Evolution of Universal Grammar
Pia Göser
Universität Tübingen Seminar: Sprachevolution
Dozent: Prof. Jäger
11.02.2010
Structure
• Historical outline
• Arguments for Universal Grammar
– Language and Grammar – Learning Theory
– Language Acquisition
• Evolutionary Processes
– Language
– Universal Grammar
Universal Grammar
Brief history of a theory...
• Basic: Researches on language acquisition
• From ‘20s: behavioristic approach
Problem: „poverty of stimulus“
• ‘60s/‘70s: Alternative model: nativistic approach
Noam Chomsky: innate human mechanism
• Still controversial topic
mathematical approach tries to explain it
What is language?
Formal language theory:
• Generative system
• Set of sentences
– A sentence is a string of symbols
– there are infinite many sentences (countable)
• Finite Languages (are infinite – countable)
• Infinite Languages (are infinite – uncountable)
set of languages is uncountable
Biologically: extended phenotype of population
What is language?
What is grammar?
• Finite list of rules specifying a language
• There are infinitely many grammars (countable)
only small subset of languages can be described by a grammar (=computable languages)
What is grammar?
Relationship language-grammar
Languages, grammars and machines
Chomsky hierarchy of formal grammar automata
Phrase structure (Unrestricted) Turing machine;
description for computable languages with unrestricted rewrite-rules
Context sensitive Linear-bounded;
Turing machine can decide every sentence’s belonging to the language; TM has an infinite memory
Context free Push-down;
language can be described with computers with only one memory task
Regular (Finite State) Finite-State;
generate regular languages; subset of regular languages contains all finite languages
Learning Theory
Classical Learning Theory (Gold)
assumptions:
a) learner has to identify the target language
b) learner receives only positive examples
c)learner has access to an arbitrary large number of examples
d) learner is not limited by any consideration of computational complexity
Gold’s theorem: no algorithm can learn the set of all regular language; only by memorization
Statistical learning theory
assumptions:
a) learner will come very close to the right language with a high probability b) learner receives both positive and negative examples (distribution P) c) after a Number of “empirical data”
the learner guesses a language out of a set of languages
Theorem:
-> set of all regular and finite languages cannot be learnt
-> subsets of regular languages with finite-state automata can be learnt Basically: generalizing rules beyond one’s own experience
Learning Theory
• no learning theory can permit all languages to be
learnable
necessity of specific resrictions
innate “restricted search space” for languages
Learning a language
• Paradox of language acquisition
• Poverty of stimulus
possible answer: UG
argument: there is a learning algorithm lim A (T
N) =L
N →∞
Approaches:
a) principles and parameters theory (Chomsky)
b) optimality theory
Chomsky Hierarchy
UG as a restricted search space:
Natural languages
• finite state grammar is not capable of representing “if…then”-sentences
• at least context-free grammars are necessary
•no phrase structure (is unrestricted)
Summary: Arguments for UG
• Language acquisition
– Paradox of Language acquisition
• Learning theory:
– necessity of restricted search space
Human brain contains of an algorithm that can learn grammar there is no algorithm that can learn an unrestricted set of
grammars Human brain can only learn a certain subset of
all possible languages The theory of this subset is UG.
Cultural Evolution of Language
starting position:
population of individuals with the same UG UG specifies finite number of languages L
1,…,L
neach individual uses particular language
selective aspect:
successful communication (coherent language) results in pay-off
increased fitness
Variation:
Offspring inherit a mechanism to learn language and UG (mutation)
Offspring use this mechanism to learn the language of their parents etc
(mistakes)
Cultural Evolution of Language
Model: Two individuals communicating by L
iand L
jComponents:
Equation:
Communicative pay-off for user of Li Fij Relative abundance of speakers of Li xi Learning Matrix (Probability of an offspring
speaking Lj with Li parents) Qij
Fitness of Li n
fi = ∑xjFij
j=1
Average fitness (grammatical coherence) of population
�(x)=∑ifi(x)xi total population size is consant: ∑ixi=1 - (x) � xj
Evolution of UG
starting position:
population of individuals with different UGs (U
I… U
M) each individual uses particular U
each U
Iadmits a subset of n grammars
selective aspect:
ability to generate coherent language
Variation:
small search space increases probability of linguistic coherence
Evolution of UG
Model: Two individuals communicating with U
Iund U
JAdditional Components:
Equation:
in the limit of no mutation: only one UG
Relative abundance of individuals with UJ speaking Lj
xJj Probability of genetical mutation from UI to UJ WIJ