Lecture18-ComplexSystems4

Lecture18-ComplexSystems4 - Complex Linguistic Systems...

Info iconThis preview shows page 1. Sign up to view the full content.

View Full Document Right Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: Complex Linguistic Systems Psych 215L: Language Acquisition What is the generative system that creates the observed (structured) data of language (ex: metrical phonology)? Observable data: stress contour EMphasis Lecture 18 Complex Systems Complex Linguistic Systems Why look at English stress? What is the generative system that creates the observed (structured) data of language (ex: metrical phonology)? Legate & Yang (2011) Observable data: stress contour - Some empirical data on children’s stress knowledge is available (H L) H EM pha sis (S S)S EM pha sis EMphasis - English stress is a volatile area theoretically (a hard system to capture because of all the exceptions); developmental work could shed light on the target state for adults (e.g., see also Pearl (2011), who worries about this) - Because of the need to capture both “core” and “exceptional” data, the system is very interesting from a developmental point of view (H L L) EM pha sis (S S S) EM pha sis A quantitative description of productivity A quantitative description of productivity Legate & Yang (2011) Legate & Yang (2011) Idea: Want to describe productive data more compactly by using abstract rules, and leave less productive data to be memorized Logic: If a rule has too many exceptions, it takes a really long time to find the right thing to do. In fact, it may take longer than simply listing out each word explicitly (that is, treating everything as if it were an “exception” that had to be memorized). Elsewhere Condition, using a serial search procedure Basic format: Does it match this exception, or this one, or this one? If not, go to the ELSEWHERE condition because it must follow the normal rule. {order by type frequency} m ore frequent exceptions accessed faster IF word is elephant, then… IF word is octopus, then… IF word is teddy bear, then… … ELSE { do something else} Exceptions vs. Exceptions Rule A quantitative description of productivity A quantitative description of productivity Legate & Yang (2011) Legate & Yang (2011) Logic: If a rule has too many exceptions, it takes a really long time to find the right thing to do. In fact, it may take longer than simply listing out each word explicitly (that is, treating everything as if it were an “exception” that had to be memorized). Logic: If a rule has too many exceptions, it takes a really long time to find the right thing to do. In fact, it may take longer than simply listing out each word explicitly (that is, treating everything as if it were an “exception” that had to be memorized). What to do for giraffe? What to do for giraffe? Time = # exceptions + rule application Exceptions giraffe? No… giraffe? No… vs. giraffe? No… giraffe? No… Rule Time = # exceptions + rule application Exceptions Exceptions Rule Okay - apply rule. vs. Time = # exceptions giraffe? giraffe? Exceptions giraffe? giraffe? giraffe? No… No… No… No… Yes! A quantitative description of productivity A quantitative description of productivity Legate & Yang (2011) Legate & Yang (2011) Suppose we have N lexical items to deal with, with m exceptions. T(N, m) = time to wait to find what to do, on average, if rule is productive. Normal items (N-m of them) will have to wait for m exceptions to be looked at before the rule can be applied. Exceptional items processed faster (<m). Suppose we have N lexical items to deal with, with m exceptions. T(N, N) = time to wait to find what to do, on average, if rule is not productive. All items (N of them) will be searched for (and eventually found) in the list of “exceptions”. Exceptions vs. Exceptions Rule Exceptions vs. Exceptions Rule A quantitative description of productivity English Past Tense Example Legate & Yang (2011) Legate & Yang (2011) Tolerance Principle: A rule will be represented as productive if T(N,m) < T(N,N). Regular past tense: +ed But also many irregulars: Yang (2005): This happens when m < N / ln N How many regulars needed to make +ed a productive rule? Productive representation Exceptions Rule Non-productive representation vs. Exceptions kissed hugged missed tugged catch-caught, go-went, drink-drank, … Productive representation Exceptions Rule Non-productive representation vs. Exceptions English Past Tense Example Legate & Yang (2011) Regular past tense: +ed But also many irregulars: Legate & Yang (2011) kissed hugged missed tugged catch-caught, go-went, drink-drank, … If irregulars (m) < N (total) / ln N (total), then it’ s productive. Productive representation Exceptions Non-productive representation vs. Exceptions Rule Legate & Yang (2011) kissed hugged missed tugged catch-caught, go-went, drink-drank, … If there are 120 irregulars, then need 800 total (680 regulars) for that to be true: 120 =~ 800 / ln 800. Turns out there are way more than this. Productive representation Exceptions Non-productive representation vs. Exceptions English Stress: Nouns vs. Verbs Legate & Yang (2011), based on Halle & Vergnaud (1987) & Hayes (1995) Also works to explain why there is German feminine plural (+n) productivity, and why there is irregular English unproductivity (e.g. drink-drank, sink-sank, …), despite there being several verbs following irregular patterns. Productive representation Rule Regular past tense: +ed But also many irregulars: Rule Tolerance Principle Success Exceptions English Past Tense Example Non-productive representation vs. Exceptions Nouns: main stress final if VV (kangaroo), otherwise next to last if VV/VC (horizon), otherwise antepenult (Canada) Verbs: main stress final if VV/VC (introduce), otherwise next to last (consider) English Stress: Nouns vs. Verbs Legate & Yang (2011), based on Halle (1998) Nouns: main stress antepenult if two light syllables follow (Canada), otherwise penult if VV/VC penult (horizon, stalagmite), otherwise final (malachite) Learning English Stress Legate & Yang (2011): Learner Assumptions Verbs: main stress final if VV/VC (introduce), otherwise penult (develop) - phonemes are known (experimental evidence that this happens by age 1: Werker & Tees 1983, Kuhl et al. 1992) - phonotactic rules are in place that allow for native-like syllabification (experimental evidence this happens by 9 months: Jusczyk et al. 1992) - word segmentation is online (experimental evidence this could happen as early as 7.5 months: Jusczyk & Aslin 1995) - detection of main stress in a word is online (evidence that basic stress patterns (foot-headed-left (trochaic) vs. foot-headed-right (iambic)) are known by 7.5 months: Jusczyk, Cutler & Redanz 1993, Jusczyk, Houston & Newsome 1999.) Learning English Stress Learning English Stress Legate & Yang (2011): Not cue-based learning (Dresher 1999) Legate & Yang (2011): How Tolerance Principle learning works - cue-based approach: metrical parameters are set in an ordered sequence, based on cues observed in the linguistic input - “we envision the learner experimenting and evaluating the core metrical hypotheses in an incremental fashion as he processes linguistic data, and the learner chooses the grammar most highly valued with respect to the present data…” - L & Y argue: “A major motivation for learning as a sequence of decisions is to uphold the idealization of the child as a deterministic learner. For instance, suppose the child has not yet determined the quantity sensitivity of the language: if he proceeds to the stress placement parameters in a quantity sensitive language such as English, he might as well need to retreat from these parameters. But this idealization of deterministic learner is both empirically problematic and formally unnecessary…More important, and more general to the theory of language and language learning, is the issue of balancing generalizations with exceptions.” “Each grammar Gi, then, can be associated with a tuple (Ni, mi ), the number of words (Ni) it could apply to, and the number of words that contradict it (mi).” Learning English Stress Learning English Stress Legate & Yang (2011): Input Legate & Yang (2011): What’ s being learned 4.5 million tokens (26,700 types) pulled from all of CHILDES (2000) -only use the 20,000 types that are either nouns or verbs “For the present paper, we only consider the placement of the main stress. Since the pronunciation dictionary marks primary, secondary, as well as no stress, we mark the former as 1 and collapse the latter two as 0. For instance, the word animals will be represented syllabically as LLC with the stress contour of 100.” - “ introduce” would be marked as 001 - “ stalagmite” would be marked as 010 “Since nouns and verbs have somewhat different stress patterns, considering them together will pose a realistic test for any model that seeks systematic regularity amidst a heterogeneous mix of patterns.” Assume morphological knowledge: “Based on the consistent developmental evidence that the inflectional morphology is acquired relatively early−in some languages very early−we assume that the learner is capable of parsing words into morphological structures and considering their roles in the acquisition of stress.” - necessary if dividing nouns from verbs, but unclear children have this until about 3 years old in English… Learning English Stress Legate & Yang (2011): The input distribution from an extrapolated early vocabulary “For simplicity, we only consider two specific points of stress development, one designed to capture the childʼs stress system under a very small vocabulary and the other when the child has already learned enough words to potentially match the target state.” Learning English Stress Legate & Yang (2011): The input distribution for an extrapolated later vocabulary Considering words that occur at least once per million: 4047 nouns, 2402 verbs [some overlap, so only 5763 distinct types total] How many follow a quantity-insensitive, trochaic grammar? 4533 This leaves 5763-4533 = 1230 exceptions. How many exceptions can 5763 types tolerate? 5763/ln(5763) =~ 668. This grammar is no longer productive. “The distribution in Table 1 is clearly consistent with a quantity insensitive trochaic [ft-hd-left] system [assuming feet of size 2 syllables]. A total of 402 words can tolerate 402/ ln 402 = 67 exceptions where in fact there are 26 [01, 010, 100, 1000] . Interestingly, children learning English and similar languages go through an initial stage, which terminates at about 2;0, during which the child is limited to a maximum bisyllabic template with the primary stress falling on the first [syllable]…” Option? Try quantity-sensitive grammars instead. Learning English Stress Legate & Yang (2011): The morphology assumption plays out Learning English Stress Legate & Yang (2011): Evaluating adult target grammars “…we assume that the learner is capable of relating inflectional forms of verbs to their stem forms, but is incapable of parsing derivational forms into decomposable pieces (words such as growth and government will be treated as morphologically simplex)…” “…we assume that the learner has correctly learned that inflectional suffixes do not trigger stress shift - a t ask easily accomplished, again, by the use of the productivity model: there are no exceptions to the lack of stress shift with inflectional morphology. In other words, the child treats all inflectional forms of walk (i.e., walk, walks, walked and walking) as walk for the purposes of stress acquisition…” [Worth thinking about?: what age the child is able to recognize this and when stress acquisition happens] Learning English Stress Legate & Yang (2011): Other useful results Why the diminuitive “ee” doesn’ t get stressed (kitty, doggie/doggy): If lexical decomposition is allowed (kitty = kit + diminuitive “ee”), child can decide on productive rule that never stresses the diminuitive, based on experience with diminuitive forms in the input. Learning when morphological affixes cause main stress shifting: (document vs. documentary, tone vs. tonic, r idicule vs. ridiculous) +lex = distinguish nouns from verbs (perhaps using semantic knowledge?) +stem = inflectional decomposition of verbs (walking = walk + ing) HV87 = Halle & Vergnaud 1987, Hayes 1995 grammar H98 = Halle 1998 grammar Basic results: H98 grammar is productive while HV87 is not, but has more exceptions if inflectional decomposition isn’t allowed ...
View Full Document

Ask a homework question - tutors are online