Structures and strings

1
Structures and strings

• Virginia Savova
• Johns Hopkins University

2
Thank you

• My adviser Robert Frank.
• My committee members
• Paul Smolensky
• Fred Jelinek
• Sanjeev Khudanpur
• Robert Berwick (here all the way from Boston)
• All of you (for being here so early)

3
Structures and strings

• What should structure tell us about ordered word
  sequences?
• grammaticality status
• interpretation
• agreement relations
• word order

Separate linearization component
4
Dimensions of structure
interpretation agreement relations
Dominance
grammaticality
Precedence
word order
5
Strict correspondence
Asymmetric c-command precedence Kayne, 1994
6
Why SC is a good idea

• Apparent parsimony
• No need for an external linearization component
• Strong claim
• dominance is all that matters, precedence is
  epiphenomenal
• cross-linguistic variation is confined to the
  dominance relation

7
Why SC is a bad idea

• Gotta Move
• Difference in precedence ? difference in
  dominance (even if no independent evidence
  exists)
• Can derive any word order unless movement is
  required to have other consequences.
• Non-canonical word order structural complexity

8
Troubling consequences

• Some languages are structurally more complex than
  others at spell-out
• Latin requires movement in simple clauses

9
Troubling consequences

• Mapping at LF must be language-specific

LOVED (Brutus, Caesar)
10
Troubling consequences

• Tree-sequence representation with spell-out at
  different levels

Spell-out English
Spell-out Latin
LF
11
What is the alternative?

• Let structure represent only dominance.
• Determine word order through a linearization
  procedure on the basis of
• structure (i.e. dominance)
• discourse
• morpho-phonology

12
Why is that a bad idea

• Must specify an external linearization component
  different for each language
• where is UG?
• Cross-linguistic variation is no longer confined
  to the dominance relation
• Can derive any word order

13
Why is that a good idea

• Can simplify our notion of structure
• discourse features can be distinguished from
  syntactic feature (no Top, Foc heads)
• word order difference need not imply structural
  difference
• precedence need not be defined on structures
• tree sequences are no longer required
• Cross-linguistic variation can be confined to the
  lexicon the precedence relation

14
Resolving the negative side

• I will specify an external linearization
  component which
• accounts for cross-linguistic variation in word
  order
• is restricted w.r.t. possible word orders
• (putting UG back into linearization)

15
Exploiting the positive side

• I will err on the side of structural simplicity
  by
• defining a hierarchical component based on
  Dependency Grammars
• dispensing with the notion of sequence-type
  structures
• externalizing discourse features

16
Dependency structure(Tesniere 1959, Hays 1964,
Gaifman 1965)

• Simple structural representation
• a set of binary asymmetric relations among
  sentential elements .
• Preliminary conditions
• Every word is dependent (modifies, is conditioned
  upon, is selected by) some other word.
• Exactly one word is (dependent on) ROOT.

17
Constituent versus Dependency Structures
Head of VP
Head of NP
18
Constituent versus Dependency Structures
19
Conditions on Dependency Structure

• Widely-accepted formal conditions
• Every word ? exactly one other word (head).
• Exactly one word ? ROOT.
• No cycles
• Projectivity no crossing arrows

Dependency structures are trees
Hudsons Adjacency principle A word can be
separated from its head only by a sister and
subordinates
20
Conditions on Constituent Structure

• Every node has a single mother
• Only one root
• No cycles
• No crossing branches

21
Word order in dependency structure

• Head-Dependent ordering constraints
• Spec Head Comp (English)
• Spec Comp Head (Latin)
• Linearization procedure
• bought
• George bought camel
• George bought a camel

bought
comp
spec
George
camel
spec
a
22
Word order in Constituent Structure

• Head Comp
• Spec Head
• Spec Comp
• Kaynes Antisymmetry These fall out of dominance
• PP These are binary parameters

23
Dependency grammar

• Lexicon (ROOT, buy, John, the, camel, sleep)
• Syntactic categories (PoS)
• (NNSG,NPL, NPROP VVTR, VINTR D, P, Adv, Adj,
  ROOT)
• Subcategorization frames for PoS
• Subcat frame ordered list of obligatory deps
  (typearguments)
• VTR 1N, 2N VINTR 1N NSG 1D
  NPROP ROOT1V
• Optional modifiers for PoS
• V Adv, P N Adj, P

24
Dependency (tree) structure Generation

• A possible generative procedure
• Insert ROOT
• Recursively
• For each unmarked node in the structure, insert
  dependents that satisfy some subcat frame for
  that node and mark it.
• For each marked node, insert a finite number of
  allowed modifiers
• For each PoS label, insert a lexical entry

25
Dependency structures are not trees

• Subject is dependent on both auxiliary and verb
• Heads compete for adjacency

George doesnt buy camels
26
Dependency structures are not projective

• Discourse-related word order variation
• Locality competes with discourse

Projectivity doesnt hold (arrows cross)
27
The structural sequence solution

• One solution structure is a graph-tree sequence
  (Hudson, 1995)
• subcategorization ( modification) relations
• morphosyntactic feature-driven relations
• trees related through transformation

preserve M links remove conflicting S links
Linearize
George doesnt buy camels
28
The conflict and competition solution

• Evidence of competition in word order
• locality against discourse-motivated order
• multiple heads against one another
• Conflicts are universal
• Resolutions are language-specific

Optimality Theory
29
Separating structures from strings

• Generalized acyclic graphs
• Linearization optimization over conflicting
  constraints

Other factors
30
Word Order Optimization

• Underlying representation
• structure
• discourse features
• Candidate surface representations
• all possible permutations of elements
• Constraints
• Local (Head-Dep) ordering
• Discourse constraints

31
Conditions on Dependency Structure

• New (less-restrictive) formal conditions
• Every word ? exactly one other word (head).
• Exactly one word ? ROOT.
• No cycles
• Projectivity no crossing arrows

Dependency structures are no longer trees but
directed acyclic graphs
32
Dependency structure II Generation

• A possible generative procedure
• Insert ROOT
• Recursively
• For each unmarked node in the structure, insert
  dependents that satisfy some subcat frame for
  that node and mark it. If any such dependent is
  already appropriately inserted, create a
  secondary dependent
• For each marked node, insert a finite number of
  allowed modifiers
• For each PoS label, insert a lexical entry

ROOT
VAUX
VTR
NPL
NPROP
Adv
D
33
Word order in compound tense clauses
preserve M links remove conflicting S links
Linearize
George doesnt buy camels
34
Typology of simple-tense word order

• Four basic word orders are uncontroversial
• SOV
• SVO
• VSO
• VOS
• Find a set of four constraints such that each
  word order violates at most one

35
(No Transcript)
36
Cognitive basis of constraints

• S O
• V O
• Adjacent(O,V)
• Adjacent(V, Edge)

Agent Patient Compute function inside
out Consistency, boundary marking
Tomlin 1986
37
Generalized constraints

• Precedence constraints

HdCmp The head must precede its complement
Observed in SVO, VSO, VOS SpcCmp The
specifier must precede the complement Observed
in SOV, SVO, VSO

• Adjacency constraints

HdCmp The head must be adjacent to its
complement Observed in SOV, SVO, VOS HdEdge
The head must be at the edge Observed in SOV,
VSO, VOS
38
(No Transcript)
39
Word order of compund subclauses

• Subordinate clause structure
• complementizer that
• dependency between subject and auxiliary
• subject is specifier of Aux and V
• V complement of Aux

comp
spec
comp
spec
40
Word order of compund subclauses

• Candidate set
• 5! 120 permutations
• Four constraints
• 6 possible winners
• 114 permutations are universally banned!
• (inferior under any strict ranking)

that
41
Word order of compound subclauses
5! 120 permutations 6 possible winners
42
Word order of main clauses with auxiliaries

• Must derive one additional word order
• SAuxOV
• German
• SVO/ SAuxOV/SOV/SOVAux
• Dinka
• SVO/SAuxOV/VSO/AuxSOV
• harmonically bounded by SOVAux, AuxSOV and SAuxVO

43
Word order of main clauses with auxiliaries

• Head Movement Empty head (C) must merge with its
  complement
• Consequence Long Head Movement is banned!
• One additional constraint
• cEdge
• (Empty) C should not be aligned with the edge of
  the utterance

C
44
Word order of main clauses with auxiliaries
45
Predicted cross-linguistic word order typology
4 unattested types
46
Rare word orders
47
Rare word orders

• What is the role of frequency in linguistic
  theory?
• rare languages might be an experimental error
• different frequencies might be due to historical
  accident
• might be due to the relative fitness of a
  language type

48
Constraints as soft biases in iterative learning

• Iterated learning model

Kirby 2001
learner1
learner2
hypothesis
hypothesis
data
data
Griffiths and Kalish 2005

• Iterated Bayesian learning
• finite discrete hypothesis space
• prior over hypotheses
• speakers generate samples from hypothesis

49
Prior of satisfied constraints
Equal weight SOV 3 22 SVO 3 22 VSO
3 22 VOS 3 22 OVS 1 6 OSV 1 6

• HdCmp
• (VO)
• HdEdge
• (...V or V...)
• HdCmp
• (VO)
• SpcCmp
• (SO)

50
Consequences for frequency

• In the limit, the distribution of language types
  converges to prior
• Before convergence
• 70 of the time after 15 iterations, the
  distribution of language types features the four
  high prior languages outnumbering the low prior
  languages
• the frequencies of high-prior languages differ
• ? the real world distribution is an instance of
  such distribution

51
Typical pre-convergence distribution
52
Summary of this talk

• Structure should be order-free.
• How can word order variation and universal
  tendencies be captured outside of the structural
  component?
• An account of cross-linguistic variation of basic
  word order
• simple-tense clauses
• complex-tense clauses
• subordinate clauses
• main clauses
• A possible link to frequency

53
Summary of things not in this talk

• An account of within language word order
  variation
• questions versus declaratives
• inversion in yes-no questions
• wh-fronting
• clitic versus non-clitic pronouns
• restrictions on collocations

54
The big picture

• Cross-linguistic variation
• is surprising if language is evolutionarily
  selected for/innate
• can be explained if individual languages are
  viewed as different solutions to an optimization
  problem
• Word order variation is due to multiple solutions
  of dependency structure to string mapping under
  cognitive constraints