Synchronous context-free grammar
Edinburgh MT lecture 12: Synchronous context-free grammar
Jul 25, 2015
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Synchronous context-free grammar
Phrase-based models
•Exact decoding is NP-hard.
•As a consequence of arbitrary permutation…
•…but real permutations are not arbitrary!
•Parameterization of reordering is weak.
•No generalization!
Garcia and associates .
Garcia y asociados .Carlos Garcia has three associates .
Carlos Garcia tiene tres asociados .his associates are not strong .
sus asociados no son fuertes .Garcia has a company also .
Garcia tambien tiene una empresa .its clients are angry .
sus clientes estan enfadados .the associates are also angry .
los asociados tambien estan enfadados .
la empresa tiene enemigos fuertes en Europa .
the company has strong enemies in Europe .the clients and the associates are enemies .
los clientes y los asociados son enemigos .the company has three groups .
la empresa tiene tres grupos .its groups are in Europe .
sus grupos estan en Europa .the modern groups sell strong pharmaceuticals .
los grupos modernos venden medicinas fuertes .the groups do not sell zanzanine .
los grupos no venden zanzanina .the small groups are not modern .
los grupos pequenos no son modernos .
Garcia and associates .
Garcia y asociados .Carlos Garcia has three associates .
Carlos Garcia tiene tres asociados .his associates are not strong .
sus asociados no son fuertes .Garcia has a company also .
Garcia tambien tiene una empresa .its clients are angry .
sus clientes estan enfadados .the associates are also angry .
los asociados tambien estan enfadados .
la empresa tiene enemigos fuertes en Europa .
the company has strong enemies in Europe .the clients and the associates are enemies .
los clientes y los asociados son enemigos .the company has three groups .
la empresa tiene tres grupos .its groups are in Europe .
sus grupos estan en Europa .the modern groups sell strong pharmaceuticals .
los grupos modernos venden medicinas fuertes .the groups do not sell zanzanine .
los grupos no venden zanzanina .the small groups are not modern .
los grupos pequenos no son modernos .
Same pattern:NN JJ → JJ NN
Phrase-based models do not capture this generalization.
Context-free grammar
Context-free grammar
S → NP VPNP → watashi waNP → hako woVP → NP V V → akemasu
Context-free grammar
S → NP VPNP → watashi waNP → hako woVP → NP V V → akemasu
S
Context-free grammar
S → NP VPNP → watashi waNP → hako woVP → NP V V → akemasu
S
Context-free grammar
S → NP VPNP → watashi waNP → hako woVP → NP V V → akemasu
S
NP VP
Context-free grammar
S → NP VPNP → watashi waNP → hako woVP → NP V V → akemasu
S
NP VP
Context-free grammar
S → NP VPNP → watashi waNP → hako woVP → NP V V → akemasu
S
NP VP
watashi wa
Context-free grammar
S → NP VPNP → watashi waNP → hako woVP → NP V V → akemasu
S
NP VP
watashi wa
Context-free grammar
S → NP VPNP → watashi waNP → hako woVP → NP V V → akemasu
S
NP VP
NP Vwatashi wa
Context-free grammar
S → NP VPNP → watashi waNP → hako woVP → NP V V → akemasu
S
NP VP
NP Vwatashi wa
Context-free grammar
S → NP VPNP → watashi waNP → hako woVP → NP V V → akemasu
S
NP VP
NP Vwatashi wa
hako wo
Context-free grammar
S → NP VPNP → watashi waNP → hako woVP → NP V V → akemasu
S
NP VP
NP Vwatashi wa
hako wo
Context-free grammar
S → NP VPNP → watashi waNP → hako woVP → NP V V → akemasu
S
NP VP
NP Vwatashi wa
hako wo akemasu
Context-free grammar
S → NP VPNP → watashi waNP → hako woVP → NP V V → akemasu
S
NP VP
NP Vwatashi wa
hako wo akemasu
Context-free grammar
S → NP VPNP → watashi waNP → hako woVP → NP V V → akemasu
S
NP VP
NP Vwatashi wa
hako wo akemasu
watashi wa hako wo akemasu
Synchronous context-free grammar
S → NP VPNP → watashi waNP → hako woVP → NP V V → akemasu
Originally: syntax-directed translation (Lewis & Stearns 1966; Aho and Ullman 1969)
Synchronous context-free grammar
S → NP VPNP → watashi waNP → hako woVP → NP V V → akemasu
S → NP VPNP → INP → the boxVP → V NP V → open
Originally: syntax-directed translation (Lewis & Stearns 1966; Aho and Ullman 1969)
Synchronous context-free grammar
S → NP VPNP → watashi waNP → hako woVP → NP V V → akemasu
S → NP VPNP → INP → the boxVP → V NP V → open
Originally: syntax-directed translation (Lewis & Stearns 1966; Aho and Ullman 1969)
Japanese is SOV. English is SVO.
Synchronous context-free grammar
S → NP1 VP2 / NP1 VP2
NP → watashi wa / I NP → hako wo / the boxVP → NP1 V2 / V2 NP1
V → akemasu / open
Originally: syntax-directed translation (Lewis & Stearns 1966; Aho and Ullman 1969)
Synchronous context-free grammar
S → NP1 VP2 / NP1 VP2
NP → watashi wa / I NP → hako wo / the boxVP → NP1 V2 / V2 NP1
V → akemasu / open
REQUIREMENT:one-to-mapping
between source and target nonterminals,
indicated by coindexes
Originally: syntax-directed translation (Lewis & Stearns 1966; Aho and Ullman 1969)
Synchronous context-free grammar
S → NP1 VP2 / NP1 VP2
NP → watashi wa / I NP → hako wo / the boxVP → NP1 V2 / V2 NP1
V → akemasu / open
Synchronous context-free grammar
S → NP1 VP2 / NP1 VP2
NP → watashi wa / I NP → hako wo / the boxVP → NP1 V2 / V2 NP1
V → akemasu / open
S S
Synchronous context-free grammar
S → NP1 VP2 / NP1 VP2
NP → watashi wa / I NP → hako wo / the boxVP → NP1 V2 / V2 NP1
V → akemasu / open
S S
Synchronous context-free grammar
S → NP1 VP2 / NP1 VP2
NP → watashi wa / I NP → hako wo / the boxVP → NP1 V2 / V2 NP1
V → akemasu / open
S S
Synchronous context-free grammar
S → NP1 VP2 / NP1 VP2
NP → watashi wa / I NP → hako wo / the boxVP → NP1 V2 / V2 NP1
V → akemasu / open
S S
NP VP NP VP
Synchronous context-free grammar
S → NP1 VP2 / NP1 VP2
NP → watashi wa / I NP → hako wo / the boxVP → NP1 V2 / V2 NP1
V → akemasu / open
S S
NP VP NP VP
Synchronous context-free grammar
S → NP1 VP2 / NP1 VP2
NP → watashi wa / I NP → hako wo / the boxVP → NP1 V2 / V2 NP1
V → akemasu / open
S S
NP VP NP VP
Synchronous context-free grammar
S → NP1 VP2 / NP1 VP2
NP → watashi wa / I NP → hako wo / the boxVP → NP1 V2 / V2 NP1
V → akemasu / open
S S
NP VP NP VP
watashi wa I
Synchronous context-free grammar
S → NP1 VP2 / NP1 VP2
NP → watashi wa / I NP → hako wo / the boxVP → NP1 V2 / V2 NP1
V → akemasu / open
S S
NP VP NP VP
watashi wa I
Synchronous context-free grammar
S → NP1 VP2 / NP1 VP2
NP → watashi wa / I NP → hako wo / the boxVP → NP1 V2 / V2 NP1
V → akemasu / open
S S
NP VP NP VP
NP V V NPwatashi wa I
Synchronous context-free grammar
S → NP1 VP2 / NP1 VP2
NP → watashi wa / I NP → hako wo / the boxVP → NP1 V2 / V2 NP1
V → akemasu / open
S S
NP VP NP VP
NP V V NPwatashi wa I
Synchronous context-free grammar
S → NP1 VP2 / NP1 VP2
NP → watashi wa / I NP → hako wo / the boxVP → NP1 V2 / V2 NP1
V → akemasu / open
S S
NP VP NP VP
NP V V NPwatashi wa I
Synchronous context-free grammar
S → NP1 VP2 / NP1 VP2
NP → watashi wa / I NP → hako wo / the boxVP → NP1 V2 / V2 NP1
V → akemasu / open
S S
NP VP NP VP
NP V V NPwatashi wa I
hako wo the box
Synchronous context-free grammar
S → NP1 VP2 / NP1 VP2
NP → watashi wa / I NP → hako wo / the boxVP → NP1 V2 / V2 NP1
V → akemasu / open
S S
NP VP NP VP
NP V V NPwatashi wa I
hako wo the box
Synchronous context-free grammar
S → NP1 VP2 / NP1 VP2
NP → watashi wa / I NP → hako wo / the boxVP → NP1 V2 / V2 NP1
V → akemasu / open
S S
NP VP NP VP
NP V V NPwatashi wa I
akemasu openhako wo the box
Synchronous context-free grammarS S
NP VP NP VP
NP V V NPwatashi wa I
akemasu openhako wo the box
Synchronous context-free grammarS S
NP VP NP VP
NP V V NPwatashi wa I
akemasu openhako wo the box
watashi wa hako wo akemasu
Synchronous context-free grammarS S
NP VP NP VP
NP V V NPwatashi wa I
akemasu openhako wo the box
watashi wa hako wo akemasu I open the box
Translation as parsing
watashi wa hako wo akemasu
Translation as parsingS
NP VP
NP Vwatashi wa
akemasuhako wo
watashi wa hako wo akemasu
Translation as parsingS
NP VP
NP Vwatashi wa
akemasuhako wo
S
NP VP
V NPI
open the box
watashi wa hako wo akemasu
Translation as parsingS
NP VP
NP Vwatashi wa
akemasuhako wo
S
NP VP
V NPI
open the box
watashi wa hako wo akemasu I open the box
Preliminaries
S → NP VPNP → watashi wa NP → hako woVP → NP V V → akemasu
Transform source grammar into Chomsky normal form:all productions in form X → w or X → YZ.
Preliminaries
S → NP VPNP → watashi wa NP → hako woVP → NP V V → akemasu
Transform source grammar into Chomsky normal form:all productions in form X → w or X → YZ.
S → NP VPNP → X Y
X → watashi Y → wa NP → Z W Z → hako W → woVP → NP V V → akemasu
Preliminaries
S → NP VPNP → watashi wa NP → hako woVP → NP V V → akemasu
Transform source grammar into Chomsky normal form:all productions in form X → w or X → YZ.
S → NP VPNP → X Y
X → watashi Y → wa NP → Z W Z → hako W → woVP → NP V V → akemasu
Q: how do synchronous productions interact with
this transformation?
Decoding
Decoding
•A binary-branching (i.e. CNF) grammar can produce a Catalan number of parses of an input sentence.
Decoding
•A binary-branching (i.e. CNF) grammar can produce a Catalan number of parses of an input sentence.
O((2n)!
(n + 1)!n!)
Decoding
•A binary-branching (i.e. CNF) grammar can produce a Catalan number of parses of an input sentence.
•Dynamic programming to the rescue!
O((2n)!
(n + 1)!n!)
ParsingNN → duck, pato
PRP → I, yo
VBD → saw, vi
PRP$ → her, ella
NP → PRP$1 NN2, PRP$1 NN2
VP → VBD1 NP2, VBD1 NP2
S → PRP1 VP2, PRP1 VP2
PRP → her, su
VB → duck, agacharseSBAR → PRP1 VB2, PRP1 VB2
VP → VBD1 SBAR2, VBD1 SBAR2
ParsingNN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
Parsing
I1 saw2 her3 duck4
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
Parsing
I1 saw2 her3 duck4
Xi,i+1 � (wi+1 = w) ⇤ (X ⇥ w)NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
Parsing
I1 saw2 her3 duck4
Xi,i+1 � (wi+1 = w) ⇤ (X ⇥ w)NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
Parsing
I1 saw2 her3 duck4
Xi,i+1 � (wi+1 = w) ⇤ (X ⇥ w)NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
Parsing
I1 saw2 her3 duck4
PRP0,1 � (w1 = I) ⇤ (PRP ⇥ I)Xi,i+1 � (wi+1 = w) ⇤ (X ⇥ w)NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
Parsing
I1 saw2 her3 duck4
PRP0,1 � (w1 = I) ⇤ (PRP ⇥ I)Xi,i+1 � (wi+1 = w) ⇤ (X ⇥ w)
PRP0,1
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
Parsing
I1 saw2 her3 duck4
Xi,i+1 � (wi+1 = w) ⇤ (X ⇥ w)
PRP0,1
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
Parsing
I1 saw2 her3 duck4
Xi,i+1 � (wi+1 = w) ⇤ (X ⇥ w)
PRP0,1
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR VBD1,2
Parsing
I1 saw2 her3 duck4
Xi,i+1 � (wi+1 = w) ⇤ (X ⇥ w)
PRP0,1
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR VBD1,2
PRP$2,3
Parsing
I1 saw2 her3 duck4
Xi,i+1 � (wi+1 = w) ⇤ (X ⇥ w)
PRP0,1
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR VBD1,2
PRP$2,3
PRP2,3
Parsing
I1 saw2 her3 duck4
Xi,i+1 � (wi+1 = w) ⇤ (X ⇥ w)
PRP0,1
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR VBD1,2
PRP$2,3
PRP2,3
NN3,4
VB3,4
Parsing
I1 saw2 her3 duck4
Xi,i+1 � (wi+1 = w) ⇤ (X ⇥ w)
PRP0,1
Xi,j � Yi,k ⇤ Zk,j ⇤ (X ⇥ Y Z)
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR VBD1,2
PRP$2,3
PRP2,3
NN3,4
VB3,4
Parsing
I1 saw2 her3 duck4
Xi,i+1 � (wi+1 = w) ⇤ (X ⇥ w)
PRP0,1
Xi,j � Yi,k ⇤ Zk,j ⇤ (X ⇥ Y Z)
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR VBD1,2
PRP$2,3
PRP2,3
NN3,4
VB3,4
Parsing
I1 saw2 her3 duck4
Xi,i+1 � (wi+1 = w) ⇤ (X ⇥ w)
PRP0,1
Xi,j � Yi,k ⇤ Zk,j ⇤ (X ⇥ Y Z)
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR VBD1,2
PRP$2,3
PRP2,3
NN3,4
VB3,4
Parsing
I1 saw2 her3 duck4
Xi,i+1 � (wi+1 = w) ⇤ (X ⇥ w)
PRP0,1
Xi,j � Yi,k ⇤ Zk,j ⇤ (X ⇥ Y Z)
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR VBD1,2
PRP$2,3
PRP2,3
NN3,4
VB3,4
Parsing
I1 saw2 her3 duck4
Xi,i+1 � (wi+1 = w) ⇤ (X ⇥ w)
PRP0,1
Xi,j � Yi,k ⇤ Zk,j ⇤ (X ⇥ Y Z)NP2,4 � PRP$2,3 ⇤NN3,4 ⇤ (NP⇥ PRP$ NN)
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR VBD1,2
PRP$2,3
PRP2,3
NN3,4
VB3,4
Parsing
I1 saw2 her3 duck4
Xi,i+1 � (wi+1 = w) ⇤ (X ⇥ w)
PRP0,1
Xi,j � Yi,k ⇤ Zk,j ⇤ (X ⇥ Y Z)NP2,4 � PRP$2,3 ⇤NN3,4 ⇤ (NP⇥ PRP$ NN)
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR VBD1,2
PRP$2,3
PRP2,3
NN3,4
VB3,4
NP2,4
Parsing
I1 saw2 her3 duck4
Xi,i+1 � (wi+1 = w) ⇤ (X ⇥ w)
PRP0,1
Xi,j � Yi,k ⇤ Zk,j ⇤ (X ⇥ Y Z)
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR VBD1,2
PRP$2,3
PRP2,3
NN3,4
VB3,4
NP2,4
Parsing
I1 saw2 her3 duck4
Xi,i+1 � (wi+1 = w) ⇤ (X ⇥ w)
PRP0,1
Xi,j � Yi,k ⇤ Zk,j ⇤ (X ⇥ Y Z)
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR VBD1,2
PRP$2,3
PRP2,3
NN3,4
VB3,4
NP2,4 SBAR2,4
Parsing
I1 saw2 her3 duck4
Xi,i+1 � (wi+1 = w) ⇤ (X ⇥ w)
PRP0,1
Xi,j � Yi,k ⇤ Zk,j ⇤ (X ⇥ Y Z)
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR VBD1,2
PRP$2,3
PRP2,3
NN3,4
VB3,4
NP2,4 SBAR2,4
VP1,4
Parsing
I1 saw2 her3 duck4
Xi,i+1 � (wi+1 = w) ⇤ (X ⇥ w)
PRP0,1
Xi,j � Yi,k ⇤ Zk,j ⇤ (X ⇥ Y Z)
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR VBD1,2
PRP$2,3
PRP2,3
NN3,4
VB3,4
NP2,4 SBAR2,4
VP1,4
Parsing
I1 saw2 her3 duck4
Xi,i+1 � (wi+1 = w) ⇤ (X ⇥ w)
PRP0,1
Xi,j � Yi,k ⇤ Zk,j ⇤ (X ⇥ Y Z)
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR VBD1,2
PRP$2,3
PRP2,3
NN3,4
VB3,4
NP2,4 SBAR2,4
VP1,4
S0,4
Parsing
I1 saw2 her3 duck4
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
Parsing
I1 saw2 her3 duck4
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
Parsing
I1 saw2 her3 duck4
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
I saw her duck
NP
VP
PRP VBD PRP$ NN
S
Parsing
I1 saw2 her3 duck4
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
I saw her duck
NP
VP
PRP VBD PRP$ NN
SNP
VP
PRP VBD PRP$ NN
S
yo vi su pato
Parsing
I1 saw2 her3 duck4
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
I saw her duck
SBAR
VP
PRP VBD PRP VB
S
Parsing
I1 saw2 her3 duck4
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
I saw her duck
SBAR
VP
PRP VBD PRP VB
SSBAR
VP
PRP VBD PRP VB
S
yo vi ella agacharse
Parsing
I1 saw2 her3 duck4
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
Analysis
Parsing
I1 saw2 her3 duck4
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
Analysis
nodesO(Nn2)
O(Gn3) edges
Wait a second!
Wait a second!
•Phrase-based MT is NP-hard because of permutations (there are a factorial number).
Wait a second!
•Phrase-based MT is NP-hard because of permutations (there are a factorial number).
•SCFGs also permute sentences.
Wait a second!
•Phrase-based MT is NP-hard because of permutations (there are a factorial number).
•SCFGs also permute sentences.
•But the decoding algorithm is polynomial…
Wait a second!
•Phrase-based MT is NP-hard because of permutations (there are a factorial number).
•SCFGs also permute sentences.
•But the decoding algorithm is polynomial…
•What are we giving up for this efficiency?
PermutationsX → X1 X2, X1 X2
X → X1 X2, X2 X1
X → a, a
X → b, bX → c, cX → d, d
PermutationsX → X1 X2, X1 X2
X → X1 X2, X2 X1
X → a, a
X → b, bX → c, cX → d, d
PermutationsX → X1 X2, X1 X2
X → X1 X2, X2 X1
X → a, a
What permutations of a b c d can this
grammar produce?
X → b, bX → c, cX → d, d
PermutationsX → X1 X2, X1 X2
X → X1 X2, X2 X1
X → a, a
b a c d
b a d c
b c a d
b c d a
b d a c
b d c a
What permutations of a b c d can this
grammar produce?
d a b c
d a c b
d b a c
d b c a
d c a b
d c b a
a b c d
a b d c
a c b d
a c d b
a d b c
a d c b
c a b d
c a d b
c b a d
c b d a
c d a b
c d b a
X → b, bX → c, cX → d, d
PermutationsX → X1 X2, X1 X2
X → X1 X2, X2 X1
X → a, a
b a c d
b a d c
b c a d
b c d a
b d a c
b d c a
What permutations of a b c d can this
grammar produce?
d a b c
d a c b
d b a c
d b c a
d c a b
d c b a
a b c d
a b d c
a c b d
a c d b
a d b c
a d c b
c a b d
c a d b
c b a d
c b d a
c d a b
c d b a
X → b, bX → c, cX → d, d
PermutationsX → X1 X2, X1 X2
X → X1 X2, X2 X1
X → a, a
b a c d
b a d c
b c a d
b c d a
b d a c
b d c a
What permutations of a b c d can this
grammar produce?
d a b c
d a c b
d b a c
d b c a
d c a b
d c b a
a b c d
a b d c
a c b d
a c d b
a d b c
a d c b
c a b d
c a d b
c b a d
c b d a
c d a b
c d b a
X → b, bX → c, cX → d, d
PermutationsX → X1 X2, X1 X2
X → X1 X2, X2 X1
X → a, a
b a c d
b a d c
b c a d
b c d a
b d a c
b d c a
What permutations of a b c d can this
grammar produce?
d a b c
d a c b
d b a c
d b c a
d c a b
d c b a
a b c d
a b d c
a c b d
a c d b
a d b c
a d c b
c a b d
c a d b
c b a d
c b d a
c d a b
c d b a
X → b, bX → c, cX → d, d
PermutationsX → X1 X2, X1 X2
X → X1 X2, X2 X1
X → a, a
b a c d
b a d c
b c a d
b c d a
b d a c
b d c a
What permutations of a b c d can this
grammar produce?
d a b c
d a c b
d b a c
d b c a
d c a b
d c b a
a b c d
a b d c
a c b d
a c d b
a d b c
a d c b
c a b d
c a d b
c b a d
c b d a
c d a b
c d b a
X → b, bX → c, cX → d, d
PermutationsX → X1 X2, X1 X2
X → X1 X2, X2 X1
X → a, a
b d a c
What permutations of a b c d can this
grammar produce?
c a d b
X → b, bX → c, cX → d, d
PermutationsX → X1 X2, X1 X2
X → X1 X2, X2 X1
X → a, a
b d a c
What permutations of a b c d can this
grammar produce?
c a d b
X → b, bX → c, cX → d, d
PermutationsX → X1 X2, X1 X2
X → X1 X2, X2 X1
X → a, a
b d a c
What permutations of a b c d can this
grammar produce?
c a d b
X → b, bX → c, cX → d, d
inside-outside alignments
PermutationsX → X1 X2, X1 X2
X → X1 X2, X2 X1
X → a, a
b d a c
What permutations of a b c d can this
grammar produce?
c a d b
X → b, bX → c, cX → d, d
inside-outside alignments
X → X1 X2 X3 X4, X2 X4 X1 X3 X → X1 X2 X3 X4, X3 X1 X4 X2
PermutationsX → X1 X2, X1 X2
X → X1 X2, X2 X1
X → a, a
b d a c
What permutations of a b c d can this
grammar produce?
c a d b
X → b, bX → c, cX → d, d
inside-outside alignments
X → X1 X2 X3 X4, X2 X4 X1 X3 X → X1 X2 X3 X4, X3 X1 X4 X2
No equivalent binary-branching SCFG
PermutationsX → X1 X2, X1 X2
X → X1 X2, X2 X1
X → a, a
b d a c
What permutations of a b c d can this
grammar produce?
c a d b
X → b, bX → c, cX → d, d
inside-outside alignments
X → X1 X2 X3 X4, X2 X4 X1 X3 X → X1 X2 X3 X4, X3 X1 X4 X2
No equivalent binary-branching SCFGComplexity of many problems is exponential in rank
Parsing as deductionXi,i+1 � (wi+1 = w) ⇤ (X ⇥ w)Xi,j � Yi,k ⇤ Zk,j ⇤ (X ⇥ Y Z)
[X ! w][wi+1 = w]
[i,X, i+ 1]
[X ! Y Z][i, Y, k][k, Z, j]
[i,X, j]
For sentence w1…wn, grammar G with nonterminals N
[i,X, j] 8i, j 2 0, ..., n,X 2 N
[X ! w] 8X ! w 2 PG
[wi = w] 8i 2 1, ..., naxioms:
items:inference rules:
[1, S, n]goal:
[X ! Y Z] 8X ! Y Z 2 PG
From proof to (pseudo)codeInput: w1…wn, grammar Gfor i in 1,…,n: for X->w_i in P(G): chart[i-1,X,i] := truefor l in 2,…,n: for i in 0,…,n-l: j := i+l for k in i+1,…,j-1: for X->YZ in P(G): if chart[i,Y,k] and chart[k,Z,j]: chart[i,X,j] := truereturn chart[0,S,n]
That’s nice, but…
•We need probabilities.
•We need to compute the most probable parse.
•We need to compute expectations.
Most probable parse
I1 saw2 her3 duck4
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
Most probable parse
I1 saw2 her3 duck4
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
(1.0)(1.0)(0.3)(0.7)(1.0)(1.0)(1.0)(1.0)(0.8)(0.2)(1.0)
0.7
Most probable parse
I1 saw2 her3 duck4
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
(1.0)(1.0)(0.3)(0.7)(1.0)(1.0)(1.0)(1.0)(0.8)(0.2)(1.0)
0.3 1.0
1.01.0
1.0
0.7
Most probable parse
I1 saw2 her3 duck4
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
(1.0)(1.0)(0.3)(0.7)(1.0)(1.0)(1.0)(1.0)(0.8)(0.2)(1.0)
1.0 0.3
0.3 1.0
1.01.0
1.0
0.7
Most probable parse
I1 saw2 her3 duck4
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
(1.0)(1.0)(0.3)(0.7)(1.0)(1.0)(1.0)(1.0)(0.8)(0.2)(1.0)
1.0 0.3
0.3 1.0
1.01.0
1.0
0.7
Most probable parse
I1
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
(1.0)(1.0)(0.3)(0.7)(1.0)(1.0)(1.0)(1.0)(0.8)(0.2)(1.0)
1.0 0.3
0.3 1.0
1.01.0
1.0
0.7
Most probable parse
I1
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
(1.0)(1.0)(0.3)(0.7)(1.0)(1.0)(1.0)(1.0)(0.8)(0.2)(1.0)
0.06
1.0 0.3
0.3 1.0
1.01.0
1.0
0.7
Most probable parse
I1
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
(1.0)(1.0)(0.3)(0.7)(1.0)(1.0)(1.0)(1.0)(0.8)(0.2)(1.0)
0.06
1.0 0.3
0.3 1.0
1.01.0
1.0
0.7
Most probable parse
I1
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
(1.0)(1.0)(0.3)(0.7)(1.0)(1.0)(1.0)(1.0)(0.8)(0.2)(1.0)
0.06
1.0 0.3
0.3 1.0
1.01.0
1.0
0.7
Most probable parse
I1
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
(1.0)(1.0)(0.3)(0.7)(1.0)(1.0)(1.0)(1.0)(0.8)(0.2)(1.0)
0.8
1.0 0.3
0.3 1.0
1.01.0
1.0
0.7
Most probable parse
I1 saw2 her3 duck4
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
(1.0)(1.0)(0.3)(0.7)(1.0)(1.0)(1.0)(1.0)(0.8)(0.2)(1.0)
0.8
1.0 0.3
0.3 1.0
1.01.0
1.0
0.7
Most probable parse
I1 saw2 her3 duck4
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
(1.0)(1.0)(0.3)(0.7)(1.0)(1.0)(1.0)(1.0)(0.8)(0.2)(1.0)
0.80.56
1.0 0.3
0.3 1.0
1.01.0
1.0
0.7
Most probable parse
I1 saw2 her3 duck4
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
0.80.56
Xi,j = max(Xi,j , Yi,k � Zk,j � p(X ⇥ Y Z))
1.0 0.3
0.3 1.0
1.01.0
1.0
0.7
Rule expectations
I1 saw2 her3 duck4
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
(1.0)(1.0)(0.3)(0.7)(1.0)(1.0)(1.0)(1.0)(0.8)(0.2)(1.0)
1.0 0.3
0.3 1.0
1.01.0
1.0
0.7
Rule expectations
I1
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
(1.0)(1.0)(0.3)(0.7)(1.0)(1.0)(1.0)(1.0)(0.8)(0.2)(1.0)
1.0 0.3
0.3 1.0
1.01.0
1.0
0.7
Rule expectations
I1
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
(1.0)(1.0)(0.3)(0.7)(1.0)(1.0)(1.0)(1.0)(0.8)(0.2)(1.0)
0.06
1.0 0.3
0.3 1.0
1.01.0
1.0
0.7
Rule expectations
I1
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
(1.0)(1.0)(0.3)(0.7)(1.0)(1.0)(1.0)(1.0)(0.8)(0.2)(1.0)
0.06
1.0 0.3
0.3 1.0
1.01.0
1.0
0.7
Rule expectations
I1
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
(1.0)(1.0)(0.3)(0.7)(1.0)(1.0)(1.0)(1.0)(0.8)(0.2)(1.0)
0.86
1.0 0.3
0.3 1.0
1.01.0
1.0
0.7
Rule expectations
I1 saw2 her3 duck4
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
(1.0)(1.0)(0.3)(0.7)(1.0)(1.0)(1.0)(1.0)(0.8)(0.2)(1.0)
0.86
1.0 0.3
0.3 1.0
1.01.0
1.0
0.7
Rule expectations
I1 saw2 her3 duck4
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
NN → duck
PRP → I
VBD → saw
PRP$ → her
NP → PRP$ NN
VP → VBD NP
S → PRP VP
PRP → her
VB → duckSBAR → PRP VB
VP → VBD SBAR
(1.0)(1.0)(0.3)(0.7)(1.0)(1.0)(1.0)(1.0)(0.8)(0.2)(1.0)
0.860.602
1.0 0.3
0.3 1.0
1.01.0
1.0
0.7
Rule expectations
I1 saw2 her3 duck4
PRP0,1
VBD1,2
VP1,4
PRP$2,3 NN3,4
NP2,4
PRP2,3 VB3,4
SBAR2,4
S0,4
0.860.602
Xi,j = Xi,j + (Yi,k � Zk,j � p(X ⇥ Y Z))
Similarities
Xi,j = Xi,j + (Yi,k � Zk,j � p(X ⇥ Y Z))
Xi,j = max(Xi,j , Yi,k � Zk,j � p(X ⇥ Y Z))
Xi,j � Yi,k ⇤ Zk,j ⇤ (X ⇥ Y Z)
Similarities
Xi,j = Xi,j + (Yi,k � Zk,j � p(X ⇥ Y Z))
Xi,j = max(Xi,j , Yi,k � Zk,j � p(X ⇥ Y Z))
Xi,j = Xi,j ⇤ (Yi,k ⇥ Zk,j ⇥ (X � Y Z))
Similarities
Xi,j = Xi,j + (Yi,k � Zk,j � p(X ⇥ Y Z))
Xi,j = max(Xi,j , Yi,k � Zk,j � p(X ⇥ Y Z))
Xi,j = Xi,j ⇤ (Yi,k ⇥ Zk,j ⇥ (X � Y Z))
h{T, F},_, F,^, T i
Similarities
Xi,j = Xi,j + (Yi,k � Zk,j � p(X ⇥ Y Z))
Xi,j = max(Xi,j , Yi,k � Zk,j � p(X ⇥ Y Z))
Xi,j = Xi,j ⇤ (Yi,k ⇥ Zk,j ⇥ (X � Y Z))
h{T, F},_, F,^, T i
hR+,max, 0,⇥, 1i
Similarities
Xi,j = Xi,j + (Yi,k � Zk,j � p(X ⇥ Y Z))
Xi,j = max(Xi,j , Yi,k � Zk,j � p(X ⇥ Y Z))
Xi,j = Xi,j ⇤ (Yi,k ⇥ Zk,j ⇥ (X � Y Z))
h{T, F},_, F,^, T i
hR+,max, 0,⇥, 1i
hR+,+, 0,⇥, 1i
Similarities
Xi,j = Xi,j + (Yi,k � Zk,j � p(X ⇥ Y Z))
Xi,j = max(Xi,j , Yi,k � Zk,j � p(X ⇥ Y Z))
Xi,j = Xi,j ⇤ (Yi,k ⇥ Zk,j ⇥ (X � Y Z))
Xi,j = Xi,j � (Yi,k ⇥ Zk,j ⇥R(X ⇤ Y Z))
h{T, F},_, F,^, T i
hR+,max, 0,⇥, 1i
hR+,+, 0,⇥, 1i
Similarities
Xi,j = Xi,j + (Yi,k � Zk,j � p(X ⇥ Y Z))
Xi,j = max(Xi,j , Yi,k � Zk,j � p(X ⇥ Y Z))
Xi,j = Xi,j ⇤ (Yi,k ⇥ Zk,j ⇥ (X � Y Z))
Xi,j = Xi,j � (Yi,k ⇥ Zk,j ⇥R(X ⇤ Y Z))
h{T, F},_, F,^, T i
hA,�,0,⌦,1i
hR+,max, 0,⇥, 1i
hR+,+, 0,⇥, 1i
Similarities
Xi,j = Xi,j + (Yi,k � Zk,j � p(X ⇥ Y Z))
Xi,j = max(Xi,j , Yi,k � Zk,j � p(X ⇥ Y Z))
Xi,j = Xi,j ⇤ (Yi,k ⇥ Zk,j ⇥ (X � Y Z))
Xi,j = Xi,j � (Yi,k ⇥ Zk,j ⇥R(X ⇤ Y Z))
h{T, F},_, F,^, T i
hA,�,0,⌦,1isemiring
hR+,max, 0,⇥, 1i
hR+,+, 0,⇥, 1i
Similarities
Xi,j = Xi,j + (Yi,k � Zk,j � p(X ⇥ Y Z))
Xi,j = max(Xi,j , Yi,k � Zk,j � p(X ⇥ Y Z))
Xi,j = Xi,j ⇤ (Yi,k ⇥ Zk,j ⇥ (X � Y Z))
Xi,j = Xi,j � (Yi,k ⇥ Zk,j ⇥R(X ⇤ Y Z))
boolean
tropical
inside
h{T, F},_, F,^, T i
hA,�,0,⌦,1isemiring
hR+,max, 0,⇥, 1i
hR+,+, 0,⇥, 1i
Parsing as weighted deductionFor sentence w1…wn, grammar G with nonterminals N
[i,X, j] 8i, j 2 0, ..., n,X 2 N
[X ! w] 8X ! w 2 PG
[wi = w] 8i 2 1, ..., naxioms:
items:inference rules:
[1, S, n]goal:
[X ! w] : u [wi+1 = w] : v
[i,X, i+ 1] : u⌦ v
[X ! Y Z] : u [i, Y, k] : v [k, Z, j] : y
[i,X, j] : u⌦ v ⌦ y
[X ! Y Z] 8X ! Y Z 2 PG
From proof to (pseudo)codeInput: w1…wn, grammar Gfor i in 1,…,n: for X->w_i in P(G): chart[i-1,X,i] := u(X->w_i)for l in 2,…,n: for i in 0,…,n-l: j := i+l for k in i+1,…,j-1: for X->YZ in P(G): chart[i,X,j] += chart[i,Y,k]*chart[k,Z,j]return chart[0,S,n]
Semiring parsing•Viterbi, inside, boolean (Goodman 1999)
•Expectation and variance semirings (Li & Eisner 2009)
•Feature expectations
•Minimum Bayes Risk
•Gradients, etc.
•minimum error upper envelope (Kumar et al. 2009)
Remaining questions
•How do we get the grammar?
•How do n-gram language models fit in?
•Is this really a plausible model of translation?
•Does it actually work? Why or why not?
Related Documents
