Language as a culturally evolving system: from computer simulation - PowerPoint PPT Presentation

Language as a culturally evolving system: from computer simulation to the experiment lab V I N E Simon Kirby U R S E I H T Y T Language Evolution and Computation Research Unit O H School of Philosophy, Psychology & Language Sciences F G R E U University of Edinburgh D B I N

Why don’t we have the answers yet? • Language is our species defining characteristic • Surely we should know where it came from by now?

Why don’t we have the answers yet? • Language is our species defining characteristic • Surely we should know where it came from by now? • Language evolution is: “the hardest problem in science” (Christiansen & Kirby 2003) • Why?

There’s something special about language • It arises from 3 interacting dynamical systems • Biological evolution of a capacity for INDIVIDUAL BIOLOGICAL learning languages that are transmitted LEARNING EVOLUTION culturally over space and time in a speech community CULTURAL TRANSMISSION • Arguably, no other single behaviour in nature has such complex a source LANGUAGE • So, how do we go about tackling this?

Idealisations • A foundational idealisation: the idealised speaker/hearer in a homogenous speech community Universal properties of individual cognition • Abstract away from complex social/cultural/ population aspects • The nativist approach: • focus on individual’s faculty of language acquisition Universal properties of language

Idealisations • A foundational idealisation: the idealised speaker/hearer in a homogenous Genes “for” language speech community • Abstract away from complex social/cultural/ population aspects • The evolutionary/nativist approach: Universal properties of individual cognition • focus on individual’s faculty of language acquisition ?! • treat as a biological adaptation for communication Universal properties of language • universals of language arise from universal biological adaptation

The missing process • The individual-centric idealisation leaves out cultural transmission, precisely because we don’t understand it well • Leading questions: • Are there other idealisations we could make that would allow us to look at the full complexity of the adaptive systems at play? • Can we be sure that the basic assumptions we’ve outlined are correct? Do properties of the individual get straightforwardly “written out” in the structure of language? Are there alternative mechanisms explaining adaptive structure of language?

The iterated learning model • Cultural evolution of language is of a particular type: LEARNING INTERNAL ITERATED LEARNING: REPRESENTATION Learning by observation of behaviour in PRODUCTION another that was itself learned in the same way. OBSERVABLE BEHAVIOUR LEARNING • Simulate this process in the computer INTERNAL • Embed simple models of learners in an REPRESENTATION “environment” that they want to PRODUCTION communicate about OBSERVABLE • Agents learn from others’ behaviour (cf. BEHAVIOUR LEARNING the game “telephone”)

Quick summary of the results from simulations • Key insight: transmission bottlenecks • If a learner is given imperfect information about the language, e.g. noise, processing constraints, or simply not hearing all the data (cf. stimulus poverty) • ... cultural transmission becomes an adaptive system. • Language will adapt so that it appears to be designed to “fit” the bottleneck • Has been used to explain some of the design features of language, particularly compositionality

Design without a designer • Languages (arguably) have the appearance of design • Two obvious mechanisms to explain this: • Biological evolution (leading to the evolutionary/nativist view) • Intentional design by individuals • Computational models suggest an alternative • Cultural evolution • Consistent with the idea of the “invisible hand” (Keller 1990) • But can we demonstrate this with real human agents?

Cumulative cultural evolution in the laboratory: An experimental approach to the origins of structure in human language Simon Kirby* † , Hannah Cornish*, and Kenny Smith ‡ *School of Philosophy, Psychology, and Language Sciences, University of Edinburgh, Edinburgh EH8 9LL, United Kingdom; and ‡ Division of Psychology, Northumbria University, Newcastle-upon-Tyne NE1 8ST, United Kingdom Edited by Dale Purves, Duke University Medical Center, Durham, NC, and approved June 6, 2008 (received for review August 20, 2007)

An experimental approach • Inspired by work of Galantucci, Garrod, Fay and others, we tried to replicate simulations in the experiment lab (Kirby, Cornish & Smith, 2008) • Cultural transmission of an “alien” language 1. Start with a random artificial language 2. Ask experimental participant to learn language and test them 3. Use their output at test to teach the next participant in the experiment (and repeat)

Hypothesis • There will be cumulative cultural adaptation of the language without intentional design by the participants • Two ways of verifying this: • The language should become easier to learn • The language should become structured

The language • Simple syllable sequences for coloured moving shapes. • For example: kapihu kalu moki luki nelu kilamo • 27 “meanings” in total (3 shapes, 3 colours, 3 motions)

Procedure • Language divided randomly into two sets: • SEEN set: 14 meaning-signal pairs • UNSEEN set: remaining 13 meaning-signal pairs • Subjects trained only on SEEN set • Tested on complete set • Output on test randomly redivided into new SEEN and UNSEEN sets for next generation

Results 1.0 a) b) 0.8 10 ● 0.6 Structure Error ● ● ● 5 0.4 ● ● ● ● ● ● ● ● 0.2 ● ● ● ● 0 ● 0.0 ● ● 2 4 6 8 10 0 2 4 6 8 10 Generations Generations Language becomes easier to learn, and structured

After generation 1 hipe miniku kimei himini miwn miheniw wige nige hepinimi pemini tupim pobo kuwpi kuhepi tuge kupini weg mip mie pon mhip poi poh mpo

After generation 6 miniku poi tupin minku tupim tuge

After generation 7 miniku poi tupin tupim tuge

After generation 8 miniku poi tupin tupim tuge

After generation 9 miniku poi tupin tupim tuge

After generation 10 miniku poi tupin tupim tuge

Language adapts • Language adapts to be structured and learnable • Polysemy emerges that underspecifies the meaning-space in clever ways • Subjects are unaware of this happening

Language adapts • Language adapts to be structured and learnable • Polysemy emerges that underspecifies the meaning-space in clever ways • Subjects are unaware of this happening • But it’s not a particularly exciting form of structure! • Need a way of forcing the language to be expressive • Second experiment: simply discard words at random from SEEN set before training if they do not discriminate meanings

Results a) b) 1.0 ● ● 0.8 ● 10 ● ● ● ● ● 0.6 ● Structure Error ● 5 0.4 ● ● ● ● ● ● 0.2 ● ● ● ● 0 ● 0.0 2 4 6 8 10 0 2 4 6 8 10 Generations Generations Language becomes structured and easier to learn

Generation 1 lumonamo kinahune lahupine nelu kanehu namopihu kapihu humo lahupiki moki luneki lanepi kalu mola pihukimo nane kalakihu mokihuna kilamo kahuki neluka pilu neki pinemohu luki namola lumoka

Generation 10 n-ere-ki l-ere-ki renana n-ehe-ki l-aho-ki r-ene-ki n-eke-ki l-ake-ki r-ahe-ki n-ere-plo l-ane-plo r-e-plo n-eho-plo l-aho-plo r-eho-plo n-eki-plo l-aki-plo r-aho-plo n-e-pilu l-ane-pilu r-e-pilu n-eho-pilu l-aho-pilu r-eho-pilu n-eki-pilu l-aki-pilu r-aho-pilu Compositional language emerges. Adaptive structure that is both learnable and expressive.

Why is there adaptation? • Language is adapting to maximise its own transmissibility INTERNAL REPRESENTATION • Evolution from initial unlearnable random language to learnable structured one BOTTLENECK OBSERVED • Note: this is a blind process. Not the result of UTTERANCES intentional design by individuals. Design without designer (or biological INTERNAL evolution) REPRESENTATION • Driven by bottleneck on transmission: learners only see a subset of utterances

The poverty of the stimulus • In real world, language is unboundedly productive . This leads to an inevitable poverty of the stimulus (i.e., transmission bottleneck) • Chomsky sees this as a learning problem, suggesting the child must be born with specific knowledge of language in order to solve it • But now we can look at it another way: it is a challenge a culturally evolving language must meet. Language structure is an adaptive response to this challenge, allowing language to be reliably transmitted despite the poverty of the stimulus • The design-like features of language are not a result of us adapting to language, but rather language adapting to us.