Modelling Language Evolution Lecture/practical: The Talking Heads

1
Modelling Language EvolutionLecture/practical
The Talking Heads

• Simon Kirby
• University of Edinburgh
• Language Evolution Computation Research Unit

2
Meanings in models.

• Most models either
• Ignore meanings entirely (e.g., Elman,
  Christiansen etc.)
• Treat meanings as pre-given (e.g., Kirby etc.)
• Why have meanings at all?
• Imagine an iterated learning model without them
• what would the best language be like?
• Probably say anything or say nothing
• Need something to drive agents to be expressive

3
Problems with meanings.

• Where do meanings come from?
• If they are predefined, then the model must
  assume they are innate
• Is this justified?
• How different are different languages semantics?
• How do we learn the meaning of words?
• In models, agents are simply given the
  meaning-form pairs on a plate
• What do meanings correspond to in the real world?
• In models, they are symbolic representations
  there is no real world.

4
Enter the robots

• Luc Steels and others
• Get round these problems by using robots

5
Communication grounded in sensory-motor behaviour

• Quinn (2001) Use evolving robots
• Khepera robot
• Two motors and several proximity sensors
• Controlled by neural network
• No learning
• Weights evolved by genetic algorithm

6
Cooperative evolutionary task

• Khepera inhabit environment in pairs
• Put into environment close together but in random
  orientation
• Task (fixed time limit)
• Move final average position as far away from
  initial position as possible
• How to maximise fitness?
• Need to coordinate to both go in the same
  direction
• How can the robots coordinate?
• One must lead, the other follow

7
The evolution of after you

• Evolutionary early solution
• Two genotypes leaders and followers
• Problem if two leaders or two followers get
  picked
• Communicative solution
• Agent rotates until it faces the other
• The first to face moves forward to close range
  (this will be the follower)
• Once in range, it oscillates back and forth
• This second agent starts reversing, and the other
  follows

8
Scales up to a number of robots
9
Grounding in learned system

• Assume there is a real world out there and agents
  want to discriminate objects in the world and
  name them
• Talking heads experiment (e.g., Paul Vogt)

10
Discrimination

• World is a collection of objects (shapes on
  whiteboard)
• Represented as features Red, Green, Blue, Area,
  X, Y
• Context a set of objects on white board
• Topic one particular object
• Robots want to build a set of meanings
• Meaning is a region represented by a prototype
• A particular colour, area and location
• The category of every object is the region
  represented by its nearest prototype
• An object is discriminated if its category is
  different from all the others in the context

11
Simplified example
CONTEXT A(0.1, 0.3) B(0.3, 0.3) C(0.25,
0.15)
A
B
b
a
C
ROBOTS PROTOTYPES a(0.15, 0.25) b(0.35,
0.3)
A is categorised as a B is categorised as b C is
categorised as b
A is discriminated B and C are not
12
After discrimination

• If discrimination of the topic fails
• Add new prototype that corresponds to exactly
  that topic
• If discrimination of the topic succeeds
• Shift prototype slightly so that it moves closer
  to the topic
• If discrimination succeeds, the distinctive
  category is used to play a language game
• In some sense, the distinctive category is the
  meaning that will be communicated

13
The language game

• The lexicon is a set of meaning-word pairs with
  an association score (between 0 and 1)
• Language game speakers use a word to refer to
  the topic (possibly by invention) and hearers try
  to interpret that word
• Different game types
• Guessing game
• Observational game

14
The guessing game step 1

• Hearers try and guess the topic
• Hearers look at all objects in context only
  consider discriminated objects as possible topics
• (BUT Hearers categorise (discriminate) each
  object in context anyway)
• Look for association between each possible topic
  and the speakers word
• Select one with highest score
• Speakers provide corrective feedback
• Yes, thats the topic
• No, thats not the topic and indicate the
  correct one

15
The guessing game step 2

• Both speaker and hearer adjust lexicon
• Speaker
• Success increase association, and laterally
  inhibit other associations
• Failure decrease association
• Hearer
• Success same as speaker
• Failure to guess correct topic decrease
  association, and increase association with
  correct topic
• Failure to understand word add word

16
What decides communicative success?

• ADJUST
• Number of games
• Just colours
• Noise
• Context size
• Number of agents

17
Look into their brains what determines lexicon
size?

• LOOK AT
• Number of symbols
• Number of meanings

18
The observational game

• The guessing game relies on corrective feedback
• Instead, lets assume that learners can figure
  out what the topic is (joint attention? ToM?)
• Hearers given the topic
• Hearers play discrimination game only on topic
• Looks for association between the topic and the
  speakers word
• Speaker knows if hearer has correct association
• Both speaker and hearer adjust lexicon just as
  before

19
What difference does this make?
Repeat selected experiments