The Transition to Language: Where learning, culture

1
The Transition to Language Where learning,
culture evolution meet

• Simon Kirby
• with Kenny Smith Henry Brighton
• Language Evolution and Computation Research
  UnitTheoretical and Applied Linguistics,
  Edinburgh

http//www.ling.ed.ac.uk/simon
2
Language evolution in context

• Eight key transition events in the history of
  life on Earth
• (Maynard Smith Szathmáry 1995)
• Replicating molecules ? Populations of molecules
• Independent replicators ? Chromosomes
• RNA ? DNA
• Prokaryotes ? Eukaryotes
• Asexual clones ? Sexual populations
• Protists ? Animals, plants and fungi
• Solitary individuals ? Colonies
• Primate societies ? Human societies (Language)

Each transition involves an increase in the
complexity of the mechanisms of information
transmission. Language is a new kind of
evolutionary system.
3
The ultimate why question for linguistics
E
L
L is the set of human languages (characterised by
Language Universals)
Why are human languages the way they are and not
some other way?
4
What is evolutionary linguistics?
The study of language as a uniquely complex
interaction of dynamical systems
Acquisition
Change
Ontogenetic
Cultural
Natural Selection
Biological
The key feature of evolutionary linguistics
Answers question, why are languages the way they
are? by asking, how did they come to be that way?
5
Compositionality, a test case for evolutionary
linguistics

• Compositionality
• The meaning of an utterance is some function of
  the meaning of parts of that utterance and the
  way they are put together.
• Compositionality a good target for explanation
• Compositionality probably unique to humans
• Fundamental underlying feature of any theory of
  morphosyntax

Batali (1998, in press), Kirby (2000, 2001, in
press), Wray (2000), Hurford (2000, 2002), Kirby
Hurford (2001), Zuidema (2001), Tonkes (2002),
Brighton Kirby (2001), Brighton (2002), Nowak,
Plotkin Jansen (2000)
6
The nativist approach to explanation
Language Acquisition Device
Linguistic Data E-language
Grammar I-language

• Any learning mechanism must have some sort of
  prior bias
• A prior bias will affect the learnability of
  particular grammars
• Where does the human prior bias come from?
• Our biological endowment.
• Nativism hypothesis what we are born with
  directly determines the structure of languages
  (i.e., Language Universals)
• Vanilla (i.e., Chomskyan) nativism has nothing
  more to say. A characterisation of UG is itself
  explanatory.

7
The evolutionary psychology approach(Pinker
Bloom 1990)

• A synthesis of nativism and functionalism
• Grammar is a complex mechanism tailored to the
  transmission of propositional structures through
  a serial interface.
• Evolutionary theory offers clear criteria for
  when a trait should be attributed to natural
  selection complex design for some function, and
  the absence of alternative processes capable of
  explaining such complexity. Human language meets
  these criteria.

8
Evolutionary psychology explanation for
compositionality

• We use a compositional syntax because we are born
  with an LAD that can only learn compositional
  systems.
• Our proto-human ancestors learning mechanisms
  had the opposite bias.
• Compositionally biased learners have a selection
  advantage under certain ecological conditions.
• Therefore, a compositional LAD evolved through
  natural selection

9
Potential problems
Grammar is a complex mechanism tailored to the
transmission of propositional structures through
a serial interface. Evolutionary theory offers
clear criteria for when a trait should be
attributed to natural selection complex design
for some function, and the absence of alternative
processes capable of explaining such complexity.
Human language meets these criteria.

• Functionalist arguments notoriously tricky in
  linguistics. (Lightfoot 1999)
• Hard for natural selection to tune
  communicatively optimal innate constraints
  without Baldwin Effect, but this fails due to
  decorrelation. (Kirby Hurford 1997)
• Perhaps there are alternative processes?

10
Repeated acquisition

• Output of learning is the input to learning
• No longer a simple relationship between the
  structure of the LAD (UG) and universal
  properties of languages
• Structure of language emerges from dynamicsof
  iterated learning

Language Acquisition Device
E-Language
I-Language
Arena of Use
11
The Iterated Learning Model(a framework for
computational simulation)
Meaning-signal Pairs in (utterances from parent)
Agent (simulated individual)
Learning Algorithm
Internal linguistic representation
Meanings (generated by environment)
Production Algorithm
Meaning-signal Pairs out (to next generation)
12
Components of an ILM1. The environment

• The environment provides meanings for speakers.

A meaning space
An environment

• The meaning space has
• n-features (here 3)
• m-values (here 5)

• The environment can be
• sparse or dense,
• structured or random

13
Components of an ILM2. Signals

• Signals are strings of characters. Can vary in
  length, and size of alphabet.
• Here, signals are up to 3 long, and alphabet has
  size 5.

(1 2 3)
(4 2 3)
Compositional
(1 2 3)
(1 2 3)
(4 2 3)
d d d
a b c
( 2 )
(1 )
( 3)
(4 2 3)
Holistic
a
b
c
( 2 )
(4 )
( 3)
a
d
c
14
Components of an ILM3. Population model

• How many speakers? How many learners? Do learners
  speak to each other? How does the population turn
  over?
• Simplistic model.
• one adult speaker
• one child learner
• speaker produces signals given random meanings
  from environment
• child learns to associate meanings and signals
• adult removed, child becomes adult, new child
  introduced
• Crucial parameter how many utterances?
• potential bottleneck on transmission of language
  (i.e. poverty of the stimulus)

15
Components of an ILM4. Learning model

• A variety of models of learning in the
  literature
• Recurrent Neural Networks (Batali 1998, Tonkes
  2002)
• Feedforward Networks (Kirby 2002, Kirby Hurford
  2002, Smith 2001)
• Trigger learners (Kirby Hurford 1997, Yamauchi
  2001, Briscoe 2000)
• Instance-based learning (Batali in press)
• Heuristic CFG induction (Kirby 2000, 2001, in
  press, Zuidema 2001)
• MDL automata induction (Teal 2000, Brighton
  Kirby 2001, Brighton 2002)

16
Associative memory
17
Storage
18
Production
19
Simulation runs

• Four environments
• Dense, unstructured
• Dense, structured
• Sparse, unstructured
• Sparse, structured
• Run 1000 iterated learning simulations to
  stability (usually within 50 generations)
• Assess final language for compositionality using
  all-pairs distance correlation, c
• c is 1 when neighbourhood in meaning space is
  preserved exactly in signal space.
• c is 0 when the mapping is random.

20
Results with no poverty of the stimulus
21
Results with poverty of stimulus (40 coverage)
22
Summary

• Qualitative language evolution without any
  biological evolution or functional pressures (or
  communication!).
• Compositionality emerges without hard
  constraints.
• Sensitive to
• structure in perception of environment and
• limited training data (i.e., poverty of the
  stimulus)
• Why?
• Language itself is an evolving system.
• Languages that are generalisable from limited
  data are more stable in an iterated learning
  scenario if there is a bottleneck.
• Language adapts to aid its own survival.

23
Implications

• What does the ILM tell us about linguistic
  theory?
• Links the evolution of basic linguistic structure
  to the mechanisms that underlie language change
  creolisation
• The ILM shows that evolutionary approaches need
  not be functionalist.
• It changes the relationship between universals
  and Universal Grammar

24
UG in evolutionary psychology
E
pUG
L
UG
F
L is the set of human languages
UG is the set of learnable languages
F is the set of functional languages
pUG is the set of languages learnable by some
proto LAD
pUG evolves under pressure from natural selection
to fit F
25
UG as environment of adaptation for evolving
languages
E
L
UG
L is the set of human languages
UG is the set of learnable languages
Languages are attracted into the set L
26
Conclusions

• The transition to language is a transition to a
  new kind of evolutionary system.
• The ILM is a tool for understanding this system
• Our results so far suggest four things
• Linguistic structure may not be coded directly in
  UG
• Language evolution need not appeal to
  communicative function
• Platos Problem is not so much a problem but more
  of a requirement for the emergence of linguistic
  structure.
• We need to find out how learning, culture, and
  evolution interact, in order to relate UG and
  language universals.
• We have evolved to learn languages. Languages
  evolve to be learned.