky box…frog
Goldberg Feb ‘07Georgetown Linguistics Colloquium
7
37
Additional items at test: novel transitives(as well as novel instances of the new construction)
6 Novel instances of the new construction (involve new novel verbs; new scenes)
6 Novel instances of familiar transitive construction (also involve new novel verbs; new scenes)
Important to test to know that subjects are attending to theform of the novel construction.
38
Control conditions (2Noun & no-sound)(undergraduates, N=54)
Goldberg,Casenhiser and White (to appear, New Ideas in Psychology)
New (2N) control condition does not differ significantly from no-soundcontrol: performance on novel construction at chance; performance ontransitive construction, as expected, above chance.
39
Accuracy on novel construction(undergraduates, N=54)
And successfully assign the appearance meaning only to thenovel construction.
40
Four year olds (in progress)
Comparison of control conditions(mean age 4;5, N=20)
Four year olds likewise are (so far) at chance in both controlconditions; are, as expected, above chance on noveltransitives
41
Accuracy on novel construction(mean age 4;5, N=30)
And successfully assign the appearance meaning only to thenovel construction.
42
Many open questions remain….
Just how detailed and how robust isknowledge of the novel construction?
What role does type frequency play?
Goldberg Feb ‘07Georgetown Linguistics Colloquium
8
43
Novel construction can be learned and learnedquickly by children as young as 4;6.
More exemplars with same novel verb intraining—leads to better accuracy at test.
Actual input children receive is tailor-made injust this way.
44
Novel construction can be distinguishedfrom novel transitives.
45
But if linking rules are universal, maybe theconstructions are learnable but not learned…
46
Is it possible to account for cross-linguisticgeneralizations by appealing toindependently needed cognitiveprocesses? (cf. also Hauser, Chomsky and Fitch2002)
47
Proposed universals related toARGUMENT STRUCTURE
48
if there’s a subject and an object, and if there’s an ACTOR and an UNDERGOER
then ACTOR -> subject; UNDERGOER -> object,
except when they’re linked the opposite way, incertain (syntactically ergative) languages.
Proposed Cross-Linguistic UniversalsE.g., Dowty (1991):
Goldberg Feb ‘07Georgetown Linguistics Colloquium
9
49
Dowty: relatively weak claim
Oversimplified account of ergativity Yidiny is syntactically ergative with nominals;
syntactically accusative with pronouns (Dixon 1979)
Also, what counts as “subject” “object” differscross-linguistically (Fried 1993; Morris 1997; Croft 2001;Barðdal 2005)
50
Reformulation of Dowty’s generalization:
Actors and undergoers tend to be expressed inprominent slots
51
“conceptual accessibility” of actors (Bockand Warren 1985; Bock, Loebel and Morley 1992; Teufel and Feleki1996; Keil 1979).
undergoers/endpoints are also salientand closely attended to (Robertson and Suci 1980;Woodward 1998, 1999; Csibra et al. 1999; Gergely et al 1995;Jovanovic et al. to appear; Lakusta & Landau to appear; Regier 2003)
52
Actors are salient--Visual attention tends to be centered on the actor
in an event (Robertson and Suci, 1980).
--Agent bias (chase vs flee) (Fisher et al. 1994)
--9 month olds: attribute intentional behavior to eveninanimate objects (Csibra et al. 1999)
--16 month olds: distinguish intentional vs accidentalactions (Carpenter et al.1998).
53
Undergoers are salient--Easier to discriminate between events that have distinct
endpoints than distinct onsets (Regier and Zheng 2003)
-- 6 month olds attend more to changes of state than tochanges of motion without corresponding state-change(Woodward 1998; 1999)
--subjects use a wider range of more specific verbs todescribe endpoint-focused actions than onset-focusedactions (Landau, 2003).
--Eng and Fr speakers are more likely to mention goal-directed paths of motion than atelic paths whendescribing video clips (Pourcel, 2004).
54
Reformulation of Dowty’s generalization:
Actors and undergoers tend to be expressed inprominent slots
********************************************************Tendency is explained by the fact that we attend to actors
and undergoers.
Particular constructions allow for exceptions (e.g., passive)
Goldberg Feb ‘07Georgetown Linguistics Colloquium
10
55
Another generalization:
# of arguments expressed=
# of semantic arguments
The isomorphic mapping principle
(Lidz et al. 2003)
56
Examples of general tendency in English
Subj V Obj Obj2X Y Z
X causes Y to receive Z
Subj V Obj RPX Y Z
X causes Y to become Z
Subj V Obj PPX Y Z
X causes Y to move Z
Subj V PPX Y
X moves (to) Y
FormMeaning
# of arguments expressed = # of semantic arguments
57
Do we need a generalization that isspecific to language?
58
Grice (1975): Maxim of Quantity: Say as much, and onlyas much, as is needed for the communicative goal.
: Pragmatic assumption in all kinds of linguistic and non-linguistic communicative acts.
(cf. also Paul 1889; Zipf 1935; Horn 1984)
59
Pragmatic Mapping Generalizations(Goldberg, 2004, Cognition)
A) The arguments that are expressed are interpreted to berelevant to the message being conveyed.
B) Any semantic arguments in the event being conveyed thatare relevant and non-recoverable from context must beovertly indicated.
60
…Pragmatic generalization
Expressed Relevant Relevant & Non-recoverable Identifiable
Goldberg Feb ‘07Georgetown Linguistics Colloquium
11
61
Pragmatic generalizations say nothing aboutarguments that are recoverable or irrelevant.
In fact, languages and constructions withinlanguages treat these arguments variably…
62
Recoverable arguments are typically omittedRecoverable arguments are typically omittedcross-linguisticallycross-linguistically
JapaneseJapanese
[] [] [] [] eiga eiga ni ni sasoimasitasasoimasita.. movie movie to to invited invited ““(He) invited (him) to a movie.(He) invited (him) to a movie.””
(Walker, Iida, Cote 1993)(Walker, Iida, Cote 1993)
KoreanKorean, , ChineseChinese, Hindi, Hungarian, Kannada,, Hindi, Hungarian, Kannada,LaosLaos……
63
2: (the tiger, the tiger’s prey) 1
The deprofiled objectconstruction(e.g., The tiger killed again)
2: (Pat, Pat’s killer) 1Short Passives(e.g., Pat was killed)
# semantic arguments# argumentsexpressed
[
Particular constructions allow for recoverableand/or irrelevant arguments to be omitted, even inEnglish:
64
3: (Pat, the toast, butter/spread) 2
Semantic “incorporation”constructions(e.g., Pat buttered the toast)
# semantic arguments# argumentsexpressed
[
Pragmatic Mapping principle also allows forarguments to be identified without appearing as anovert complement:
65
So Isomorphic Mapping principle does nothold, but Pragmatic Mappinggeneralizations do.
66
…Proposed universals of argument structure
Dowty’s linking generalizations explained by attentional properties of
humans
Tendency for # of args to equal # ofcomplements explained by general Gricean pragmatics
Goldberg Feb ‘07Georgetown Linguistics Colloquium
12
67
Constraining generalizationsStatistical preemptionDegree of openness
68
Other generalizations:
Argument per Subevent Principle General cognitive motivation for tendency
Subj-Aux inversion Island constraints Distributional facts about resultatives, ditransitives Inclusive interpretation of “or” within conditionals
69
…Why constructions are learned
Constructions are good predictors of overallsentence meaning
Sorting task Bencini & Goldberg Journal of Memory and Language
2000; Gries and Wulff 2004 Corpus analysis
(Goldberg, Casenhiser, Sethuraman Journal of ChildLanguage, 2005)
Constructions are primed in production
70
Argument structure constructions (“linking rules”)can be learned quickly
Skewed input (what children actually receive)facilitates construction-learning
(Constructions are good predictors of sentence-meaning) (Indirect negative evidence) (Memory for language)
Conclusions…
71
Cross-linguistic generalizations may stemfrom:general cognitive (e.g., attentional) biasespragmatic (e.g., Gricean) effects, (or may follow from semantic/pragmatic facts
about the constructions involved)
Conclusions…
72
Devin Casenhiser
Tiffani White Dave Harris Lisa Glukhovsky
NSF