1 Bidirectional optimization from the perspective of experimental pragmatics Reinhard Blutner Universiteit van Amsterdam June 11, 2007 ∙ ZAS Berlin
1
Bidirectional optimization from the perspective of experimental pragmatics
Reinhard BlutnerUniversiteit van Amsterdam
June 11, 2007 ∙ ZAS Berlin
0 Introduction
“With four parameters I can fit an elephant, and with five I can make him wiggle his trunk”
(John von Neumann)
3
Grice and His Followers
Relevance Theory
Presumptive Meanings
Neo-Gricean
Theories (Horn, Atlas)
OT-Pragmatics
4
Experimental Pragmatics
“Properly devised experimental evidence can be highly pertinent to the discussion of pragmatic issues, and pragmatics might greatly benefit from becoming familiar with relevant experimental work and from contributing to it ”(Noveck & Sperber 2007, p. 210)
5
Optimality Theoretic Pragmatics
Like generally in OT: no artificial separation between competence and performance– NL comprehension as interpretive optimization
– NL production as expressive optimization
Open issues: – Are the two optimization processes integrated
with each other (bidirectional optimization)?
– Are there asymmetries between comprehension and production?
– The role of fossilization?
6
The Idea of Fossilization
‘Invited Inferences’ (Geis & Zwicky 1971). Mechanism of conventionalization for implicatures
Short-circuited implicatures (Morgan 1978; Horn & Bayer 1984)
Lexicalization (Cole 1975) Traugott (1989…2005) applied the idea to
explain language change (conventionalization and language change)
Levinson (2000) und Mattausch (2004) used the idea for explaining the development of binding principles.
7
Outline
1. Pronouns/Reflexives with children and adults
1.1 Hendriks & Spenader‘s bidirectional processing account
1.2 The fossilization account
2. R-expressions/Pronouns with young and elderly adults
3. All/Some (Scalar Implicatures) with children and adults
4. Conclusions
1 Pronouns/Reflexives
In an important recent article Hendriks and Spenader (2004) give a new interpretation of children‘s delay of the comprehension of pronouns. I discuss the validity of this interpretation and present an alternative account in terms of iterated learning
9
The Pronoun Interpretation Problem
(1) Bert saw himself(2) Bert saw him
Children correctly interpret reflexives like adults from the age of 3;0 but they continue to perform poorly on the interpretation of pronouns even up to the age of 6;6 (50 % errors)
E.g. Jakubowicz (1984); Koster and Koster (1986); Chien and Wexler (1990); McDaniel, Smith Cairns and Hsu (1990); McDaniel and Maxfield (1992).
10
Children‘s Production of Pronouns and Reflexives
(3) I hit myself.
(4) John hit me
(5) * I hit me.
Bloom et al. (1994): Even in the youngest age groups investigated (ranging from 2;3 or 2;4 to 3;10), the children consistently used the pronoun me to express a disjoint meaning (99.8% correct), while they used the reflexive myself to express a coreferential interpretation (93.5% correct).
Conclusion: very young children have competence in binding principles.
11
The problem
Usually, comprehension of a given form precedes production of this form
– Bates, Dale and Thal 1995; Benedict 1979; Clark 1993; Fraser, Bellugi and Brown 1963; Goldin-Meadow, Seligman and Gelman 1976; Layton and Stick 1979.
Thus how do we reconcile children’s poor performance on comprehension tasks with their near-perfect production data?
12
Previous accounts
Reject the comprehension data (Bloom et a. 1994)– the tasks used in the comprehension experiments
did not adequately test children’s grammatical competence
Dissociation between a comprehension grammar and a production grammar.– requires some ad hoc stipulations
Revise the binding principles, making a distinction between coindexation and coreference (Chien and Wexler 1990; Grodzinsky and Reinhart 1993). – This is based on the observation that children seem
to correctly interpret pronouns in the scope of quantified noun phrases.
1.1 Hendriks & Spenader‘s account
“Our account, formulated in the framework of Optimality Theory handles the comprehension data as well as the production data by arguing that children acquire the ability to reason about alternatives available to other conversation participants relatively late. It is this type of bidirectional reasoning, we argue, that is necessary for correctly interpreting pronouns.” (H&S 2004)
14
Optimality Theory as a Framework
Contraint-Hierarchy:C1 >> C2 >> C3 Evaluator
Output
Input
Generator
1 2 3 4 5Candidates
15
Unidirectional OT
Consider two directions of optimization (Hearer-oriented, Speaker-oriented)
Use the same set of constraints and the same ranking for both perspectives
Hence, the evaluator evaluates pairs of representations (e.g. form-meaning pairs)
16
Constraints(slidely modifying Burzio 1998)
PRINCIPLE A: A reflexive must be bound locally REFERENTIAL ECONOMY:
Avoid R-expressions >> Avoid pronouns >> Avoid reflexives
proself
disj conj
proself
disj conj
PRINCIPLE A REFERENTIAL ECONOMY
17
Assuming a Ranking
PRINCIPLE A >> REFERENTIAL ECONOMY
Hearer‘s perspective: one optimal interpretation for self but two optimal interpretations for pro.
Speaker‘s perspective: correct unique form for each interpretation.
proself
disj conj
pro
self
disj
conj
18
Bidirectional OT
Strong bidirection requires that an interpretation is optimal for a form (interpretive optimization) and that a form is optimal for an interpretation (expressive interpretation)
In the example there is only one bidirectionally optimal form-meaning pair but two optimal interpretation pairs: (pro, disj) and (self, disj)pro
self
disj conj
19
Delayed Bidirection
The proposal is that children begin with unidirectional optimization, and only later acquire the ability to optimize bidirectionally.
A child must, when hearing a pronoun, reason about what other non-expressed forms the speaker could have used, compare the interpretation associated with the pronoun and realize that a coreferential meaning is better expressed with a reflexive. Then, by a process of elimination, the child must realize the pronoun should be interpreted as disjoint.
Optimizing bidirectionally inherently involves reasoning about alternatives not present in the current situation, which may be a skill acquired very late, thus explaining the lag in acquisition.
20
Delayed Bidirection
Unidirectional Optimization
Bidirectional Optimization
What‘s essential for this solution is that the hearer has to take a potential speaker into account
proself
disj conj
pro
self
disj
conj
21
Advantages
The authors are able to derive Principle B effects from Principle A alone, through bidirectional optimization.
The analysis clearly distinguishes the task of a speaker from the task of a hearer. As a result the analysis is able to model different results for production and comprehension.
Besides the stipulation of the constraints and their ranking no other stipulations are required
The approach nicely combines a pragmatic explanation with a processing account (lack of processing resourses)
22
Disadvantages
The constraints are partly stipulated - no constraint grounding
Theory of Mind (Perner, Leekam and Wimmer 1987) requires awareness of other conversation participant’s choices. Hence, theory of mind is based on controlled rather than automatic processing. However, the effects of pronoun processing are automatic rather than controlled. There is no explicit hint for mind reading capacities in such tasks
1.2 A Reinterpretation in Terms of
Learning/Fossilization
I will propose a reinterpretation of the Hendriks/Spenader account based on the idea that the ranked system of constraints is changed during learning.Rather than stipulating a change from unidirectional to bidirectional processing I account for the effects of (weak) bidirection by changing the constraint ranking.
24
Measuring the Success of Communication
Speaker‘s strategy: given the possible utterance meanings m, the OT system specifies a function S(m)
Hearer‘s strategy: given the possible language forms F, the OT system specifies a function H(F)
1 if m = H(S(m)) U(S,H,m) =
0 elsewhere EU(S,H) = P(mi) U(S,H,mi)
25
Learning as utility optimization
Learning consists in improving the value of expected utility.
In OT-learning theories the ranking of a given system of constraints is (stepwise) changed
Learning leads to a stable outcome if the relevant EU(s) reach its maximum value
26
Very Simple Algorithm
m f m’Speaker
Hearer
m = m’ ?
If yes, nothing happens
If no, adjustment:
All constraints that favour (f, m) over (f, m’) are promoted
All constraints that favour (f, m’) over (f, m) are demoted
27
Application: Fossilization Principle A: self conjoint Referential Economy: self >> pro Principle B: pro disjoint; ….
proself
disj conj
pro
self
Constraint B strengthened
disj
conj
conj
Speaker sel
f
Hearernothing happens
conjconj
Speaker pro
Hearernothing happens
disjdisj
28
More Natural Constraints
Bias Constraints– PRINCIPLE A: A reflexive must be bound locally
– PRINCIPLE B: A pronominal is free (in its governing cat)
Markedness Constraints– DISJOINT REFERENCE: disj > conj
– EXPRESSIVE ECONOMY: pro > selfproself
disj conj
proself
disj conj Bias
ConstraintsMarkedness Constraints
AB
29
Why delayed principle B?
Teacher proself
disj conj
pro
selfdisj conj
AB Learner
pro self
disj
conj
A
proself
disj conj
B
L0
L1: A-first
(ambiguous pro
L1: B-first
(ambiguous self
0.1 0.2 0.3 0.4 0.5Prob for Conj
0.6
0.7
0.8
0.9
1
EU
A-first
B-first
Initial State
30
The Final Step
Teacher
proself
disj conj
proself
disj conj
AB
proself
disj conj
A
L0
L1: A-first
L2
proself
disj conj
AB
31
Conclusions
Two different views– Processing account (unidirectional vs. bidirectional
processing)
– Fossilization account (applying OT learning theory). This view is memory-based
Conceptual advantages of the fossilization account
Two kinds of fossilization– Individual fossilization via learning/automatication
on an ontogenetic time scales (seconds-years)– Cultural fossilization via iterated learning / cultural
evolution on a historical time scale (years-centuries)
Meanings are partly conventionalized within speech communities and partly negotiated anew during each individual interaction (Traugott & Dasher 2002)
32
Explaining Universals
Functional Formal
Genetic evolution
Evolutionary Psychology
(Pinker)
Minimalist program
(Chomsky)
Cultural evolution
Recruitment theory (Steels)
Iterated learning (Kirby, Hurford,
Zuidema)
2 R-Expressions/Pronouns
A recent article by Petra Hendriks, Christina Englert, & Ellis Wubs * investigates whether choosing the appropriate referring expression requires taking into account the hearer’s perspective, as is predicted under some versions of bidirectional OT but is unexpected under other versions.
* Age differences in adults’ use of referring expressions (unpublished manuscript, University
of Groningen 2007)
34
Sentence Elicitation Study
A woman hold-ing an ice cream cone is walking past a road sign.
The woman comes across a girl.
She gives the girl an ice cream cone.
The girl is eating from the ice cream cone.
Well, the woman passes again an ice cream van.
The woman buys another ice cream come.
Topic shift Target Picture
she
35
Main Results
36
Discussion Elderly adults produce (non-recoverable) pronouns
significantly more often than young adults when referring to the old topic in the presence of a new topic.
With respect to the comprehension task, no significant differences were found between elderly and young adults.
These results support the hypothesis that speakers optimize bidirectionally and take into account hearers when selecting a referring expression.
If the use of a pronoun will lead to an unintended interpretation by the hearer, the speaker will use an unambiguous definite noun phrase instead.
Because elderly adults are more limited in their processing capacities as speakers they will not always be able to reason about the hearer’s choices.
37
Constraints
PROTOP: Pronouns refer to topics REFERENTIAL ECONOMY *
Avoid R-expressions >> Avoid pronouns
R
proN-Top
Top
Rpro
N-Top Top
PROTOPIC REFERENTIAL ECONOMY
* REFERENTIAL ECONOMY = EXPRESSIVE ECONOMY in the case of comparing pronouns & R-expressions
38
A bidirectional optimization account that fits the observations
REFERENTIAL ECONOMY >> PRO TOP
Hearer‘s perspective: one optimal interpretation for pro and one R-dependent optimal interpretations for R-expression (not represented!)
Speaker‘s perspective: pro is the the optimal form for both the Top and the N-Top interpretation
Bidirectional Optimization: Speaker choses pro for Top and a R-expressions for N-Top.
Rpro
N-Top
Top
R
pro
N-Top
Top
39
Discussion
Fits nicely with the earlier approach for pro/self Prediction for elicitation studies: elderly adults
behave similar to children ☺ Prediction for pronoun/reflexive interpretation
studies: elderly adults behave similar to children ☹
Conceptual problem: What is the motivation for the ‘inverted ranking’ REFERENTIAL ECONOMY >> PRO TOP ??
(In the paper, Hendriks et al. wrongly assume the earlier ranking PRO TOP >> REFERENTIAL ECONOMY
which doesn’t fit their data)
40
A fossilization account that fits the observations
EXPRESSIVE ECONOMY, TOPIC
Child can not learn anything if it takes the listener‘s role only!
Rpro
N-Top
Top
R
pro
N-Top
TopSpeaker pr
o
Hearernothing happens
TopTop
Speaker R
Hearernothing happens
TopN-Top
41
EXPRESSIVE ECONOMY , TOPIC
Rpro
N-Top
Top
R
pro
N-Top
Top
PRO TOPstrengthened
Speaker pr
o
Hearernothing happens
TopTop
Speaker pro
HearerTopN-Top
A fossilization account that fits the observations
R
42
Discussion
The fossilization accounts is able to describe the difference between children and adults
However, it predicts that elderly adults behave similarly to younger adults in case of R-expressions/pronouns
Wrong prediction! They should behave like children. Defossilization doesn‘t make any sense here.
3 All/Some (Scalar Implicatures)
Recent experimental work by Noveck & Sperber investigates the case of scalar implicatures. Their experimental method has helped sharpen a theoretical debate and has provided uniquely relevant evidence.
44
Some and all
Experimental Pragmatics: Noveck u.a. – Some elephants live in the zoo (appropriate) yes 90% 99%
– All elephants live in the zoo (inappropriate) no 99% 99%
– Some elephants have trunks (inappropriate) yes 85% 41%
– All elephants have trunks (appropriate) yes 99% 96%
– Some elephants have wings (absurd) no 99% 98%
– All elephants have wings (absurd) no 99% 99%
Why do children sometimes think more logical than adults?
Adults10-11
45
Potential Answers
RT (see Noveck)
Chierchia’s defaults
OT pragmatics has two potential answers
(1) Metalinguistic ability for perspective changing (bidirectional reasoning) not yet developed
(2) Fossilization not yet progressed
RT (see Noveck)
Chierchia’s defaults
OT pragmatics has two potential answers
(1) Metalinguistic ability for perspective changing (bidirectional reasoning) not yet developed
(2) Fossilization not yet progressed
46
The implementation of answer 1
Lexical Constraint A: all Set-inclusion Strength: all >> some
Bidirectional Solutions
someall
some
all
47
Answer 2: Fossilization Lexical Constraint A: all Set-inclusion Strength: all >> some Potential lexical Constraint B: some Set-intersection;
….someall
some
all
Speaker all
Hearernothing happens
Constraint B strengthened
Speaker som
e
Hearernothing happens
48
Comparing two cases of blocking
proself
disj conj
pro
self
disj
conj
someall
some
all
7 years old
12 years old
49
Emergence of Bidirection or Fossilization?
According to the solution of evolutionary psychology (processing account) the crucial developmental stages should appear synchronously for the different domains
According to the fossilization solution (iterated learning) the time course of the development is not necessarily synchronized but may crucially depends on factors of frequency and other use factors
The processing view predicts similarities between the behavior of children and elderly adults
– True for the production of (non-recoverable) pronouns – False for interpretation of pronouns
– Unclear for scalar implicatures
4 General Conclusions
The idea of fossilization as a starting point for resolving puzzles in experimental pragmatics
- Some elephants have a trunk: why children sometimes think more logical than adults (Noveck)
- The acquisition of binding principles: why children sometimes misinterpret pronouns while correctly producing them (Hendriks & Spenader)
- Production of (non-recoverable) pronouns when referring to the old topic in the presence of a new topic (Hendriks, Englert, & Wubs)
51
Final Scores: Comprehension/Generation
Bidir. Processing
(Hendriks et al.)
Asymmetric OT
(Zeevat)
OT with
Fossilization
Pronouns & Refl
Children
Young adults
Elderly adults
+/+
+/+
−/+
−/+
+/+
+/+
+/+
+/+
+/+
R-Exp & ReflChildren
Young adults
Elderly adults
+/+
+/+
+/+
+/+
+/+
+/+
+/+
+/+
−/+
All & SomeChildren
Young adults
Elderly adults
+/+
+/+
?/+
+/+
+/+
?/+
+/+
+/+
?/+
52
Possible Solution
Combining the idea of Fossilization with asymmetric OT
In asymmetric OT the speaker takes the listener into account but not vice versa
We need independent motivation for that. At the moment it’s a data fitting only!
With four parameters I can fit an elephant, and with five I can make him wiggle his trunk