Iterated Complexity Bias

Transcript

1. Molly Lewis Michael C. Frank 25 July 2015 Conceptual complexity and the evolution of the lexicon
2. dog calculator sesquipedalian length Accounts appealing to form – Frequency (Zipf, 1936) – Predictability in context (Piantadosi, Tily, & Gibson, 2011; Mahowald et al., 2013) Accounts appealing to meaning – Not considered because language thought to be arbitrary (Saussure, 1916)
3. Complexity Bias Languages encode conceptually more complex meanings with longer linguistic forms. tupabugorn (Lewis, Sugarman, & Frank, 2014)
4. Complexity Bias: Two pieces of converging evidence r= 0.69 1 2 3 4 5 6 7 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Word Length (characters) Complexity Rating tupabugorn Generalizes cross-linguistically (Lewis, Sugarman, & Frank, 2014) Productive complexity bias Lexical complexity bias
5. Where does this bias come from? Ecological pressures shape language use, and through transmission, ultimately shape language structure. (Christiansen & Chater, 2008; in press) Language use timescale (minutes) Language change timescale (many years) t
6. What are the ecological pressures? The Communicative Hypothesis: Linguistic structure emerges from the interaction of speaker constraints (e.g., memory) and listener constraints (i.e., expressivity). Speaker: memory constraints energetic constraints Listener: expressivity constraint (Zipf, 1949; Regier, Kay, & Kherpertal, 2011; Kirby, et al. 2015)
7. The Communicative Hypothesis: A problem of linkage Complexity bias in individual speakers over time leads to the same regularity emerging in the structure of the lexicon through memory errors. Productive complexity bias Lexical complexity bias r= 0.69 1 2 3 4 5 6 7 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Word Length (characters) Complexity Rating
8. Present studies Study 1: Do memory errors lead to the emergence of a complexity bias? Study 2: How do these errors shape the structure of the lexicon through transmission?
9. Word learning Study 1: Lexical learning (10 word-object pairs x4) Word recall
10. Quintile 1 2 3 4 5 Least complex Most complex Objects
11. Study 1: Lexical learning Generated random lexicon – Words: 3, 5, 7, 9, 11 characters ([CV]+C syllables) – Objects: 2 from each complexity quintile Complexity bias: Shorten words for simple objects, lengthen words for complex objects ninop nin ninop ninopen
12. Study 1: Results Words are shortened for simple objects • • • • • −0.5 0.0 0.5 1.0 1.5 1 2 3 4 5 Complexity quintile Number characters removed Shorter words Complexity N = 50
13. Study 2: Iterated lexical learning Iterated learning paradigm – a method for simulating language change (e.g., Kirby, Cornish, & Smith, 2008) Gave the labels generated by participants to a new set of participants Iterated for total of 7 generations 50 participants/generation
14. An example chain damitobup nilobup nilobup nilobop nilobop nilop nilop nilop bipag bipag bippenbog buttenbug buttenbop bittenbop bittenbop bittenbop Gen. 0 Gen. 7
15. Study 2: Iterated lexical learning Word forms become more stable Across generations, words tended to: ①become easier to remember ②shorten ③increase in bigram transitional probability ④decrease in variability • • • • • • • • • • • • • • 0.00 0.25 0.50 0.75 1.00 1 2 3 4 5 6 7 Generation Proportion correct Mean accuracy
16. Levenshtein edit distance Metric of similarity between two strings The minimum number of character edits necessary to transform one string to another can à cat [1] can à calculator [8] Normalized edit distance: distance between actual vs. guessed word, normalized by length of longest
17. • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • 0.0 0.1 0.2 0.3 0.4 0.5 1 2 3 4 5 6 7 Generation Normalized edit distance Edit metric • • • • • • • • Levenshtein distance insertions substitutions deletions Levenshtein edit distance Study 2: Iterated lexical learning Change decreases across time Levenshtein edit distance substitutions deletions insertions
18. Study 2: Complexity Bias Length: cumulative characters removed (CCR) CCR Gen_N = length Gen_0 – length Gen_N Controls for variability in input length Same pattern for raw number of characters
19. Study 2: Iterated lexical learning Complexity bias persists across time 1 2 3 4 5 6 7 • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • 0 1 2 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 Complexity quintile Cumulative characters removed Complexity bias across generations Shorter words Complexity
20. If memory is the driver of change, need errors to change system – Chains with greater cross-generational change in lexical forms tend to show an increase in complexity bias over time. – When errors are made, consistent with complexity bias Pressure to simplify suppresses complexity bias – Because task not communicative, insufficient pressure against compression But, why doesn’t it strengthen?
21. Conclusion Communicative Hypothesis as origin of complexity bias -- but, problem of linkage In an iterated learning paradigm, find: – Language becomes more stable over time – Memory errors lead to complexity bias – Change in bias may be related to memory demands and listener pressures
22. Thank you