1. Chapter 1. Introduction From “The Computational Nature of Language Learning and Evolution” Summarized by Seok Ho-Sik

2. © 2009 SNU CSE Biointelligence Lab 2 Contents 1.1 Language Acquisition 1.2 Variation – Synchronic and Diachronic 1.3 More Example of Change 1.3.1 Phonetic and Phonological Change 1.3.2 Syntactic Change 1.4 Perspective and Conceptual Issues 1.4.1 The Role of Learning 1.4.2 Populations versus Idiolects 1.4.3 Gradualness versus Abruptness (or the S-Shaped Curve) 1.5 Evolution in Linguistics and Biology 1.5.1 Scientific History 1.6 Summary of Results 1.6.1 Main Insights 1.7 Audience and Connections to Other Fields 1.7.1 Structure of the Book

3. © 2009 SNU CSE Biointelligence Lab 3 Introduction (1/4) How to discriminate between well-formed expressions and ill- formed ones? An English-speaking community  V: vocabulary, E i : the set of acceptable sentences for individual i.  I i : the internal system of rules characterizing the linguistic knowledge, E =  i E i : the set of sentences that would be considered grammatically acceptable by everyone. English learning  : internal system of rules acquired by a child.  : mirroring E If all children acquired the language of their parents, languages would not change with time!

4. © 2009 SNU CSE Biointelligence Lab 4 Introduction (2/4) The set E in 878 A.D. is quite different from what it is today. OV sentences

5. © 2009 SNU CSE Biointelligence Lab 5 Introduction (3/4) Language acquisition vs. Language change  A child learns the language of its parents successfully.  We have language change – the language of a community drifting over generational time.

6. © 2009 SNU CSE Biointelligence Lab 6 Introduction (4/4) Given the children attempt to learn the language of their parents and caretakers  Why do language changes with time?  Why don’t they remain stable over all time?  How fast do they change?  What are the factors that influence change?

7. © 2009 SNU CSE Biointelligence Lab 7 1.1 Language Acquisition (1/4) The question: how we come to acquire our native language. Learning a language: developing a system of rules (a grammar) on the basis of linguistic examples encountered during the learning period.  A: D  g (D: data, g: grammar)  Generalization: of all the different grammars, the child develops a particular one – one that goes beyond the data and one that is remarkably similar to that of its parents in normal and homogeneous environments. Logical problem of language acquisition  Target grammar: is a target grammar from a class of possible target grammars  Example sentences: example sentences are generated by the target grammar and presented to the learner.  Hypothesis grammars h  H : hypothesis grammars drawn from a class of possible hypothesis grammars constructing on the basis of exposure to example sentences.  Learning algorithm A : an effective procedure by which grammars from H are selected.

8. © 2009 SNU CSE Biointelligence Lab 8 1.1 Language Acquisition (2/4) In a natural language-acquisition setting,  Learners are exposed to sentences of the ambient language as a result of spoken interaction with the world  Their linguistic experience consists of example sentences  Constituting primary linguistic data. Successful generalization to novel sentences is the key aspect of language acquisition.  The learner’s hypothesis converges to the target ( in some sense indicated by the distance metric ) as the data goes to infinity Language acquisition: the process by which a grammar is learned so that when novel sentences are encountered, the learner will be able to correctly judge their grammaticality and in fact will be able to produce ones of their own as well.

9. © 2009 SNU CSE Biointelligence Lab 9 1.1 Language Acquisition (3/4) A number of different aspects of this framework  1. G or H could represent grammars for syntax in more traditional generative linguistic traditions such as Government and Binding theory (GB), Minimalism, Head-driven Phase Structure Grammar (HPSG), Lexical-Functional Grammar (LFG), Three Adjoining Grammar (TAG). It might represent syntactic grammars with less traditional notational systems.  2. The example sentences might be strings of lexical items, annotated lexical strings, parse trees, pairings of syntactic structure with semantic representation.  3. The learning algorithm will depend upon the representations used for grammars in H and examples s i. Depending upon the domain and the phenomena of interest, an appropriate notational system for grammars and a cognitively plausible learning algorithm is used in formal explorations in th language acquisition.

10. © 2009 SNU CSE Biointelligence Lab 10 1.1 Language Acquisition (4/4) Various learning-theoretic frameworks  Probably Approximately Correct Framework, inductive inference framework…  The inherent difficulty of inferring an unknown target from finite resources.  Tabula rasa learning is not possible.  The learner do not entertain every possible hypothesis consistent with the data but only a limited class of hypothesis  This class of hypothesis is the class of possible grammars a learner can conceive  These constrain the range of possible language that can be invented and spoken  It is Universal Grammar (UG). The inherent intractability of learning a language in the absence of any constraints implies that the only profitable direction is to try and figure out what the appropriate constraints are.  Linguistic theory: the nature of the constraints underlying H.  Psychological learning theory: plausible learning algorithms A

11. © 2009 SNU CSE Biointelligence Lab 11 1.2 Variation – Synchronic and Diachronic (1/2) The variation across languages  Two languages might have  Different lexicons but similar syntactic systems.  Similar lexicons but different syntactic systems.  Similar lexical and syntactic properties yet have very different phonological systems. Synchronic variation: the variation across individuals in space at any fixed point in time. Diachronic variation: the possible behaviors of linguistic systems changing with time.  This book concerns itself with variation along a different dimension – the variation in the language of spatially linguistic communities over generational time  There is a need for a dynamical systems changing with time.

12. © 2009 SNU CSE Biointelligence Lab 12 1.2 Variation – Synchronic and Diachronic (2/2) A homogeneous linguistic community where all adults speak L h1 If a child is given a infinite number of sentences The next generation would consist of homogeneous speakers of L h1 If a child is given N sentences A small portion  end up acquiring L h2 What happens to  ? - Increasing - Decreasing - Stable state - Bouncing back and forth Depending on - How similar the L h1 and L h2 are? - What is the size of N? - What is the learning algorithm? - What it the probability with which sentences are presented to learner?

13. © 2009 SNU CSE Biointelligence Lab 13 1.3.1 Phonetic and Phonological Change Various possibilites  Individual learning by children; tendencies by speakers, listeners, and learner to avoid gaps and reduce homophonies. Why would rules arise and be lost?  Variation existing in the population.  How might learning by children, frequency of usage of different forms, and variation in the population interact to create the circumstances under which a rule might be gained and the circumstances under which a rule might be lost?

14. © 2009 SNU CSE Biointelligence Lab 14 1.3.2 Syntactic Change Pro-drop: in some language of the world, the pronominal subject of a sentence has to be present in the surface form for the sentence to be deemed grammatical in that language. Modern Italian allows one to drop the subject if the putative subject can be unambiguously inferred by pragmatic or other considerations. Old French used to allow pro-drop, while Modern French does not  Then why?

15. © 2009 SNU CSE Biointelligence Lab 15 1.4 Perspective and Conceptual Issues 1. Language has form  Formal Language Theory to describe linguistic form and linguistic structures. 2. Language is learned  Learning Theory to characterize the problem of language acquisition and learning. 3. Languages vary  Dynamic Systems to characterize the diachronic evolution of linguistic populations over time. What is the precise nature of relationship between language acquisition and language change?  Language change is contingent on language learning.  …if languages were learned perfectly by the children of each generation, then language would not change (Henry Sweet, 1899). Gradualness vs. Abruptness  The rate and time course of language changes.  S-shaped curve denoting the change in linguistic behavior over successive generations.  A given change begins quite gradually; after reaching a certain point, it picks up momentum and proceeds at a much faster rate and finally tails off slowly before reaching completion. The result is an S-curve,… Long-term change in a language is complicated by compounding factors  Sociopolitical considerations  The undue influence of one person or groups of persons might result in the propagation of their linguistic preference.

16. © 2009 SNU CSE Biointelligence Lab 16 1.5 Evolution in Linguistic and Biology In the case of human language, the mode of transmission is  Learning rather than genetic reproduction  Learning by children result in linguistic similarity between parent and child.  Children acquire their language based on the linguistic composition of the parental generation at large  language evolution is more like epidemiology or ecology than like Mendelian genetics. Because individual learning is the mechanism by which language is transmitted from the speakers of one generation of those of the next, the theory of learning will play a central role in the development of the evolutionary models  the linguistic evolution models are distinct from those that encountered on evolutionary biology. Sufficient conditions for evolution by natural selection In the case of language evolution The principle of variationIn any population, there is variation in linguistic behavior. The principle of heredity This variation is in part inheritable through the mechanism of inheritance is based on learning from larger group of people. The principle of differential fitness Fitness may be viewed as the differential transmission of linguistic variants.

17. © 2009 SNU CSE Biointelligence Lab 17 1.6 Summary of Results (1/4) 1. Linguistic behavior and underlying linguistic knowledge may be characterized as a formal system. H : the range of such formal systems. 2. Variation within any population (at time t ) may be characterized by a probability distribution P t on H. For h  H, P t (h) is the proportion of individuals using the system h. 3. An individual child born within the population will acquire language based on a learning algorithm A that maps its primary linguistic data on linguistic systems. Stipulation (1), (2), (3) taken together will make it possible to deduce a map P t  P t+1 that characterizes how linguistic variation evolves over time.

18. © 2009 SNU CSE Biointelligence Lab 18 1.6 Summary of Results (2/4) The logic of language evolution. There is grammatical variation in the population of parents. g 1, g 2, g 5 are some of the grammatical systems attested in the parental generation as shown (D 1 and D 2 are two different data sets that two different children receive). Each child has the same language learning algorithm A that maps these different linguistic experiences on to different grammatical systems (g and h respectively). Thus there is variation in the next generation of speakers. The sense in which natural selection plays a role in language evolution is unclear. No commitment is made at this point to the precise nature of g’s or the learning algorithm A.

19. © 2009 SNU CSE Biointelligence Lab 19 1.6 Summary of Results (3/4) Learning and Evolution  Learning at individual level and evolution at the population level are related.  The questions:  Difference between many different learning algorithms and their evolutionary consequences.  The role of critical age periods (the maturation parameter) in learning and evolution.  Difference between learning algorithms that learn from the input provided by a single individual vs. algorithms that learn from the input provided by the community.  Bifurcations in the History of Language  The role of bifurcations in population dynamics as an explanatory construct to account for major transitions in language.  (1) the dynamics is typically nonlinear, and (2) there are bifurcations which may be interpretable in linguistic terms as the change of language from one seemingly stable mode to another.

20. © 2009 SNU CSE Biointelligence Lab 20 1.6 Summary of Results (4/4)  Natural Selection and the Emergence of Language  What is the relationship between communicative efficiency and fitness, social connectivity, learnability, and the emergence of shared linguistic systems?  The complexity of the class of possible grammars H, the size of the learning set n, and the confidence  are all related.  If one learns from parents alone, then natural selection based on communicative fitness is necessary for the emergence of a shared linguistic system.  If one learns from the community at large, then natural selection is not necessary.