using iterated learning to reveal biases for well
play

Using iterated learning to reveal biases for well-structured - PowerPoint PPT Presentation

Using iterated learning to reveal biases for well-structured meanings in language Jon W. Carr Centre for Language Evolution University of Edinburgh Linguistics and English Language Postgraduate Conference 2016 What shapes language? Language


  1. Using iterated learning to reveal biases for well-structured meanings in language Jon W. Carr Centre for Language Evolution University of Edinburgh Linguistics and English Language Postgraduate Conference 2016

  2. What shapes language? Language

  3. What shapes language? Language Expressivity

  4. What shapes language? Language Expressivity Learnability Kirby, Tamariz, Cornish, & Smith, 2015, Cognition

  5. What shapes language? Language Informativeness Simplicity Kirby, Tamariz, Cornish, & Smith, 2015, Cognition Kemp & Regier, 2012, Science

  6. What shapes language? Language Deutlichkeitsstreben Bequemlichkeitsstreben Kirby, Tamariz, Cornish, & Smith, 2015, Cognition Kemp & Regier, 2012, Science Gabelentz, 1901

  7. Models of learning vs communication Transmission chain Dyadic interaction Transmission chain with dyadic interaction

  8. Learning vs communication Dyadic interaction Transmission chain with dyadic interaction pihino nemone piga kawake egewawu egewawa egewuwu ege kapa gakho wuwele nepi mega megawawa megawuwu wulagi newhomo kamone gaku hokako gamenewawu gamenewawa gamenewuwu gamene Kirby, Tamariz, Cornish, & Smith, 2015, Cognition

  9. How do learning and communication shape the structure of semantic categories?

  10. Previous work Carr, J. W., Smith, K., Cornish, H., & Kirby, S. (2016). The cultural evolution of structured languages in an open-ended, continuous world. Cognitive Science . doi:10.1111/cogs.12371

  11. Discrete meaning space egewawu egewawa egewuwu ege mega megawawa megawuwu wulagi gamenewawu gamenewawa gamenewuwu gamene

  12. Open-ended meaning space

  13. Experiment 1 Training Test Training Test Training Test output etc… input output input output input DYNAMIC SET 0 DYNAMIC SET 1 DYNAMIC SET 2 etc… STATIC SET STATIC SET Generation 1 Generation 2 Generation 3 Experiment 2 Training Communicative Training Communicative Training Communicative input output input output input output etc… DYNAMIC SET 0 DYNAMIC SET 1 DYNAMIC SET 2 etc… STATIC SET STATIC SET Generation 1 Generation 2 Generation 3

  14. fama

  15. a m a p fama

  16. a m a p fama fod

  17. a m a p muaki fama fod

  18. a m a p muaki kazizui fama fod kazizizui k a z i z i z u

  19. Generation Generation Generation Generation Generation Generation Generation Generation Generation Generation Generation 10 1 2 5 3 4 6 7 8 9 0

  20. Generation Generation Generation Generation Generation Generation Generation Generation Generation Generation 10 7 6 5 8 4 3 9 2 1

  21. Generation Generation Generation Generation Generation Generation Generation Generation Generation 10 3 4 6 5 7 8 9 2

  22. Generation Generation Generation Generation Generation Generation Generation Generation 10 9 8 7 6 5 4 3

  23. Generation Generation Generation Generation Generation Generation Generation 10 9 8 7 6 5 4

  24. Generation Generation Generation Generation Generation Generation 10 9 8 7 6 5

  25. Generation Generation Generation Generation Generation 10 9 8 7 6

  26. Generation Generation Generation Generation 10 9 8 7

  27. Generation Generation Generation 10 9 8

  28. Generation Generation 10 9

  29. Generation 10

  30. Generation 10 mamo pika

  31. Conclusions Experiment 1 showed that cultural evolution can deliver languages that categorize the meaning space under pressure from learnability. This happens by losing categories and structuring the space in such a way that is easy to learn. Experiment 2 combined a pressure for learnability and a pressure for expressivity derived from a genuine communicative task. This gave rise to languages that use both categorization and string-internal structure to be both learnable and expressive.

  32. Ongoing work

  33. Informativeness Natural category systems “provide maximum information with the least cognitive effort” (Rosch, 1999). Regier et al. formalize informativeness as “communicative cost”. The most informative category system is one that minimizes communicative cost.

  34. Informativeness Natural category systems “provide maximum information with the least cognitive effort” (Rosch, 1999). Regier et al. formalize informativeness as “communicative cost”. The most informative category system is one that minimizes communicative cost. Negative logarithm of average within- category similarity, summed for all possible targets. � � � − � · � � − ��� � ( � � , � ) | � � | � ∈ � � ∈ � � Convex Random

  35. Informativeness Convex Random Convex Random Convex Random

  36. Three predictions Maximize number of categories The use of a single category has the highest cost and is therefore the least informative system; placing every item in its own category reduces the cost to 0 (maximum informativeness). Maximize dimensionality Representing 64 items using three dimensions is more informative than representing 64 items using one dimension (for a given number of categories). Maximize convexity A convex category system is always better than or equal to a non-convex system in terms of minimizing communicative cost. Convex category structures are optimal.

  37. Except: The learnability tradeoff Maximize number of categories But: Learning an infinite number of categories is not possible given finite time and cognitive resources. Maximize dimensionality But: Representing categories using infinite feature dimensions would be impossible to process. However: Maximize convexity Convexity leads to systems that are both more informative and potentially easier to learn. Thus, the property of convexity seems to be particularly interesting (Gärdenfors, 2000, 2014).

  38. Stims: Shepard circles 147.0° 2.57 rad 172.71° 3.01 rad 198.43° 3.46 rad 224.14° 3.91 rad 249.86° 4.36 rad 275.57° 4.81 rad 301.28° 5.26 rad 327.0° 5.71 rad 25 px 50 px 75 px 100 px 125 px 150 px 175 px 200 px

  39. Stims: Shepard circles 147.0° 2.57 rad 172.71° 3.01 rad 198.43° 3.46 rad 224.14° 3.91 rad 249.86° 4.36 rad 275.57° 4.81 rad 301.28° 5.26 rad 327.0° 5.71 rad 25 px 50 px 75 px 100 px 125 px 150 px 175 px 200 px

  40. Stims: Shepard circles 147.0° 2.57 rad 172.71° 3.01 rad 198.43° 3.46 rad 224.14° 3.91 rad 249.86° 4.36 rad 275.57° 4.81 rad 301.28° 5.26 rad 327.0° 5.71 rad 25 px 50 px 75 px 100 px 125 px 150 px 175 px 200 px

  41. Stims: Shepard circles 147.0° 2.57 rad 172.71° 3.01 rad 198.43° 3.46 rad 224.14° 3.91 rad 249.86° 4.36 rad 275.57° 4.81 rad 301.28° 5.26 rad 327.0° 5.71 rad 25 px 50 px 75 px 100 px 125 px 150 px 175 px 200 px

  42. Squares and Stripes: Three category systems Angle-only Size-only Angle & Size Easy to learn but low informativeness Informative but hard to learn

  43. Results (so far…)

  44. Results: Training trajectory

  45. Results: Test performance Training material Participant’s test outcome

  46. Results: Angle only

  47. Results: Angle only

  48. Results: Size only

  49. Results: Size only

  50. Results: Angle & Size

  51. Results: Angle & Size

  52. Results: Dimension preference Angle-only system Size-only system Angle & Size system

  53. Results: Dimension preference Angle-only system Size-only system Angle & Size system

  54. Results: Dimension preference Angle-only system Size-only system Angle & Size system Angle and size equally important

  55. Results: Dimension preference Angle-only system Size-only system Angle & Size system Angle more important than size

  56. Results: Dimension preference Angle-only system Size-only system Angle & Size system Size more important than angle

  57. Next steps Why do people find the size-only condition so hard – is it just something weird with these stims? What happens when the task is iterated in a transmission chain? Prediction: Everyone shifts to the angle-only system because it’s easiest Prediction: lots of noise Prediction: loss of categories What happens when you introduce a communicative task? Prediction: Everyone shifts to the angle & size system because it’s the most informative.

  58. Thanks!

Recommend


More recommend