1 Lecture 21: From language change to language evolution Professor Robert C. Berwick Menu • Administrivia: Lab 5-6; Project endgame • The illogical problem of language change • Social language learning and its implications • How did language evolve? 6.863J SP11 6.863J SP11 … L1 L2 L2 L1 T [2, 2] or T[2,1] αt(1) T [1, 1] or T[1,2] αt+1(1) The picture with 1 ‘gene’, 2 ‘alleles’ (languages) 6.863J SP11 Dynamics of the IL model Question: Can the IL model yield distinct language ‘species’? Answer: No! It yields neither stable language species, nor admits the possibility of rapid language change But why?2 6.863J SP11 Given IL, what is the steady-state language mixture after many generations? (Perhaps surprising) outcome: just a single equilibrium (steady-state) from all initial conditions (possible initial distributions of Italian, English speakers), a 50-50 mix of the two languages: (Plausible? Do you think that if we all sat around in this villa long enough we’d all wind up in this state after about 10 generations?) 6.863J SP11 Dynamics of Iterated Learning model: cannot account for language stability 0 95% 5% 50% Despite effective learnability (learnable w/ pr 1–ε), a homogeneous language community cannot be maintained & degenerates over time to a mix And because the update equation is linear, there can be no bifurcations, (or phase transitions) only gradual change wrt correspondingly smooth gradience in the external environment 6.863J SP11 Where did the IL model go wrong? • Single parent problem (single agent inheritance) • No ‘maturation time’ for learning (take as long as you want) • Replace these idealizations with 3 more biologically realistic ones: 1. Two or more “parents” (Social Learning) 2. Finite time to mature or learn a language (# examples, K) 3. Threshold “firing” model for learning (like gene expression levels) 6.863J SP11 2b. Social Learning: from 2 parents, siblings, peers…3 6.863J SP11 Social Learning 6.863J SP11 Comparing Iterated & Social Learning • We now choose a learning algorithm; Assume just 2 languages (as earlier) • Learning algorithm: cue-based learning (though this turns out not be relevant) • Examines the sentences it gets for surface evidence of a particular linguistic parameter setting • For example, whether a language displays argument-function form (e.g., with objects following verbs, as in Italian, English, …) or argument-function form (German, Bangla, Japanese…) • Wait until we see a certain % of such ‘cues’ (almost like gene expression levels) Hans hat das Buch gelesen Hans read the book verb object verb object 6.863J SP11 Cue-based learning 6.863J SP11 Evolutionary Dynamics of Cue-based Learning4 6.863J SP11 Evolutionary dynamics for SL, continued 6.863J SP11 3b. “Social Learning” Model: Summary • Replace these idealizations with more biologically realistic ones: – Two or more “parents” (Social Learning) – Finite time to mature or learn language (# examples, K) – Threshold “firing” model for learning (like gene expression levels) (A ‘cue’ for L1 for our purposes here is the set difference {L1\ L2}, e.g., C={German–English}, e.g., Ich weiß dass Karl kauft das Buch/I know that Karl buys the book) • The state update equation for αt+1 is a polynomial of degree K: Where p is the probability of producing a trigger out of the totality of sentences spoken 6.863J SP11 The Social Learning model has a much richer dynamics than single parenting • Polynomial update equation: so potential for bifurcation (phase transitions, chaos) • Hence possible abrupt changes wrt external input cues • Possibly multiple fixed points (stable language mixtures other than a 50-50 mix) • Can this happen? • Yes, and this establishes ‘attractor basins’ that form the ‘clusters’ corresponding to language species 6.863J SP11 Dynamical system analysis: phase diagram of fixed points, for particular maturation time Frequency of L2 speakers = –verb 2nd language (eg, English), % All L2 regime All L1 regime Frequency of L2 cues, p, % 100 0 100 = L1speakers = L2 speakers Blue Red 100% L2speakers 100% L1speakers5 6.863J SP11 Dynamical system analysis Frequency of L2 speakers = –verb 2nd language (eg, English), % Frequency of L2 cues, p, % 100 0 100 = L1speakers = L2 speakers Blue Red 100% L2speakers 100% L1speakers For L2 cues from 0 until ‘critical pt’ 0.42, all L1 speakers 6.863J SP11 Dynamical system analysis Frequency of L2 speakers = –verb 2nd language (eg, English), % Frequency of L2 cues, p, % 100 0 100 = L1speakers = L2 speakers Blue Red 100% L2speakers 100% L1speakers Past critical point 0.44, a mix of L1, L2 speakers, depending on initial conditions 6.863J SP11 Dynamical system analysis: stable mix of 2 languages Frequency of L2 speakers = –verb 2nd lanauge (eg, English), % Frequency of L2 cues, p, % 100 0 100 = L1speakers = L2 speakers Blue Red 100% L2speakers 100% L1speakers 10-90% L1–L2 mix 6.863J SP11 Dynamical system analysis: unstable point Frequency of L2 speakers y = –verb 2nd language (eg, English), % Frequency of L2 cues, p, % 100 0 100 = L1speakers = L2 speakers Blue Red 100% L2speakers 100% L1speakers 30-70% L1–L2 mix decays to 100% L16 6.863J SP11 Empirical application: loss of “Verb second” in Middle English Historical empirical assumptions • Middle English changed from a Verb 2nd (V2) language to a non-Verb 2nd language during the 13th century, and never looked back • But why? • Lightfoot, Roberts: There were two English-like languages in an historical contact situation: (1) Northern (Middle) English; and (2) Southern (Middle) English • Northern (Middle) English was a V2 language; limited to subjects with pronouns (we, she, …); Southern (Middle) English was a non-V2 language 6.863J SP11 English History • The grammars for the two languages generated different surface phenotypes (= ‘cues’) • Some sentences possible only in Northern English: [In Amsterdam][verb saw] we many students (verbal elements in 2nd position) • Some sentences possible in both, when the subject was not a pronoun: In Amsterdam saw John the students • Some examples possible only in Southern English: In Amsterdam we saw many students (not verb-2nd) • Hypothesis (linguistic form): When speakers of Northern
View Full Document
Unlocking...