Theories and Models of Language Change A Study of Self-Organization - - PowerPoint PPT Presentation

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

Theories and Models of Language Change

Session 6: Models II - Emergence of Universals Roland Mühlenbernd June 9, 2015

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

Review: Universal Darwinism

Mechanisms of universal evolution:

• 1. variation: continuing abundance of different elements
• 2. selection : number/probability of copies of elements -

depending on interaction between element features and environmental features

• 3. replication: reproduction/copying of elements

What is the role of linguistic universals in an evolutionary model of language change?

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

Language Change - Broad and Narrow Sense

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

Linguistic Universals

◮ are patterns that occur systematically across natural

languages

◮ can be distinguished between

◮ absolute (e.g. all languages have nouns and verbs) ◮ implicational (e.g. if a language is spoken, it has vowels

and consonants)

◮ are given on all linguistic levels:

◮ phonology (e.g. symmetry of sound inventories) ◮ syntax (Greenberg universals, e.g. SOV → postpositional) ◮ semantics (Swadesh list, natural semantic metalanguage

and its semantic primitives, basic color terms)

◮ pragmatics (e.g. generalized implicatures, speech acts)

◮ are given as innate cognitive structures or realized for

functional reasons under communicative aspects?

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

Universals of Sound Systems (Exercise 1)

The UCLA Phonological Segment Inventory Database (UPSID) contains 921 different speech sounds with the following values (over all 451 languages of the database):

◮ average number of phonemes: 20 to 37 ◮ minimal number of phonemes: 11 (Rotokas, Pirahã) ◮ maximal number of phonemes: 141 (!X˜

u) Example: Hawaiian phonemes: a, e, i, o, u, p, k, m, n, w, l, h, P

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

Universals of Sound Systems

Phonetic inventories of the world’s languages exhibit

◮ frequent and rare sounds: f([m]) = 94%, f([ö]) = 1% ◮ a tendency to symmetric inventories:

◮ f([ď] | [O]) = 83%; f([ď] | ¬[O]) = 18% ◮ f([t]) = 40%; f([t] | [d]) = 83%

Sound sequences of the world’s languages exhibit

◮ frequent and rare syllable structures: universal V, CV ◮ follow a sonority hierarchy sonority class value plosives 1 fricatives 2 nasals 3 liquids 4 approximates 5 closed vowels 6

• pen vowels

7

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

Universals of Sound Systems (Exercise 2)

The repertoire and use of speech sounds in human languages are constrained. According to de Boer, explanations for those regularities can by divided in the following two classes:

◮ they are based on physiological features of the human

vocal tract and articulators

◮ they are based on the human audible frequencies ◮ they are based on innate human cognitive capabilities √ ◮ they are based on functional constraints of a good

communication system √

◮ enable learnable and robust communication ◮ redundancy and predictability ◮ easily to distinguish and to produce

Note: It is hard to find out how specific capacities might have become innate

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

Universals of Sound Systems

Sound systems of human languages are often optimized for criteria such as

◮ acoustic distinctiveness (especially vowels) ◮ articulatory ease (especially consonants)

→ f([m]) = 94%, f([ö]) = 1%

Fig: realization of German vowels Fig: vowel systems of human languages

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

Universals of Sound Systems

How do sound systems become optimized?

◮ when children learn, they don’t explicitly optimize, but

imitate members of their community

◮ their imitation is closer to the source material than

necessary for successful communication (see dialects)

◮ suggestion: optimization is caused by self-organization in

a population of language users

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

Self-Organization in a Language Community

Hypothesis: “...the structure of human vowel systems is determined by self-organization in a population under constraints of perception and production.”

Bart de Boer (2000), Self-organization in vowel systems

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

Self-Organization in a Language Community

Conditions for a model of self-organization

• 1. emergence of organization on a global scale: most

members have same inventories

• 2. emergence due to interaction between members, not
• ptimization-actions of single members
• 3. non-local influence between members: no direct access to
• ther members’ inventories
• 4. no pre-wired knowledge: members start with empty

inventories; or as tabula rasa

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

Self-Organization in a Language Community

Why a computer simulation model for a system of self-organization?

• 1. phenomenon of self-organization is hard to predict from

just the description of the system

• 2. computer simulations help to investigate life-like

phenomena

• 3. computer model formulates a hypothesis, its simulation

synthesizes the phenomenon

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

de Boer’s Model

Properties of the model:

• 1. population of agents
• 2. agents can produce, perceive and remember speech sounds

in a human-like way

• 3. agents interact with others by imitating them
• 4. agents update their repertoires in dependence of

interaction outcome (imitation success)

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

de Boer’s Model: Articulatory Space

Fig: Articulatory space (from de Boer 1999, page 64)

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

de Boer’s Model: Acoustic Space

Fig: Acoustic space (from de Boer 1999, page 42)

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

de Boer’s Model: Space Conversion

Fig: Space Conversion (from de Boer 2000, page 7)

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

de Boer’s Model: Agent Architecture

Fig: Agent architecture (from de Boer 2000, page 6)

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

de Boer’s Model: Agent Architecture (Exercise 3)

De Boer is using a prototype model for the agents to store

• vowels. In his model a prototype has the following features:

◮ non-static √ ◮ space-segmenting ◮ addable/removable √ ◮ unerasable ◮ articulatory/acoustic aspect √ ◮ feature-based

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

de Boer’s Model: Agent Interaction

Two randomly chosen agents interact via imitation games:

• 1. one as initiator, one as imitator
• 2. initiator i) selects random vowel var from repertoire, and

ii) synthesizes it to vac

• 3. imitator receives noisy v′

ac, chooses best imitation v′′ ar and

synthesizes it to v′′

ac

• 4. initiator does the same and checks if the resulting vowel

matches with var :yes → success, otherwise failure

Fig: Exemplary play of the imitation game (from de Boer 2000, page 12)

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

de Boer’s Model: Agent Interaction (Exercise 4)

In de Boer’s simulation model agents play an imitation game, which can be successful or not. What two possible courses of action can be performed by the imitator, if the game was not successful?

◮ a bad vowel prototype (low success/use ratio) is shifted

closer to the perceived signal √

◮ a new vowel prototype is added, which is a good imitation

• f the perceived signal √

◮ the vowel prototype that is closest to the perceived signal

is replaced by a new one, that is a good imitation of the perceived signal

◮ a randomly chosen vowel prototype is shifted closer to the

perceived signal

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

de Boer’s Model: Agent Interaction

Agent update mechanisms

Fig: Agent update mechanisms (from de Boer 2000, page 12)

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

Simulation and Analysis Results

◮ Simulation run: apparently realistic systems emerge

Fig: Exemplary run (from de Boer 2000, page 14)

◮ Analyses over 1000 runs, 5000 games each:

◮ high success rates, realistic inventory sizes ◮ energy of systems is near-optimal Fig: Statistical analysis (from de Boer 2000, page 15)

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

Simulation and Analysis Results

Classification of emerged systems

Fig: 6-vowel system classification (from de Boer 2000, page 21)

Result: systems A, B, C and E resemble the four most common 6-vowel inventories of human languages, (Schwartz et al. 1997)

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

Conclusion (Exercise 5)

De Boer’s mentions that production and perception might not be the only factors that determine the shape of human vowel systems, and he mentions further possible factors, like:

◮ analogy1 √ ◮ contextual effects ◮ historical factors2 √ ◮ learnability √ ◮ morphosyntactic structure

1Phonetic analogy describes the development from a sporadic rule to a

general phonological rule.

2e.g. language contact, borrowing

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

Conclusion (Exercise 6)

“The work described here has shown that it is possible to explain the universal tendencies of vowel systems as a re- sult of self-organization in a population under constraints of perception and production. This eliminates the need to pos- tulate an innate predisposition towards certain vowel sys- tems, as well as the need for explicit optimization by lan- guage users.”

Roland Mühlenbernd Introduction: The

• Evolut. Approach

Universals of Sound Systems A Study of Self-Organization

Agent Architecture Agent Interaction Simulation Results Conclusion

Further Studies Homeworks

Homeworks

◮ Read the article ‘The emergence of linguistic structure -

an overview of the Iterated Learning Model’ (Kirby and Hurford , 2002)

◮ solve the appropriate exercises given on ILIAS