treebanking for the
play

Treebanking for the grammar Acquisitionist my goldfish convinced - PowerPoint PPT Presentation

Sep 19, 2023 •222 likes •341 views

The learning problem A child exposed to sentences acquires a Treebanking for the grammar Acquisitionist my goldfish convinced me that linguists drive more expensive cars than banjo players (or: some times I wished I had a treebank) what did

  1. The learning problem A child exposed to sentences acquires a Treebanking for the grammar Acquisitionist my goldfish convinced me that linguists drive more expensive cars than banjo players (or: some times I wished I had a treebank) what did your goldfish convince you that linguists drive __ ? * who did your goldfish convince you that linguists drive more Jeffrey Lidz expensive cars than __ ? University of Maryland Department of Linguistics What determines the nature of projection of a system from a corpus? 2 Constraints on projection What the LAD typically does Rich, innate hypothesis space makes limited kinds of structures possible Chomsky 1965, Fodor 1966, Pinker 1979, ... Language Acquisition Input Grammar Device Language Acquisition Grammar Input Device Define space of possible grammars Accounts for how acquired grammar extends beyond experience 3 4

  2. reconstruction effects What the LAD typically does Miss Cruella knows which picture of herself Janie put up herself = Janie/Miss Cruella Language Acquisition Miss Cruella knows how proud of herself Janie was Input Grammar Device herself = Janie/*Miss Cruella 4-yr-olds know that reconstruction is optional for moved arguments but obligatory for moved predicates Define space of possible grammars Accounts for how acquired grammar extends Despite fact that speech to children includes no examples beyond experience of anaphors contained in wh-phrases Leddon & Lidz 2005 6 Inside a learning mechanism Inside a learning mechanism Encoding Encoding Hypothesis Hypothesis of of (Statistical) (Statistical) Input Space Input Space Experience Experience Inference Inference Innate structure is designed to learn Is structure of grammar reflected in input? What information is available for statistical inference? 7 8

  3. Kannada syntax & acquisition Kannada Ditransitives Is the structure of the grammar reflected in input to the learner? Is there any information in the input that Kannada learners could use to identify underlying 40m speakers structure? 9 10 Kannada Ditransitives Kannada Ditransitives Hari Rashmi-ge pustaka-vannu kalis-id-a Hari pustaka-vannu Rashmi-ge kalis-id-a Hari Rashmi-dat book-acc send-pst-3sm Hari book-acc Rashmi-dat send-pst-3sm Hari Rashmi-ge pustaka-vannu kalisi-koTT-a Hari pustaka-vannu Rashmi-ge kalisi-koTT-a Hari book-acc Rashmi-dat send-BEN-3sm Hari Rashmi-dat book-acc send-BEN-3sm ‘Hari sent the book to Rashmi’ ‘Hari sent the book to Rashmi’ 11 12

  4. Dimensions of variation Benefactive = DOC Possession +Ben -Ben Hari bangaloor-ige pustakavannu kalisida Hari pustakavannu bangaloor-ige kalisida Theme - Goal YES YES * Hari bangaloor-ige pustakavannu kalisikoTTa * Hari pustakavannu bangaloor-ige kalisikoTTa Goal - Theme YES YES Hari sent(*BEN) the book to Bangalore cf Eng: *Hari sent Bangalore the book 13 14 Possession? Prominence? -Ben +Ben -Ben +Ben Theme - Goal NO YES Possession NO YES Goal - Theme NO YES Prominence ?? ?? Structural correlates of ±Ben 15 16

  5. Binding Asymmetries Binding Asymmetries When can goal bind into theme? When can goal bind into theme? Hari pratiyobba lekhanige avaLa lekhanavannu kalisida Hari pratiyobba lekhanige avaLa lekhanavannu kalisikoTTa * Hari avaLa lekhanavannu pratiyobba lekhanige kalisida Hari avaLa lekhanavannu pratiyobba lekhanige Hari sent every author her article kalisikoTTa Hari sent(BEN) every author her article -Ben: only when goal precedes theme +Ben: always - word order is irrelevant 17 18 What about Kannada Binding Asymmetries When can theme bind into goal? -Ben +Ben Hari adara lekhanige pratiyondu lekhanavannu kalisida Possession No Yes Hari pratiyondu lekhanavannu adara lekhanige kalisida Prominence ?? GOAL Hari sent every article to its author Goal is syntactically prominent in +Ben -Ben: always - word order is irrelevant 19 20

  6. Binding Asymmetries Prominence in Kannada When can theme bind into goal? -Ben +Ben Hari adara lekhanige pratiyondu lekhanavannu kalisikoTTa Possession No Yes Hari pratiyondu lekhanavannu adara lekhanige kalisikoTTa Prominence Theme Goal Hari sent-BEN every article to its author Theme is syntactically prominent in -Ben +Ben: only when theme precedes goal 21 22 Underlying Configurations Two Questions Are children aware of asymmetries in prominence? The test: Kannada-learning 4-yr-olds subj subj V V How do learners identify which NP is prominent in Theme Goal Goal Theme ±Ben? Locative Ditransitive Possession Ditranstive English: Prep Dative English: DOC Kannada: -BEN Kannada: +BEN 23 24

  7. Method Bound True The girl gave every boy his horse Truth Value Judgement Task (Crain & McKee 1985) 6 Act out story with props & toys Meaning 1: TRUE 5 every boy gets his own Puppet watches story 3 horse Puppet gives summary 1 4 statement Meaning 2: FALSE every boy gets the alien’s Child tells Puppet if he’s horse right or wrong 2 25 26 Bound True Bound False girl every boy his horse return-(BEN) The girl gave every boy his horse girl his horse every boy return-(BEN) 6 6 Meaning 1: TRUE Meaning 1: FALSE 5 5 every boy gets his own 3 every boy gets his own horse 3 horse 1 4 Meaning 2: FALSE 1 4 Meaning 2: TRUE every boy gets the alien’s every boy gets the alien’s horse 2 horse 2 27 28

  8. Bound False Kannada Predictions girl every boy his horse return-(BEN) girl his horse every boy return-(BEN) -Ben +Ben 6 Possession No Yes Meaning 1: FALSE 5 every boy gets his own Prominence Theme Goal 3 horse 1 4 Meaning 2: TRUE every boy gets the alien’s horse 2 29 30 Kannada Predictions Kannada Predictions If 4-yr-olds understand the relation between If 4-yr-olds understand the relation between morphological form and syntactic prominence, then morphological form and syntactic prominence, then +BEN: Goal can always bind into Theme +BEN: Goal can always bind Theme -BEN: Goal can only bind into Theme when Goal -BEN: Goal can only bind Theme when Goal comes first comes first Bound True condition: “yes” except in Bound False condition: “no” except in -Ben Theme-Goal -Ben Theme-Goal 31 32

  9. Results: Bound True Results: Bound False 100 100 75 75 % judged % judged Ben Affix Ben Affix No Ben Affix No Ben Affix true true 50 50 25 25 0 0 Theme-Goal Goal-Theme Theme-Goal Goal-Theme The girl gave every boy his horse The girl gave every boy his horse Viau&Lidz 2008 33 Viau&Lidz 2008 34 Questions English Analogy Binding data is sparse Is there evidence of binding possibilities in speech to children? Gigaword NYT parsed corpus of English 385M words, 15M sentences I wish I had a Kannada treebank to answer this! V [QP] [ PP to [ NP ...pro...]] = 129 phi-feature match = 0 If no evidence of binding possibilities, is there some other information that could be used? V [QP] [ NP ...pro...] = 87 phi-feature match = 0 I wish I had a Kannada treebank to answer this too 35 36

  10. Syntactic Configurations A sketch of learning Inferences from surface form to abstract structure cued by statistical asymmetry subj subj animacy asymmetry in goal arguments V V Theme Goal Inanimate goal 12 times more likely to be used as PP- Goal Theme dative. (Bresnan & Nikitina 2003) Locative Ditransitive Possession Ditranstive if p(Str1 | inanimate goal) > p(Str2 | inanimate goal), English: Prep Dative English: DOC then Str1 = locative dative & Kannada: -BEN Kannada: +BEN Str 2 = possession dative Learning consists in selecting the appropriate abstract representation on the basis of surface form 37 38 A sketch of learning Theoretical Conclusion Skewed distribution of inanimate goals is a good Learners use statistics to select a grammatical cue to underlying structure system with properties not exhibited in the statistics themselves Abstract representations must leave a statistical This inference is licensed only if signature in surface form of language in order to be a) the representations are antecedently acquired available Learners use this statistical signature to identify b) learners expect structures to leave a grammatical, but only because they know what to detectable statistical signature on the input look for 39 40

  11. Practical Conclusion The acquisitionist wants to know when children can rely on their input to learn specific generalizations vs. need to project beyond their input what information is available to drive inferences from surface form to abstract grammatical structure Treebanks can be highly informative for answering these questions 41

More recommend