Co-evolution of language and mindreading: a computational exploration
View/ Open
Date
02/07/2019Author
Woensdregt, Marieke Suzanne
Metadata
Abstract
Language relies on mindreading: in order to use it successfully we need to be able
to entertain and recognise communicative intentions. Mindreading abilities in turn
profit from language, as language provides a means for expressing mental states
explicitly, and for transmitting our knowledge of mental states to others. Given this
interdependence, it has been hypothesised that language and mindreading have
co-evolved. In this thesis I formalise the relationship between language and mindreading
in a computational model, in order to explore under what circumstances a
co-evolutionary dynamic between the two skills could have gotten off the ground.
In Chapter 3 I present an agent-based model which combines referential signalling
with perspective-taking, where perspective-taking instantiates a very simple
form of mindreading. In this model, agents’ communicative behaviour is probabilistically
determined by an interplay between their language and their perspective
on the world. The literal variant of these agents (explored in Chapters 3 and 4)
consists of speakers who produce utterances purely based on their own language
and perspective, and listeners who interpret these utterances using what they’ve
learned about the speaker’s perspective through interaction. The pragmatic variant
of these agents in contrast (explored in Chapters 5 and 6) consists of speakers who
optimise their utterances by maximising the probability that the listener will interpret
them correctly (assuming the listener shares their perspective), and listeners
who interpret these utterances by reasoning about such a speaker, again using what
they’ve learned about the speaker’s perspective through interaction. Learning is not
straightforward however, because agents’ languages and perspectives are private
(i.e. not directly observable to other agents). Instead, the Bayesian learners in this
model only get to observe a speaker’s utterances in context, from which they have
to simultaneously infer the speaker’s language and perspective. Simulation results
show that learners can overcome this joint inference problem by bootstrapping one
from the other, but that the success of this process depends on how informative the
speaker’s language is.
This leads to an evolutionary question: If the co-development between language-learning
and perspective-learning relies on agents being exposed to an informative
language, how could a population of such agents evolve an informative language
from scratch? I address this question with an iterated learning version of the
model described above, combined with different selection pressures. Simulation
results with literal agents (presented in Chapter 4) show that an informative language
emerges not just if the population is subjected to a selection pressure for
communication, but also under selection for accurate perspective-inference. Under
both pressures, the emergence of an informative language leads not just to more
successful communication, but also to more successful perspective-inference. This
is because sharing an informative language with others provides agents with information
about those others’ perspectives (note that agents’ innate ability to learn
about others’ perspectives does not change over generations). Simulation results
with pragmatic agents (presented in Chapter 5) show the same co-evolutionary dynamics
as literal agents, with the difference that they can achieve equally high levels
of success at communicating and inferring perspectives with much more ambiguous
languages, because they can compensate for suboptimal languages using their pragmatic
ability. Finally, in Chapter 6 I explore under what circumstances such pragmatic
agents could have evolved; that is, under what circumstances being a pragmatic communicator
provides an evolutionary advantage over being a literal communicator.
Taken together, the model results presented in this thesis suggest firstly that
co-evolution between language and mindreading could have gotten off the ground
under any circumstances which created a need for either improved communication
or improved insight into others’ minds. Secondly, the results suggest that such a co-evolutionary
dynamic could have been driven largely by cultural evolution; where
mindreading improves by virtue of evolving a language.