T ree A djoining Gramma rs Natural Language Syntax with T - - PowerPoint PPT Presentation

▶
t ree a djoining gramma rs natural language syntax with t
SMART_READER_LITE
LIVE PREVIEW

T ree A djoining Gramma rs Natural Language Syntax with T - - PowerPoint PPT Presentation

Aug 16, 2022 637 likes •911 views

T ree A djoining Gramma rs Natural Language Syntax with T A G Laura Kallmey er & Timm Lihte HHU Dsseldo rf WS 2012 17.10.2012 Natural Language Syntax with T A G 1 Outline 1 The derivation tree 2 Design

slide-1
SLIDE 1 T ree A djoining Gramma rs Natural Language Syntax with T A G Laura Kallmey er & Timm Li hte HHU Dsseldo rf WS 2012 17.10.2012 Natural Language Syntax with T A G 1
slide-2
SLIDE 2 Outline 1 The derivation tree 2 Design p rin iples fo r elementa ry trees 3 Sample derivations Natural Language Syntax with T A G 2
slide-3
SLIDE 3 Derivation trees (1): The
  • ntext
NP P eter VP NP↓ VP V NP↓ repaired NP the fridge VP AD V VP* easily derived tree derivation tree VP NP VP P eter AD V VP easily V NP repaired the fridge repaired Peter 1 easily 2 the_fridge 22 Natural Language Syntax with T A G 3
slide-4
SLIDE 4 Derivation trees (2): T A G derivations a re uniquely des rib ed b y derivation trees. The derivation tree
  • ntains:
no des fo r all elementa ry trees used in the derivation, and edges fo r all adjun tions and substitutions p erfo rmed throughout the derivation, and edge lab els indi ating the ta rget no de
  • f
the rewriting
  • p
eration. Whenever an elementa ry tree γ rewrites the no de at Go rn address p in the elementa ry tree γ′ , there is an edge from γ′ to γ lab eled with p . Natural Language Syntax with T A G 4
slide-5
SLIDE 5 Derivation trees (3): Go rn adresses F
  • r
the no de addresses
  • f
elementa ry trees, Go rn addresses a re used: The ro
  • t
has address ǫ (o r 0), and the i th daughter
  • f
the no de with address p has address pi . 1 2 3 21 22 31 311 312 Natural Language Syntax with T A G 5
slide-6
SLIDE 6 Linguisti analyses with L T A G What is an elementa ry tree, and what is its shap e? elementa ry trees

?

⇐ =

synta ti /semanti p rop erties
  • f
linguisti
  • bje ts

⇒

Synta ti design p rin iples from [F rank, 2002℄: Lexi alization F undamental T A G Hyp
  • thesis
(FTH) Condition
  • n
Elementa ry T ree Minimalit y (CETM)

θ

  • Criterion
fo r T A G

⇒

Semanti design p rin iples [Ab eill and Ramb
  • w,
2000 ℄

⇒

Design p rin iple
  • f
e onomy Natural Language Syntax with T A G 6
slide-7
SLIDE 7 Design p rin iples (1): Lexi alization Ea h elementa ry tree has at least
  • ne
non-empt y lexi al item, its lexi al an ho r.

⇒

All widely used gramma r fo rmalisms supp
  • rt
some kind
  • f
lexi alization! Reasons fo r lexi alization: F
  • rmal
p rop erties: A nite lexi alized gramma r p rovides nitely many analyses fo r ea h string (nitely ambiguous). Linguisti p rop erties: Synta ti p rop erties
  • f
lexi al items an b e a ounted fo r mo re dire tly . P a rsing: The sea r h spa e during pa rsing an b e delimited (gramma r ltering). [S hab es and Joshi, 1990 , Joshi and S hab es, 1991 ℄ Natural Language Syntax with T A G 7
slide-8
SLIDE 8 Design p rin iples (2): F undamental T A G Hyp
  • thesis
F undamental T A G Hyp
  • thesis
(FTH) Every synta ti dep enden y is exp ressed lo ally within an elementa ry tree. [F rank, 2002 ℄ synta ti dep enden y valen y/sub atego rization mo di ation binding . . . exp ressed within an elementa ry tree terminal leaf (i.e. lexi al an ho r) nonterminal leaf (substitution no de and fo
  • tno
de) ma rking an inner no de fo r
  • bligato
ry adjun tion Natural Language Syntax with T A G 8
slide-9
SLIDE 9 Design p rin iples (3): Cond.
  • n
Elementa ry T ree Minimalit y Condition
  • n
Elementa ry T ree Minimalit y (CETM) The synta ti heads in an elementa ry tree and their p roje tions must fo rm the extended p roje tion
  • f
a single lexi al head. [F rank, 2002 ℄ Note: W e
  • nly
use simple, non-extended p roje tions! XP X head
  • S|
VP . . . VP . . . . . . V . . . . . . sleeps . . . Natural Language Syntax with T A G 9
slide-10
SLIDE 10 Design p rin iples (4): θ
  • Criterion
fo r T A G

θ

  • Criterion
(T A G version) a. If H is the lexi al head
  • f
an elementa ry tree T, H assigns all
  • f
its θ
  • roles
in T. b. If A is a frontier non-terminal
  • f
elementa ry tree T, A must b e assigned a θ
  • role
in T. [F rank, 2002 ℄

= ⇒

V alen y/sub atego rization is exp ressed
  • nly
with nonterminal leaves! S NP VP V sleeps , VP V VP* seems Natural Language Syntax with T A G 10
slide-11
SLIDE 11 Mo di ation and fun tional elements Ho w to insert mo diers (easily) and fun tional elements ( omplementizers, determiners, do-auxilia ries, ...)? Either b y sepa rate auxilia ry trees (e.g., XT A G gramma r),
  • r
as
  • -an ho
r in the elementa ry tree
  • f
the lexi al item they a re asso iated with. S Comp S that NP VP V sleeps S NP VP V AP sleeps A easily Natural Language Syntax with T A G 11
slide-12
SLIDE 12 Mo di ation and fun tional elements In XT A G, mo diers and fun tional elements a re generally rep resented b y auxilia ry trees.

⇒

F
  • tno
des/A djun tions indi ate b
  • th
  • mplementation
and mo di ation.

⇒

Enhan ement
  • f
the CETM: (see [Ab eill and Ramb
  • w,
2000 ℄)
  • re
tree (follo wing CETM) + spine S Comp S* that VP VP* AP A easily Natural Language Syntax with T A G 12
slide-13
SLIDE 13 Prin iples related to semanti s See [Ab eill and Ramb
  • w,
2000℄. Predi ate-a rgument
  • urren e:
Ea h elementa ry tree asso iated with a p redi ate
  • ntains
a non-terminal leaf fo r ea h
  • f
its a rguments. Semanti an ho ring: Elementa ry trees a re not semanti ally void (to, that.) Comp
  • sitional
p rin iple: An elementa ry tree
  • rresp
  • nds
to a single semanti unit. Natural Language Syntax with T A G 13
slide-14
SLIDE 14 Design p rin iple
  • f
e onomy Design p rin iple
  • f
e onomy The elementa ry trees a re shap ed in su h a w a y , that the size
  • f
the elementa ry trees and the size
  • f
the gramma r is minimal. Natural Language Syntax with T A G 14
slide-15
SLIDE 15 Sample derivations Complementation with: NPs, PPs, adje tives, lauses (raising,
  • ntrolling),
... Mo di ation with: PPs, adje tives, pa rti les, temp
  • ral
lauses, relative lauses, ... Natural Language Syntax with T A G 15
slide-16
SLIDE 16 Sample derivations: NP
  • mplements
(1) John buys Bill a b
  • k.
Elementa ry trees: NP N John S NP↓ VP V NP↓ NP↓ buys NP N Bill NP Det NP* a NP N b
  • k
Derivation tree: buys John 1 Bill 22 book 23 a

ǫ

Natural Language Syntax with T A G 16
slide-17
SLIDE 17 Sample derivations: Sentential
  • mplements
(1) (2) Bill hop es that John wins. Elementa ry trees: NP N Bill S NP↓ VP V S∗ hop es S Comp S* that S NP↓ VP V wins NP N John Derivation tree: wins that

ǫ

hopes

ǫ

Bill 1 John 1 Natural Language Syntax with T A G 17
slide-18
SLIDE 18 Sample derivations: Sentential
  • mplements
(2) (3) John seems to lik e Bill. Elementa ry trees: VP V VP∗ seems S NP↓ VP VP NP↓ V to lik e Derivation tree: to_like John 1 seems 2 Bill 22 Natural Language Syntax with T A G 18
slide-19
SLIDE 19 Sample derivations: Sentential
  • mplements
(3) (4) John exp e ts [ Bill to win ℄. Elementa ry trees: S NP↓ VP V S∗ exp e ts S NP↓ VP V to win Derivation tree: to_win expe ts

ǫ

John 1 Bill 1 Natural Language Syntax with T A G 19
slide-20
SLIDE 20 Sample derivations: Sentential
  • mplements
(4) Question: Why is the sentential
  • bje t
rep resented as a fo
  • tno
de? The sentential
  • bje t
is realised as a fo
  • t
no de in
  • rder
to allo w extra tions: (5) Who do es John exp e t to win? Elementa ry trees: VP V VP* do es S NP↓ VP V S∗ exp e t S NP↓ S NP VP

ǫ

V to win Natural Language Syntax with T A G 20
slide-21
SLIDE 21 Sample derivations: Multiple an ho rs Multiw
  • rd
exp ressions and light verb
  • nstru tions
an b e rep resented b y elementa ry trees with multiple an ho rs: (6) John exp e ted [Ma ry to mak e a
  • mment℄.
S NP↓ VP V NP to mak e N
  • mment
NP Det NP∗ a Natural Language Syntax with T A G 21
slide-22
SLIDE 22 Sample derivations: Mo diers (7) The go
  • d
student pa rti ipated in every
  • urse
during the semester. NP Det NP* the N AP N∗ A go
  • d
NP N student S NP↓ VP V PP pa rti ipated P NP↓ in VP VP∗ PP P NP↓ during Natural Language Syntax with T A G 22
slide-23
SLIDE 23 Sample derivations: Relative lauses (8) The dog [who ate the ak e℄. NP Det NP* the NP N dog NP NP* S NP↓ VP V NP↓ ate Problem: Extrap
  • sed
relative lauses: (1) Someb
  • dy
i lives nea rb y [who i has a CD-burner℄. Natural Language Syntax with T A G 23
slide-24
SLIDE 24 Derivation trees = Semanti dep enden y stru ture ? The derivation tree is not alw a ys the semanti dep enden y stru ture, due to: indis ernibilit y
  • f
  • mplementation
and mo di ation in adjun tion, and missing links. Example fo r a missing link: (2) John laims [Bill seems to win℄ to_win laims

ǫ

John 1 Bill 1 seems 2 Natural Language Syntax with T A G 24
slide-25
SLIDE 25 Ab eill, A. and Ramb
  • w,
O. (2000). T ree adjoining gramma r: An
  • verview.
In Ab eill, A. and Ramb
  • w,
O., edito rs, T ree A djoining Gramma rs: F
  • rmalisms,
Linguisti Analyses and Pro essing , volume 107
  • f
CSLI Le ture Notes , pages 168. CSLI Publi ations, Stanfo rd. F rank, R. (2002). Phrase Stru ture Comp
  • sition
and Synta ti Dep enden ies . MIT Press, Camb ridge,MA. Joshi, A. K. and S hab es, Y. (1991). T ree-Adjoining Gramma rs and lexi alized gramma rs. T e hni al Rep
  • rt
MS-CIS-91-22, Depa rtment
  • f
Computer and Info rmation S ien e, Universit y
  • f
P ennsylvania. S hab es, Y. and Joshi, A. K. (1990). P a rsing with lexi alized tree adjoining gramma r. T e hni al Rep
  • rt
MS-CIS-90-11, Depa rtment
  • f
Computer and Info rmation S ien e, Universit y
  • f
P ennsylvania.