The Co-evolution of Speech and the Lexicon: The Interaction of - - PowerPoint PPT Presentation

The Co-evolution of Speech and the Lexicon: The Interaction of Functional Pressures, Redundancy, and Category Variation

Winter & Wedel (2016)

Presented by Miriam Schulz Seminar: Exemplar Theory Lecturer: Prof. Bernd Möbius 3 June 2020

Outline

1. Introduction 2. The computational model 3. Simulations 4. Conclusion

Outline

1. Introduction 2. The computational model 3. Simulations 4. Conclusion

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

The evolution of spoken language

4

Change in form Stability in function

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

The evolution of spoken language

5

Change in form Stability in function

cot caught [k၁t]

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

The evolution of spoken language

6

Change in form Stability in function If language changes constantly, how can we maintain meaning?

cot caught [k၁t]

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

A multi-level exemplar framework

7

• Speech as a repeated cycle of production and perception

(Pierrehumbert 2001)

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

A multi-level exemplar framework

8

• Speech as a repeated cycle of production and perception

(Pierrehumbert 2001)

• Word pronunciation is influenced by stored exemplars of

the word, as well as of its individual sounds

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

A multi-level exemplar framework

9

• Speech as a repeated cycle of production and perception

(Pierrehumbert 2001)

• Word pronunciation is influenced by stored exemplars of

the word, as well as of its individual sounds Pronunciation of ‘cot’

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

A multi-level exemplar framework

10

• Speech as a repeated cycle of production and perception

(Pierrehumbert 2001)

• Word pronunciation is influenced by stored exemplars of

the word, as well as of its individual sounds Pronunciation of ‘cot’ Stored exemplars of the word ‘cot’

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

A multi-level exemplar framework

11

• Speech as a repeated cycle of production and perception

(Pierrehumbert 2001)

• Word pronunciation is influenced by stored exemplars of

the word, as well as of its individual sounds Pronunciation of ‘cot’ Stored exemplars of the word ‘cot’ Stored exemplars of the sounds of ‘cot’: /k/, /၁/, /t/

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

A multi-level exemplar framework

12

• Speech as a repeated cycle of production and perception

(Pierrehumbert 2001)

• Word pronunciation is influenced by stored exemplars of

the word, as well as of its individual sounds Perception of ‘cot’

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

A multi-level exemplar framework

13

• Speech as a repeated cycle of production and perception

(Pierrehumbert 2001)

• Word pronunciation is influenced by stored exemplars of

the word, as well as of its individual sounds Perception of ‘cot’ Word-level update

• f ‘cot’

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

A multi-level exemplar framework

14

• Speech as a repeated cycle of production and perception

(Pierrehumbert 2001)

• Word pronunciation is influenced by stored exemplars of

the word, as well as of its individual sounds Perception of ‘cot’ Word-level update

• f ‘cot’

Sound-level update of /k/, /၁/, /t/

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Hidden variation

15

• Variation in pronunciation does not always impact

categorization success

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Hidden variation

16

• Variation in pronunciation does not always impact

categorization success

It doesn’t matter if I pronounce /k/

• r /x/ !

Are you saying “roca” and “roja” sound the same to you?!

English native Spanish native

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Hidden variation

17

• Variation in pronunciation does not always impact

categorization success Schematized variation of /k/ in English Schematized variation of /k/ in Spanish

/k/ /x/ /k/ /x/

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Hidden variation

18

• Variation in pronunciation does not always impact

categorization success

• Analogy of neutral or cryptic variation in biology

(Wagner 2005)

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Hidden variation

19

Cryptic variation in biology: “variation that is not visible to evolution”

• Variation in pronunciation does not always impact

categorization success

• Analogy of neutral or cryptic variation in biology

(Wagner 2005)

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Hidden variation

20

Source: Félix & Wagner (2008)

Cryptic variation in biology: “variation that is not visible to evolution”

• Variation in pronunciation does not always impact

categorization success

• Analogy of neutral or cryptic variation in biology

(Wagner 2005)

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Hidden variation

21

Source: Félix & Wagner (2008) ≈ Variation in stored exemplars

Cryptic variation in biology: “variation that is not visible to evolution”

• Variation in pronunciation does not always impact

categorization success

• Analogy of neutral or cryptic variation in biology

(Wagner 2005)

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Hidden variation

22

Source: Félix & Wagner (2008) ≈ Variation in stored exemplars

Cryptic variation in biology: “variation that is not visible to evolution”

≈ Production noise

• Variation in pronunciation does not always impact

categorization success

• Analogy of neutral or cryptic variation in biology

(Wagner 2005)

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Hidden variation

23

Source: Félix & Wagner (2008) ≈ Word recognition ≈ Variation in stored exemplars

Cryptic variation in biology: “variation that is not visible to evolution”

≈ Production noise

• Variation in pronunciation does not always impact

categorization success

• Analogy of neutral or cryptic variation in biology

(Wagner 2005)

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

This paper

Research questions:

➔ How do the distribution of word categories and the distribution of sound categories interact?

24

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

This paper

Research questions:

➔ How do the distribution of word categories and the distribution of sound categories interact? ➔ How can the system of sound categories evolve?

25

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

This paper

Research questions:

➔ How do the distribution of word categories and the distribution of sound categories interact? ➔ How can the system of sound categories evolve?

Hypothesis:

The more words a specific sound contrast distinguishes, the less likely that contrast is to be lost.

26

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

This paper

Research questions:

➔ How do the distribution of word categories and the distribution of sound categories interact? ➔ How can the system of sound categories evolve?

Hypothesis:

The more words a specific sound contrast distinguishes, the less likely that contrast is to be lost.

Method:

Simulation; use computational model as conceptual tool

27

Outline

1. Introduction 2. The computational model 3. Simulations 4. Conclusion

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

The computational model: setup

• Two agents

29

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

The computational model: setup

30

• Two agents
• Each with a mental lexicon

○ e.g. a 4-word-lexicon: {ba, pa, bi, pi}

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

The computational model: setup

• Two agents
• Each with a mental lexicon

○ e.g. a 4-word-lexicon: {ba, pa, bi, pi}

• Every word is seeded with some exemplars

○ e.g. ba: [ba1, ba2, …, ban]

31

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

The computational model: setup

• Two agents
• Each with a mental lexicon

○ e.g. a 4-word-lexicon: {ba, pa, bi, pi}

• Every word is seeded with some exemplars

○ e.g. ba: [ba1, ba2, …, ban]

• Each exemplar varies along two continuous

phonetic dimensions

○ Let’s assume that dimension 1 = voicing (/b/ vs. /p/); dimension 2 = vowel height (/a/ vs. /i/)

32

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

The computational model: setup

• Two agents
• Each with a mental lexicon

○ e.g. a 4-word-lexicon: {ba, pa, bi, pi}

• Every word is seeded with some exemplars

○ e.g. ba: [ba1, ba2, …, ban]

• Each exemplar varies along two continuous

phonetic dimensions

○ Let’s assume that dimension 1 = voicing (/b/ vs. /p/); dimension 2 = vowel height (/a/ vs. /i/) ○ /ba/ ≈ (30,30) vs. /pi/ ≈ (70,70)

33

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

The computational model: setup

• Two agents
• Each with a mental lexicon

○ e.g. a 4-word-lexicon: {ba, pa, bi, pi}

• Every word is seeded with some exemplars

○ e.g. ba: [ba1, ba2, …, ban]

• Each exemplar varies along two continuous

phonetic dimensions

○ Let’s assume that dimension 1 = voicing (/b/ vs. /p/); dimension 2 = vowel height (/a/ vs. /i/) ○ /ba/ ≈ (30,30) vs. /pi/ ≈ (70,70) VOT

34

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

The computational model: setup

• Two agents
• Each with a mental lexicon

○ e.g. a 4-word-lexicon: {ba, pa, bi, pi}

• Every word is seeded with some exemplars

○ e.g. ba: [ba1, ba2, …, ban]

• Each exemplar varies along two continuous

phonetic dimensions

○ Let’s assume that dimension 1 = voicing (/b/ vs. /p/); dimension 2 = vowel height (/a/ vs. /i/) ○ /ba/ ≈ (30,30) vs. /pi/ ≈ (70,70) VOT tongue height

35

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

The computational model: setup

• Two agents
• Each with a mental lexicon

○ e.g. a 4-word-lexicon: {ba, pa, bi, pi}

• Every word is seeded with some exemplars

○ e.g. ba: [ba1, ba2, …, ban]

• Each exemplar varies along two continuous

phonetic dimensions

○ Let’s assume that dimension 1 = voicing (/b/ vs. /p/); dimension 2 = vowel height (/a/ vs. /i/) ○ /ba/ ≈ (30,30) vs. /pi/ ≈ (70,70) VOT tongue height

36

e.g. VOT e.g. tongue height

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

The computational model: dialogue

37

Speaker Listener

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

The computational model: dialogue

38

Speaker

• Choose word from vocabulary

/ba/

Listener

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

The computational model: dialogue

39

Speaker

• Choose word from vocabulary
• Select exemplar for production: more

recent exemplars have higher activation value (memory decay)

/ba/: (23, 29) /ba/ /ba/: [(23,29), (25,31), (30,30), ...]

Listener

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

The computational model: dialogue

40

Speaker

• Choose word from vocabulary
• Select exemplar for production: more

recent exemplars have higher activation value (memory decay)

• Apply two biases:

a. Random production noise

/ba/: (23, 29) /ba/ /ba/: [(23,29), (25,31), (30,30), ...] (23,29) → (21,33)

Listener

noise: (–2,+4)

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

The computational model: dialogue

41

Speaker

• Choose word from vocabulary
• Select exemplar for production: more

recent exemplars have higher activation value (memory decay)

• Apply two biases:

a. Random production noise b. Similarity bias

/ba/: (23, 29) /ba/ /ba/: [(23,29), (25,31), (30,30), ...] /ba/: [(23,29), (25,31), ...] /bi/: [(31,79), (28,75), ….] (23,29) → (21,33)

Listener

noise: (–2,+4)

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

The computational model: dialogue

42

Speaker

• Choose word from vocabulary
• Select exemplar for production: more

recent exemplars have higher activation value (memory decay)

• Apply two biases:

a. Random production noise b. Similarity bias

/ba/: (23, 29) /ba/ /ba/: [(23,29), (25,31), (30,30), ...] /ba/: [(23,29), (25,31), ...] /bi/: [(31,79), (28,75), ….] (23,29) → (21,33)

Listener

noise: (–2,+4)

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

The computational model: dialogue

43

Speaker

• Choose word from vocabulary
• Select exemplar for production: more

recent exemplars have higher activation value (memory decay)

• Apply two biases:

a. Random production noise b. Similarity bias

/ba/: (23, 29) /ba/ /ba/: [(23,29), (25,31), (30,30), ...] /ba/: [(23,29), (25,31), ...] /bi/: [(31,79), (28,75), ….] (23,29) → (21,33)

Listener

noise: (–2,+4)

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

The computational model: dialogue

44

Speaker

• Choose word from vocabulary
• Select exemplar for production: more

recent exemplars have higher activation value (memory decay)

• Apply two biases:

a. Random production noise b. Similarity bias

/ba/: (23, 29) (21, 33) → /ba/ !

Listener

• Categorize word based on incoming

sound noise: (–2,+4)

(21, 33) → /ba/ ! /ba/ /ba/: [(23,29), (25,31), (30,30), ...] /ba/: [(23,29), (25,31), ...] /bi/: [(31,79), (28,75), ….] (23,29) → (21,33)

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

The computational model: dialogue

45

Speaker

• Choose word from vocabulary
• Select exemplar for production: more

recent exemplars have higher activation value (memory decay)

• Apply two biases:

a. Random production noise b. Similarity bias

/ba/: (23, 29) (21, 33) → /ba/ !

Listener

• Categorize word based on incoming

sound

• Store as new exemplar of the

identified category noise: (–2,+4)

(21, 33) → /ba/ ! /ba/: [(21,33), (19,29,) ...] /ba/ /ba/: [(23,29), (25,31), (30,30), ...] /ba/: [(23,29), (25,31), ...] /bi/: [(31,79), (28,75), ….] (23,29) → (21,33)

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

The computational model: dialogue

46

Speaker

• Choose word from vocabulary
• Select exemplar for production: more

recent exemplars have higher activation value (memory decay)

• Apply two biases:

a. Random production noise b. Similarity bias

/ba/: (23, 29) (21, 33) → /ba/ !

Listener

• Categorize word based on incoming

sound

• Store as new exemplar of the

identified category → Anti-ambiguity bias: prioritize distinctive outputs noise: (–2,+4)

(21, 33) → /ba/ ! /ba/: [(21,33), (19,29,) ...] (21,33) more likely than (45,56) to be stored under /ba/ /ba/ /ba/: [(23,29), (25,31), (30,30), ...] /ba/: [(23,29), (25,31), ...] /bi/: [(31,79), (28,75), ….] (23,29) → (21,33)

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

The computational model: dialogue

47

Speaker

• Choose word from vocabulary
• Select exemplar for production: more

recent exemplars have higher activation value (memory decay)

• Apply two biases:

a. Random production noise b. Similarity bias

/ba/: (23, 29) (21, 33) → /ba/ !

Listener

• Categorize word based on incoming

sound

• Store as new exemplar of the

identified category → Anti-ambiguity bias: prioritize distinctive outputs noise: (–2,+4)

(21, 33) → /ba/ ! /ba/: [(21,33), (19,29,) ...] (21,33) more likely than (45,56) to be stored under /ba/ /ba/ /ba/: [(23,29), (25,31), (30,30), ...] /ba/: [(23,29), (25,31), ...] /bi/: [(31,79), (28,75), ….] (23,29) → (21,33)

Switch roles & repeat...

Outline

1. Introduction 2. The computational model 3. Simulations 4. Conclusion

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Simulation 1: the impact of vocabulary size

49

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Simulation 1: the impact of vocabulary size

50

Simulation results after 500 time steps

e.g. VOT e.g. tongue height

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Simulation 1: the impact of vocabulary size

51

Less variation with larger vocabulary

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Simulation 1: the impact of vocabulary size

52

Less variation with larger vocabulary Higher error rate with larger vocabulary

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Simulation 2: the impact of redundancy

53

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Hypothesis: a redundant system will be less constrained by functional load, so more redundancy → more variation

Simulation 2: the impact of redundancy

54

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Hypothesis: a redundant system will be less constrained by functional load, so more redundancy → more variation

Simulation 2: the impact of redundancy

55

(A) add a third, independent dimension, e.g. another vowel contrast, such as front-back → new vowel space: /i~ш~æ~α/

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Hypothesis: a redundant system will be less constrained by functional load, so more redundancy → more variation

Simulation 2: the impact of redundancy

56

(A) add a third, independent dimension, e.g. another vowel contrast, such as front-back → new vowel space: /i~ш~æ~α/

three dimensions two dimensions Legend:

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Hypothesis: a redundant system will be less constrained by functional load, so more redundancy → more variation

Simulation 2: the impact of redundancy

57 three dimensions two dimensions

(A) add a third, independent dimension, e.g. another vowel contrast, such as front-back → new vowel space: /i~ш~æ~α/ (B) add a second “syllable”, e.g. /bapi/

Legend:

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Legend:

Hypothesis: a redundant system will be less constrained by functional load, so more redundancy → more variation

Simulation 2: the impact of redundancy

58 three dimensions two dimensions two syllables

• ne syllable

(A) add a third, independent dimension, e.g. another vowel contrast, such as front-back → new vowel space: /i~ш~æ~α/ (B) add a second “syllable”, e.g. /bapi/

Legend:

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Simulation 3: hidden variation

59

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Simulation 3: hidden variation

60

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Simulation 3: hidden variation

61

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Simulation 3: hidden variation

62

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Simulation 3: hidden variation

63

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Simulation 3: hidden variation

64

Hidden variation creates pathways for future change

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Simulation 4: the anti-ambiguity bias

65

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Simulation 4: the anti-ambiguity bias

66

Wedel (2012)

Outline

1. Introduction 2. The computational model 3. Simulations 4. Conclusion

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Simulating the evolution of spoken language

68

Change in form Stability in function

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

69

Change in form Stability in function

Similarity biases Anti-ambiguity bias

Simulating the evolution of spoken language

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

70

Change in form Stability in function

Similarity biases Anti-ambiguity bias Relative

• ptimum of

variation

Simulating the evolution of spoken language

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Conclusion

• Goal: simulate variation & evolution of the sound system

71

cot caught [k၁t]

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Conclusion

• Goal: simulate variation & evolution of the sound system
• Framework: Exemplar theory as a model of evolutionary change through

production-perception loop

72

cot caught [k၁t]

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Conclusion

• Goal: simulate variation & evolution of the sound system
• Framework: Exemplar theory as a model of evolutionary change through

production-perception loop

• Method: simulation using an exemplar-based architecture

73

cot caught [k၁t]

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Conclusion

• Goal: simulate variation & evolution of the sound system
• Framework: Exemplar theory as a model of evolutionary change through

production-perception loop

• Method: simulation using an exemplar-based architecture
• Key findings:

○ More words → less variation, due to anti-ambiguity bias

74

cot caught [k၁t]

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Conclusion

• Goal: simulate variation & evolution of the sound system
• Framework: Exemplar theory as a model of evolutionary change through

production-perception loop

• Method: simulation using an exemplar-based architecture
• Key findings:

○ More words → less variation, due to anti-ambiguity bias ○ Extra syllables or phonetic dimensions → more variation, due to increased redundancy

75

cot caught [k၁t]

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Conclusion

• Goal: simulate variation & evolution of the sound system
• Framework: Exemplar theory as a model of evolutionary change through

production-perception loop

• Method: simulation using an exemplar-based architecture
• Key findings:

○ More words → less variation, due to anti-ambiguity bias ○ Extra syllables or phonetic dimensions → more variation, due to increased redundancy ○ Hidden variation as a pathway to exploit new dimensions for more efficiency

76

cot caught [k၁t]

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Conclusion

• Goal: simulate variation & evolution of the sound system
• Framework: Exemplar theory as a model of evolutionary change through

production-perception loop

• Method: simulation using an exemplar-based architecture
• Key findings:

○ More words → less variation, due to anti-ambiguity bias ○ Extra syllables or phonetic dimensions → more variation, due to increased redundancy ○ Hidden variation as a pathway to exploit new dimensions for more efficiency

• Implications:

○ Selection at word level impacts selection at sound level!

77

cot caught [k၁t]

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Conclusion

• Goal: simulate variation & evolution of the sound system
• Framework: Exemplar theory as a model of evolutionary change through

production-perception loop

• Method: simulation using an exemplar-based architecture
• Key findings:

○ More words → less variation, due to anti-ambiguity bias ○ Extra syllables or phonetic dimensions → more variation, due to increased redundancy ○ Hidden variation as a pathway to exploit new dimensions for more efficiency

• Implications:

○ Selection at word level impacts selection at sound level! ○ Structure of human languages shaped by cultural evolution

78

cot caught [k၁t]

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

References

Félix, M. A., & Wagner, A. (2008). Robustness and evolution: concepts, insights and challenges from a developmental model system. Heredity, 100(2), 132-140. Wedel, A. (2012). Lexical contrast maintenance and the organization of sublexical contrast systems. Language and Cognition, 4(4), 319-355. Winter, B., & Wedel, A. (2016). The Co-evolution of Speech and the Lexicon: The Interaction of Functional Pressures, Redundancy, and Category Variation. Topics in cognitive science, 8(2), 503–513. https://doi.org/10.1111/tops.12202 Winter & Wedel (2016) 3 June 2020 Miriam Schulz

79

Questions & Discussion

Winter & Wedel (2016) 3 June 2020 Miriam Schulz

Sample discussion topics

➔ Computational modeling as a tool for linguistic investigation ➔ A single-agent production-perception feedback loop? (see note 2) ➔ The biological metaphor of cryptic/neutral variation

81