SlideShare a Scribd company logo
Tree-­‐Based	
  Machine	
  Transla0on	
  
	
  
Synchronous	
  Context-­‐Free	
  Grammar	
  
Introduced	
  by	
  Akiva	
  Miura,	
  AHC-­‐Lab	
  
2015/06/18	
  
15/06/18	
 2015©Akiva	
  Miura	
  	
  	
  AHC-­‐Lab,	
  IS,	
  NAIST	
 1	
MT	
  Study	
  Group	
  
Contents	
  
15/06/18	
 2015©Akiva	
  Miura	
  	
  	
  AHC-­‐Lab,	
  IS,	
  NAIST	
 2	
6.2	
  Synchronous	
  Context-­‐Free	
  Grammar	
  
	
  6.2.1	
  Characteris0cs	
  
	
  6.2.2	
  Training	
  
	
  6.2.3	
  Syntac0c	
  Labels	
  
	
  6.2.4	
  Features	
  
	
  6.2.5	
  Decoding
SCFG	
  
15/06/18	
 2015©Akiva	
  Miura	
  	
  	
  AHC-­‐Lab,	
  IS,	
  NAIST	
 3	
Synchronous	
  Context-­‐Free	
  Grammar	
  (SCFG):	
  
	
  
•  bilingual	
  extension	
  of	
  CFG	
  
•  can	
  be	
  applied	
  for	
  machine	
  transla0on	
  by	
  source	
  
language	
  side	
  parsing	
  (transducing)
Formalism	
  
15/06/18	
 2015©Akiva	
  Miura	
  	
  	
  AHC-­‐Lab,	
  IS,	
  NAIST	
 4	
SCFG	
  is	
  defined	
  as:	
  
	
  
where:	
  	
G = N Σ Δ R A
N
  
Σ
Δ
R
A
Rewrite	
  Rules	
  
15/06/18	
 2015©Akiva	
  Miura	
  	
  	
  AHC-­‐Lab,	
  IS,	
  NAIST	
 5	
* → α  β  φ ∈ R
,
α ( ) 1* 1 1
β ( ) 1* 1
φ : 1 1 1* 1 : α β
→ α  β  φ   
→ α  β
, 1 1 1 : α β
Rules	
  Example	
  
15/06/18	
 2015©Akiva	
  Miura	
  	
  	
  AHC-­‐Lab,	
  IS,	
  NAIST	
 6	
Example	
  of	
  rewrite	
  rules:	
  
	
  
S	
  →	
  <NP1	
  が	
  VP2,	
  NP1	
  VP2>	
  
VP	
  →	
  <NP1	
  を	
  V2,	
  V2	
  NP1>	
  
VP	
  →	
  <PP1	
  V2,	
  V2	
  PP1>	
  
VP	
  →	
  <NP1	
  V2,	
  V2	
  NP1>	
  
PP	
  →	
  <NP1	
  の	
  P2,	
  P2	
  NP1>	
  
NP	
  →	
  <NP1	
  の	
  NP2,	
  NP2	
  of	
  NP1>	
V	
  →	
  <開けた,	
  opened>	
  
	
  	
  	
  	
  	
  	
  	
  	
  |<座った,sat>	
  
P	
  →	
  <上に,	
  on>	
  
NP	
  →	
  <犬,	
  the	
  dog>	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  |	
  <ドア,	
  the	
  door>	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  |	
  <本,	
  the	
  book>	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  |	
  <上に,	
  the	
  upper>
Deriva0on	
  Example	
  
15/06/18	
 2015©Akiva	
  Miura	
  	
  	
  AHC-­‐Lab,	
  IS,	
  NAIST	
 7	
Example	
  of	
  deriva0on:	
  
	
  
<S1,	
  S1>	
  ⇒	
  <NP2	
  が VP3,	
  NP2	
  VP3>	
  
	
  	
  	
  ⇒	
  <犬 が	
  VP3,	
  the	
  dog	
  VP3>	
  
	
  	
  	
  ⇒	
  <犬 が	
  NP4	
  を	
  V5,	
  the	
  dog	
  V5	
  NP4>	
  
	
  	
  	
  ⇒	
  <犬 が ドア を V5,	
  the	
  dog	
  V5	
  the	
  door>	
  
	
  	
  	
  ⇒	
  <犬 が ドア を 開けた,	
  the	
  dog	
  opened	
  the	
  door>
Parse	
  Tree	
  Example	
  
15/06/18	
 2015©Akiva	
  Miura	
  	
  	
  AHC-­‐Lab,	
  IS,	
  NAIST	
 8	
Example	
  of	
  deriva0on	
  trees:	
  
	
  
犬	
NP2	
が	
 NP3	
S1	
NP4	
 を	
 V5	
ドア	
 開けた	
the dog	
NP2	
VP3	
S1	
V5	
 NP4	
opened	
 the door
Contents	
  
15/06/18	
 2015©Akiva	
  Miura	
  	
  	
  AHC-­‐Lab,	
  IS,	
  NAIST	
 9	
6.2	
  Synchronous	
  Context-­‐Free	
  Grammar	
  
	
  6.2.1	
  Characteris0cs	
  
	
  6.2.2	
  Training	
  
	
  6.2.3	
  Syntac0c	
  Labels	
  
	
  6.2.4	
  Features	
  
	
  6.2.5	
  Decoding
Normal	
  Form	
  
15/06/18	
 2015©Akiva	
  Miura	
  	
  	
  AHC-­‐Lab,	
  IS,	
  NAIST	
 10	
•  SCFG	
  has	
  almost	
  the	
  same	
  characteris0cs	
  with	
  CFG,	
  
but	
  does	
  not	
  have	
  normal	
  form	
  
Explana0on:	
  
	
  rank	
  :	
  #	
  of	
  non-­‐terminals	
  in	
  the	
  right	
  part	
  of	
  rule	
  
	
  binariza0on	
  :	
  conversion	
  of	
  rules	
  with	
  rank	
  >=	
  3	
  
	
   	
   	
  to	
  rules	
  with	
  rank	
  <=	
  2	
  
	
  
	
  Any	
  CFG	
  can	
  be	
  converted	
  to	
  Chomsky	
  Normal	
  Form,	
  
	
  but	
  SCFG	
  can’t	
  
Binariza0on	
  of	
  Rank-­‐3	
  Rules	
  
15/06/18	
 2015©Akiva	
  Miura	
  	
  	
  AHC-­‐Lab,	
  IS,	
  NAIST	
 11	
•  Any	
  Rank-­‐3	
  SCFG	
  rule	
  can	
  be	
  binarized:	
  
	
  
	
  e.g. 	
  X	
  →	
  <A1	
  B2	
  C3,	
  C3	
  B2	
  A1>	
  
	
  
	
  introducing	
  new	
  non-­‐terminal	
  X’	
  
	
   	
  X	
  →	
  <X’	
  1	
  C2,	
  C2	
  X’	
  1>	
  
	
   	
  X’	
  →	
  <A1	
  B2,	
  B2	
  A1>	
  
Binariza0on	
  of	
  Rank-­‐4	
  Rules	
  
15/06/18	
 2015©Akiva	
  Miura	
  	
  	
  AHC-­‐Lab,	
  IS,	
  NAIST	
 12	
•  Not	
  all	
  rank-­‐4	
  SCFG	
  rules	
  can	
  be	
  binarized:	
  
	
  e.g. 	
  X	
  →	
  <A1	
  B2	
  C3	
  D4,	
  C3	
  A1	
  D4	
  B2>	
  
	
   	
  X	
  →	
  <A1	
  B2	
  C3	
  D4,	
  B2	
  D4	
  A1	
  C3>	
  
	
  
	
  
A1
X
B2 C3 D4
C3 A1 D4 B2
X
A1
X
B2 C3 D4
B2 D4 A1 C3
X
these are called “inside-out”
Rela0on	
  of	
  Grammar	
  Ranks	
  
15/06/18	
 2015©Akiva	
  Miura	
  	
  	
  AHC-­‐Lab,	
  IS,	
  NAIST	
 13	
•  r-­‐CFG	
  is	
  set	
  of	
  languages	
  produced	
  by	
  rank-­‐r	
  rules	
  
•  Any	
  r-­‐CFG	
  can	
  be	
  converted	
  to	
  equivalent	
  2-­‐CFG	
  
Ø 1-­‐CFG	
  ⊊	
  2-­‐CFG	
  =	
  3-­‐CFG	
  =	
  4-­‐CFG	
  =	
  …	
  =	
  r-­‐CFG	
  
•  r-­‐SCFG	
  is	
  set	
  of	
  language	
  pairs	
  produced	
  by	
  rank-­‐r	
  rules	
  
•  3-­‐SCFG	
  can	
  be	
  converted	
  to	
  equivalent	
  2-­‐SCFG	
  
•  r-­‐SCFG	
  (r	
  ≧	
  4)	
  can	
  not	
  be	
  banarized	
  
Ø 1-­‐SCFG	
  ⊊	
  2-­‐SCFG	
  =	
  3-­‐SCFG	
  ⊊	
  4-­‐SCFG	
  ⊊	
  …	
  ⊊	
  r-­‐SCFG	
  
Contents	
  
15/06/18	
 2015©Akiva	
  Miura	
  	
  	
  AHC-­‐Lab,	
  IS,	
  NAIST	
 14	
6.2	
  Synchronous	
  Context-­‐Free	
  Grammar	
  
	
  6.2.1	
  Characteris0cs	
  
	
  6.2.2	
  Training	
  
	
  6.2.3	
  Syntac0c	
  Labels	
  
	
  6.2.4	
  Features	
  
	
  6.2.5	
  Decoding
Training	
  
15/06/18	
Automa0c	
  training	
  of	
  synchronous	
  rules:	
彼1	
 は2	
近	
  
い3	
う	
  
ち4	
 に5	
  
国	
  
会6	
 を7	
  
解	
  
散8	
  
す	
  
る9	
  
he1	
 ■	
will2	
disolve3	
 ■	
 ■	
the4	
 ■	
diet5	
 ■	
in6	
 ■	
the7	
near8	
 ■	
 ■	
future9	
 ■	
 ■	
Word Alignment	
近	
  
い3	
う	
  
ち4	
 に5	
  
国	
  
会6	
 を7	
  
解	
  
散8	
  
す	
  
る9	
  
disolve3	
 ■	
 ■	
the4	
 ■	
diet5	
 ■	
in6	
 ■	
the7	
near8	
 ■	
 ■	
future9	
 ■	
 ■	
X1	
   に5	
 X2	
  
解	
  
散8	
  
す	
  
る9	
  
dissolve3	
 ■	
 ■	
X2	
 ■	
in6	
 ■	
the7	
X1	
 ■	
Phrase Extraction ↑	
Synchronous Rule
Extraction →
Rule	
  Extrac0on	
  
15/06/18	
 2015©Akiva	
  Miura	
  	
  	
  AHC-­‐Lab,	
  IS,	
  NAIST	
 16	
These	
  rules	
  are	
  extracted	
  hierarchically,	
  
then	
  called	
  “Hierarchical	
  Phrases/Rules”	
  (Hiero)	
  
.21 2
,. 1 1 1
.,1 1
.1. R ← ∅
   2 . ∈ Φ R ← R ∪ →{ }
   2 1: → α β . . ∈ Φ       α = α α β = β β
R ← R ∪ → α α β β{ }
2
R = R
∈
∪
Rule	
  Restric0on	
  
15/06/18	
 2015©Akiva	
  Miura	
  	
  	
  AHC-­‐Lab,	
  IS,	
  NAIST	
 17	
•  Hierarchical  rule  extraction  method  is  exhaustive,  
then  the  trained  grammar  will  be  oversized  and  
very  ambiguous!
Ø  need  to  limit  the  rules:
•  minimal  phrase  pairs  for  the  same  alignment
•  span  length  limitation  (e.g.  2  ≦  length  ≦  10)
•  rule  length  limitation  (e.g.  length  ≦  5)
•  rank  of  rules  (rank  ≦  2)
•  prohibition  of  contiguous  non-‐‑‒terminals  (X1  X2)
•  including  at  least  1  word  alignment
Glue	
  Rules	
  
15/06/18	
 2015©Akiva	
  Miura	
  	
  	
  AHC-­‐Lab,	
  IS,	
  NAIST	
 18	
•  Because  of  the  span  length  limitation,  the  
grammars  might  be  impossible  to  cover  long  
sentences.
Ø  introducing  heuristically  initial  synchronous  rules  
called  “gleu  rules”:
S  →  <S1  X2,  S1  X2>
S  →  <X1,  X1>
•  for  long  distance  reordering  (such  as  En↔Ja),
we  can  introduce  also:
S  →  <S1  X2,  X2  S1>
Contents	
  
15/06/18	
 2015©Akiva	
  Miura	
  	
  	
  AHC-­‐Lab,	
  IS,	
  NAIST	
 19	
6.2	
  Synchronous	
  Context-­‐Free	
  Grammar	
  
	
  6.2.1	
  Characteris0cs	
  
	
  6.2.2	
  Training	
  
	
  6.2.3	
  Syntac0c	
  Labels	
  
	
  6.2.4	
  Features	
  
	
  6.2.5	
  Decoding
Syntac0c	
  Labels	
  
15/06/18	
 2015©Akiva	
  Miura	
  	
  	
  AHC-­‐Lab,	
  IS,	
  NAIST	
 20	
•  In	
  standard	
  Hiero	
  rules,	
  using	
  only	
  2	
  non-­‐terminals:	
  S,	
  X	
  
•  s0ll	
  very	
  ambiguous	
  (might	
  be	
  slow	
  and	
  inaccurate)	
  
Ø introducing	
  syntac0c	
  labels	
  from	
  parse	
  tree	
近	
  
い3	
う	
  
ち4	
 に5	
  
国	
  
会6	
 を7	
  
解	
  
散8	
  
す	
  
る9	
  
disolve3	
 ■	
 ■	
the4	
 ■	
diet5	
 ■	
in6	
 ■	
the7	
near8	
 ■	
 ■	
future9	
 ■	
 ■	
NP	
PP	
NP	
VP	
 IN+DT	
VP/PP	
VPVB
Contents	
  
15/06/18	
 2015©Akiva	
  Miura	
  	
  	
  AHC-­‐Lab,	
  IS,	
  NAIST	
 21	
6.2	
  Synchronous	
  Context-­‐Free	
  Grammar	
  
	
  6.2.1	
  Characteris0cs	
  
	
  6.2.2	
  Training	
  
	
  6.2.3	
  Syntac0c	
  Labels	
  
	
  6.2.4	
  Features	
  
	
  6.2.5	
  Decoding
Features	
  
15/06/18	
 2015©Akiva	
  Miura	
  	
  	
  AHC-­‐Lab,	
  IS,	
  NAIST	
 22	
•  Decoding	
  with	
  SCFG	
  also	
  uses	
  log	
  linear	
  model,	
  and	
  the	
  
features	
  are	
  almost	
  the	
  same	
  with	
  PBMT	
  
•  If	
  phrase	
  pairs	
  include	
  non-­‐terminals,	
  	
  count	
  of	
  phrases	
  is	
  
not	
  1	
  per	
  occurrence,	
  but	
  normalized	
  by	
  number	
  of	
  
matched	
  rules	
  
•  Addi0onal	
  penal0es:	
  
•  rule	
  count	
  penalty:	
  	
  
•  glue	
  rule	
  count	
  penalty:	
  	
  
   = −
   = − ∈ ∧ ∈ R{ }
Contents	
  
15/06/18	
 2015©Akiva	
  Miura	
  	
  	
  AHC-­‐Lab,	
  IS,	
  NAIST	
 23	
6.2	
  Synchronous	
  Context-­‐Free	
  Grammar	
  
	
  6.2.1	
  Characteris0cs	
  
	
  6.2.2	
  Training	
  
	
  6.2.3	
  Syntac0c	
  Labels	
  
	
  6.2.4	
  Features	
  
	
  6.2.5	
  Decoding
Decoding	
  
15/06/18	
 2015©Akiva	
  Miura	
  	
  	
  AHC-­‐Lab,	
  IS,	
  NAIST	
 24	
•  SCFG	
  decoding	
  maximizes	
  the	
  viterbi	
  deriva0on	
  with	
  
linear	
  combina0on	
  of	
  the	
  features:	
= () ,   
= () ,
',* ω ( )( )∑
',* ω ( )( )∑
≈ () ,
∈D   G    =    =
ω ( )
Transla0on	
  Forest	
  
15/06/18	
 2015©Akiva	
  Miura	
  	
  	
  AHC-­‐Lab,	
  IS,	
  NAIST	
 25	
•  Example	
  of	
  decoding:	
⽝犬0,1
が1,2
本2,3
の3,4
座った5,6
上に4,5
NP0,1 VP2,6
S0,6
PP2,5 NP2,5
NP2,3 P4,5
NP4,5
V5,6
the  dog
sat
the  upper
on
the  book
NP0,1
V5,6
NP4,5
P4,5
NP2,3
PP2,5
of
NP2,5
S0,6
↑	
  Source	
  language	
  side	
  syntax	
  parsing	
Target	
  language	
  side	
  transla0on	
  forest	
  ↑
End	
  Slide	
  
15/06/18	
 2015©Akiva	
  Miura	
  	
  	
  AHC-­‐Lab,	
  IS,	
  NAIST	
 26

More Related Content

Viewers also liked

Paris Fringe au Théâtre les Feux de la Rampe
Paris Fringe au Théâtre les Feux de la RampeParis Fringe au Théâtre les Feux de la Rampe
Paris Fringe au Théâtre les Feux de la Rampe
TheatreDesFeuxdelaRampe
 
Trên đường tìm ngọc
Trên đường tìm ngọcTrên đường tìm ngọc
Trên đường tìm ngọc
Thị Thanh Trần
 
O.M.Bugge - CV (2016)
O.M.Bugge - CV (2016)O.M.Bugge - CV (2016)
O.M.Bugge - CV (2016)
Capt. O.M.Bugge
 
Journal On LDO From IJEETC
Journal On LDO From IJEETCJournal On LDO From IJEETC
Journal On LDO From IJEETC
Sadanand Patil
 
O.m.bugge cv (2016)
O.m.bugge   cv (2016)O.m.bugge   cv (2016)
O.m.bugge cv (2016)
Capt. O.M.Bugge
 
構文情報に基づく機械翻訳のための能動学習手法と人手翻訳による評価
構文情報に基づく機械翻訳のための能動学習手法と人手翻訳による評価構文情報に基づく機械翻訳のための能動学習手法と人手翻訳による評価
構文情報に基づく機械翻訳のための能動学習手法と人手翻訳による評価
Akiva Miura
 
MOC Presentation
MOC PresentationMOC Presentation
MOC Presentation
Capt. O.M.Bugge
 
YouTube Channel Optimisations
YouTube Channel OptimisationsYouTube Channel Optimisations
YouTube Channel Optimisations
Claudio Iobbi
 
Resume of Sheikh Nayeem for Co
Resume of Sheikh Nayeem for CoResume of Sheikh Nayeem for Co
Resume of Sheikh Nayeem for Co
Sheikh Nayeem
 
Kansai MT Pivot Arekore
Kansai MT Pivot ArekoreKansai MT Pivot Arekore
Kansai MT Pivot Arekore
Akiva Miura
 
Contoh tugas simulasi
Contoh tugas simulasiContoh tugas simulasi
Contoh tugas simulasi
ghiovand
 
Teoremaluipitagora
TeoremaluipitagoraTeoremaluipitagora
Teoremaluipitagora
olimpiaanca
 

Viewers also liked (12)

Paris Fringe au Théâtre les Feux de la Rampe
Paris Fringe au Théâtre les Feux de la RampeParis Fringe au Théâtre les Feux de la Rampe
Paris Fringe au Théâtre les Feux de la Rampe
 
Trên đường tìm ngọc
Trên đường tìm ngọcTrên đường tìm ngọc
Trên đường tìm ngọc
 
O.M.Bugge - CV (2016)
O.M.Bugge - CV (2016)O.M.Bugge - CV (2016)
O.M.Bugge - CV (2016)
 
Journal On LDO From IJEETC
Journal On LDO From IJEETCJournal On LDO From IJEETC
Journal On LDO From IJEETC
 
O.m.bugge cv (2016)
O.m.bugge   cv (2016)O.m.bugge   cv (2016)
O.m.bugge cv (2016)
 
構文情報に基づく機械翻訳のための能動学習手法と人手翻訳による評価
構文情報に基づく機械翻訳のための能動学習手法と人手翻訳による評価構文情報に基づく機械翻訳のための能動学習手法と人手翻訳による評価
構文情報に基づく機械翻訳のための能動学習手法と人手翻訳による評価
 
MOC Presentation
MOC PresentationMOC Presentation
MOC Presentation
 
YouTube Channel Optimisations
YouTube Channel OptimisationsYouTube Channel Optimisations
YouTube Channel Optimisations
 
Resume of Sheikh Nayeem for Co
Resume of Sheikh Nayeem for CoResume of Sheikh Nayeem for Co
Resume of Sheikh Nayeem for Co
 
Kansai MT Pivot Arekore
Kansai MT Pivot ArekoreKansai MT Pivot Arekore
Kansai MT Pivot Arekore
 
Contoh tugas simulasi
Contoh tugas simulasiContoh tugas simulasi
Contoh tugas simulasi
 
Teoremaluipitagora
TeoremaluipitagoraTeoremaluipitagora
Teoremaluipitagora
 

Similar to MT Study SCFG

FPGAを用いた処理のロボット向けコンポーネントの設計生産性評価
FPGAを用いた処理のロボット向けコンポーネントの設計生産性評価FPGAを用いた処理のロボット向けコンポーネントの設計生産性評価
FPGAを用いた処理のロボット向けコンポーネントの設計生産性評価
Kazushi Yamashina
 
Experiences building a distributed shared log on RADOS - Noah Watkins
Experiences building a distributed shared log on RADOS - Noah WatkinsExperiences building a distributed shared log on RADOS - Noah Watkins
Experiences building a distributed shared log on RADOS - Noah Watkins
Ceph Community
 
One Problem, Two Structures, Six Solvers and Ten Years of Personnel Schedulin...
One Problem, Two Structures, Six Solvers and Ten Years of Personnel Schedulin...One Problem, Two Structures, Six Solvers and Ten Years of Personnel Schedulin...
One Problem, Two Structures, Six Solvers and Ten Years of Personnel Schedulin...
Pierre Schaus
 
FPGAの処理をソフトウェアコンポーネント化する設計ツールcReCompの高機能化の検討
FPGAの処理をソフトウェアコンポーネント化する設計ツールcReCompの高機能化の検討FPGAの処理をソフトウェアコンポーネント化する設計ツールcReCompの高機能化の検討
FPGAの処理をソフトウェアコンポーネント化する設計ツールcReCompの高機能化の検討
Kazushi Yamashina
 
Breaking the Nonsmooth Barrier: A Scalable Parallel Method for Composite Opti...
Breaking the Nonsmooth Barrier: A Scalable Parallel Method for Composite Opti...Breaking the Nonsmooth Barrier: A Scalable Parallel Method for Composite Opti...
Breaking the Nonsmooth Barrier: A Scalable Parallel Method for Composite Opti...
Fabian Pedregosa
 
Tree-based Translation Models (『機械翻訳』§6.2-6.3)
Tree-based Translation Models (『機械翻訳』§6.2-6.3)Tree-based Translation Models (『機械翻訳』§6.2-6.3)
Tree-based Translation Models (『機械翻訳』§6.2-6.3)
Yusuke Oda
 
Selection analysis using HyPhy
Selection analysis using HyPhySelection analysis using HyPhy
Metasepi team meeting: Ajhc Project Overview
Metasepi team meeting: Ajhc Project OverviewMetasepi team meeting: Ajhc Project Overview
Metasepi team meeting: Ajhc Project Overview
Kiwamu Okabe
 
NBIS RNA-seq course
NBIS RNA-seq courseNBIS RNA-seq course
NBIS RNA-seq course
Phil Ewels
 
RNA-Seq Analysis: Everything You Always Wanted to Know...and then some
RNA-Seq Analysis: Everything You Always Wanted to Know...and then someRNA-Seq Analysis: Everything You Always Wanted to Know...and then some
RNA-Seq Analysis: Everything You Always Wanted to Know...and then some
basepairtech
 
Network Measurement with P4 and C on Netronome Agilio
Network Measurement with P4 and C on Netronome AgilioNetwork Measurement with P4 and C on Netronome Agilio
Network Measurement with P4 and C on Netronome Agilio
Open-NFP
 
EIGRP
EIGRPEIGRP
EIGRP
newbie2019
 
Containerisation and Dynamic Frameworks in ICCMA’19
Containerisation and Dynamic Frameworks in ICCMA’19Containerisation and Dynamic Frameworks in ICCMA’19
Containerisation and Dynamic Frameworks in ICCMA’19
Carlo Taticchi
 
自律移動ロボット向けハード・ソフト協調のためのコンポーネント設計支援ツール
自律移動ロボット向けハード・ソフト協調のためのコンポーネント設計支援ツール自律移動ロボット向けハード・ソフト協調のためのコンポーネント設計支援ツール
自律移動ロボット向けハード・ソフト協調のためのコンポーネント設計支援ツール
Kazushi Yamashina
 
Improving Genetic Algorithm (GA) based NoC mapping algorithm using a formal ...
Improving Genetic Algorithm (GA)  based NoC mapping algorithm using a formal ...Improving Genetic Algorithm (GA)  based NoC mapping algorithm using a formal ...
Improving Genetic Algorithm (GA) based NoC mapping algorithm using a formal ...
Vinita Palaniveloo
 
Introduction to Apache Kafka
Introduction to Apache KafkaIntroduction to Apache Kafka
Introduction to Apache Kafka
Shiao-An Yuan
 
Sequencing, Alignment and Assembly
Sequencing, Alignment and AssemblySequencing, Alignment and Assembly
Sequencing, Alignment and Assembly
Shaun Jackman
 
control techniques
control techniquescontrol techniques
control techniques
Kranthi Kumar
 
p4srv6 (P4-16) design document rev1.0
p4srv6 (P4-16) design document rev1.0p4srv6 (P4-16) design document rev1.0
p4srv6 (P4-16) design document rev1.0
Kentaro Ebisawa
 
System design using HDL - Module 5
System design using HDL - Module 5System design using HDL - Module 5
System design using HDL - Module 5
Aravinda Koithyar
 

Similar to MT Study SCFG (20)

FPGAを用いた処理のロボット向けコンポーネントの設計生産性評価
FPGAを用いた処理のロボット向けコンポーネントの設計生産性評価FPGAを用いた処理のロボット向けコンポーネントの設計生産性評価
FPGAを用いた処理のロボット向けコンポーネントの設計生産性評価
 
Experiences building a distributed shared log on RADOS - Noah Watkins
Experiences building a distributed shared log on RADOS - Noah WatkinsExperiences building a distributed shared log on RADOS - Noah Watkins
Experiences building a distributed shared log on RADOS - Noah Watkins
 
One Problem, Two Structures, Six Solvers and Ten Years of Personnel Schedulin...
One Problem, Two Structures, Six Solvers and Ten Years of Personnel Schedulin...One Problem, Two Structures, Six Solvers and Ten Years of Personnel Schedulin...
One Problem, Two Structures, Six Solvers and Ten Years of Personnel Schedulin...
 
FPGAの処理をソフトウェアコンポーネント化する設計ツールcReCompの高機能化の検討
FPGAの処理をソフトウェアコンポーネント化する設計ツールcReCompの高機能化の検討FPGAの処理をソフトウェアコンポーネント化する設計ツールcReCompの高機能化の検討
FPGAの処理をソフトウェアコンポーネント化する設計ツールcReCompの高機能化の検討
 
Breaking the Nonsmooth Barrier: A Scalable Parallel Method for Composite Opti...
Breaking the Nonsmooth Barrier: A Scalable Parallel Method for Composite Opti...Breaking the Nonsmooth Barrier: A Scalable Parallel Method for Composite Opti...
Breaking the Nonsmooth Barrier: A Scalable Parallel Method for Composite Opti...
 
Tree-based Translation Models (『機械翻訳』§6.2-6.3)
Tree-based Translation Models (『機械翻訳』§6.2-6.3)Tree-based Translation Models (『機械翻訳』§6.2-6.3)
Tree-based Translation Models (『機械翻訳』§6.2-6.3)
 
Selection analysis using HyPhy
Selection analysis using HyPhySelection analysis using HyPhy
Selection analysis using HyPhy
 
Metasepi team meeting: Ajhc Project Overview
Metasepi team meeting: Ajhc Project OverviewMetasepi team meeting: Ajhc Project Overview
Metasepi team meeting: Ajhc Project Overview
 
NBIS RNA-seq course
NBIS RNA-seq courseNBIS RNA-seq course
NBIS RNA-seq course
 
RNA-Seq Analysis: Everything You Always Wanted to Know...and then some
RNA-Seq Analysis: Everything You Always Wanted to Know...and then someRNA-Seq Analysis: Everything You Always Wanted to Know...and then some
RNA-Seq Analysis: Everything You Always Wanted to Know...and then some
 
Network Measurement with P4 and C on Netronome Agilio
Network Measurement with P4 and C on Netronome AgilioNetwork Measurement with P4 and C on Netronome Agilio
Network Measurement with P4 and C on Netronome Agilio
 
EIGRP
EIGRPEIGRP
EIGRP
 
Containerisation and Dynamic Frameworks in ICCMA’19
Containerisation and Dynamic Frameworks in ICCMA’19Containerisation and Dynamic Frameworks in ICCMA’19
Containerisation and Dynamic Frameworks in ICCMA’19
 
自律移動ロボット向けハード・ソフト協調のためのコンポーネント設計支援ツール
自律移動ロボット向けハード・ソフト協調のためのコンポーネント設計支援ツール自律移動ロボット向けハード・ソフト協調のためのコンポーネント設計支援ツール
自律移動ロボット向けハード・ソフト協調のためのコンポーネント設計支援ツール
 
Improving Genetic Algorithm (GA) based NoC mapping algorithm using a formal ...
Improving Genetic Algorithm (GA)  based NoC mapping algorithm using a formal ...Improving Genetic Algorithm (GA)  based NoC mapping algorithm using a formal ...
Improving Genetic Algorithm (GA) based NoC mapping algorithm using a formal ...
 
Introduction to Apache Kafka
Introduction to Apache KafkaIntroduction to Apache Kafka
Introduction to Apache Kafka
 
Sequencing, Alignment and Assembly
Sequencing, Alignment and AssemblySequencing, Alignment and Assembly
Sequencing, Alignment and Assembly
 
control techniques
control techniquescontrol techniques
control techniques
 
p4srv6 (P4-16) design document rev1.0
p4srv6 (P4-16) design document rev1.0p4srv6 (P4-16) design document rev1.0
p4srv6 (P4-16) design document rev1.0
 
System design using HDL - Module 5
System design using HDL - Module 5System design using HDL - Module 5
System design using HDL - Module 5
 

Recently uploaded

DDS-Security 1.2 - What's New? Stronger security for long-running systems
DDS-Security 1.2 - What's New? Stronger security for long-running systemsDDS-Security 1.2 - What's New? Stronger security for long-running systems
DDS-Security 1.2 - What's New? Stronger security for long-running systems
Gerardo Pardo-Castellote
 
UI5con 2024 - Boost Your Development Experience with UI5 Tooling Extensions
UI5con 2024 - Boost Your Development Experience with UI5 Tooling ExtensionsUI5con 2024 - Boost Your Development Experience with UI5 Tooling Extensions
UI5con 2024 - Boost Your Development Experience with UI5 Tooling Extensions
Peter Muessig
 
Essentials of Automations: The Art of Triggers and Actions in FME
Essentials of Automations: The Art of Triggers and Actions in FMEEssentials of Automations: The Art of Triggers and Actions in FME
Essentials of Automations: The Art of Triggers and Actions in FME
Safe Software
 
What is Augmented Reality Image Tracking
What is Augmented Reality Image TrackingWhat is Augmented Reality Image Tracking
What is Augmented Reality Image Tracking
pavan998932
 
Vitthal Shirke Java Microservices Resume.pdf
Vitthal Shirke Java Microservices Resume.pdfVitthal Shirke Java Microservices Resume.pdf
Vitthal Shirke Java Microservices Resume.pdf
Vitthal Shirke
 
Neo4j - Product Vision and Knowledge Graphs - GraphSummit Paris
Neo4j - Product Vision and Knowledge Graphs - GraphSummit ParisNeo4j - Product Vision and Knowledge Graphs - GraphSummit Paris
Neo4j - Product Vision and Knowledge Graphs - GraphSummit Paris
Neo4j
 
GreenCode-A-VSCode-Plugin--Dario-Jurisic
GreenCode-A-VSCode-Plugin--Dario-JurisicGreenCode-A-VSCode-Plugin--Dario-Jurisic
GreenCode-A-VSCode-Plugin--Dario-Jurisic
Green Software Development
 
May Marketo Masterclass, London MUG May 22 2024.pdf
May Marketo Masterclass, London MUG May 22 2024.pdfMay Marketo Masterclass, London MUG May 22 2024.pdf
May Marketo Masterclass, London MUG May 22 2024.pdf
Adele Miller
 
原版定制美国纽约州立大学奥尔巴尼分校毕业证学位证书原版一模一样
原版定制美国纽约州立大学奥尔巴尼分校毕业证学位证书原版一模一样原版定制美国纽约州立大学奥尔巴尼分校毕业证学位证书原版一模一样
原版定制美国纽约州立大学奥尔巴尼分校毕业证学位证书原版一模一样
mz5nrf0n
 
LORRAINE ANDREI_LEQUIGAN_HOW TO USE WHATSAPP.pptx
LORRAINE ANDREI_LEQUIGAN_HOW TO USE WHATSAPP.pptxLORRAINE ANDREI_LEQUIGAN_HOW TO USE WHATSAPP.pptx
LORRAINE ANDREI_LEQUIGAN_HOW TO USE WHATSAPP.pptx
lorraineandreiamcidl
 
E-commerce Application Development Company.pdf
E-commerce Application Development Company.pdfE-commerce Application Development Company.pdf
E-commerce Application Development Company.pdf
Hornet Dynamics
 
GOING AOT WITH GRAALVM FOR SPRING BOOT (SPRING IO)
GOING AOT WITH GRAALVM FOR  SPRING BOOT (SPRING IO)GOING AOT WITH GRAALVM FOR  SPRING BOOT (SPRING IO)
GOING AOT WITH GRAALVM FOR SPRING BOOT (SPRING IO)
Alina Yurenko
 
Using Xen Hypervisor for Functional Safety
Using Xen Hypervisor for Functional SafetyUsing Xen Hypervisor for Functional Safety
Using Xen Hypervisor for Functional Safety
Ayan Halder
 
SWEBOK and Education at FUSE Okinawa 2024
SWEBOK and Education at FUSE Okinawa 2024SWEBOK and Education at FUSE Okinawa 2024
SWEBOK and Education at FUSE Okinawa 2024
Hironori Washizaki
 
GraphSummit Paris - The art of the possible with Graph Technology
GraphSummit Paris - The art of the possible with Graph TechnologyGraphSummit Paris - The art of the possible with Graph Technology
GraphSummit Paris - The art of the possible with Graph Technology
Neo4j
 
Transform Your Communication with Cloud-Based IVR Solutions
Transform Your Communication with Cloud-Based IVR SolutionsTransform Your Communication with Cloud-Based IVR Solutions
Transform Your Communication with Cloud-Based IVR Solutions
TheSMSPoint
 
Energy consumption of Database Management - Florina Jonuzi
Energy consumption of Database Management - Florina JonuziEnergy consumption of Database Management - Florina Jonuzi
Energy consumption of Database Management - Florina Jonuzi
Green Software Development
 
Revolutionizing Visual Effects Mastering AI Face Swaps.pdf
Revolutionizing Visual Effects Mastering AI Face Swaps.pdfRevolutionizing Visual Effects Mastering AI Face Swaps.pdf
Revolutionizing Visual Effects Mastering AI Face Swaps.pdf
Undress Baby
 
A Study of Variable-Role-based Feature Enrichment in Neural Models of Code
A Study of Variable-Role-based Feature Enrichment in Neural Models of CodeA Study of Variable-Role-based Feature Enrichment in Neural Models of Code
A Study of Variable-Role-based Feature Enrichment in Neural Models of Code
Aftab Hussain
 
socradar-q1-2024-aviation-industry-report.pdf
socradar-q1-2024-aviation-industry-report.pdfsocradar-q1-2024-aviation-industry-report.pdf
socradar-q1-2024-aviation-industry-report.pdf
SOCRadar
 

Recently uploaded (20)

DDS-Security 1.2 - What's New? Stronger security for long-running systems
DDS-Security 1.2 - What's New? Stronger security for long-running systemsDDS-Security 1.2 - What's New? Stronger security for long-running systems
DDS-Security 1.2 - What's New? Stronger security for long-running systems
 
UI5con 2024 - Boost Your Development Experience with UI5 Tooling Extensions
UI5con 2024 - Boost Your Development Experience with UI5 Tooling ExtensionsUI5con 2024 - Boost Your Development Experience with UI5 Tooling Extensions
UI5con 2024 - Boost Your Development Experience with UI5 Tooling Extensions
 
Essentials of Automations: The Art of Triggers and Actions in FME
Essentials of Automations: The Art of Triggers and Actions in FMEEssentials of Automations: The Art of Triggers and Actions in FME
Essentials of Automations: The Art of Triggers and Actions in FME
 
What is Augmented Reality Image Tracking
What is Augmented Reality Image TrackingWhat is Augmented Reality Image Tracking
What is Augmented Reality Image Tracking
 
Vitthal Shirke Java Microservices Resume.pdf
Vitthal Shirke Java Microservices Resume.pdfVitthal Shirke Java Microservices Resume.pdf
Vitthal Shirke Java Microservices Resume.pdf
 
Neo4j - Product Vision and Knowledge Graphs - GraphSummit Paris
Neo4j - Product Vision and Knowledge Graphs - GraphSummit ParisNeo4j - Product Vision and Knowledge Graphs - GraphSummit Paris
Neo4j - Product Vision and Knowledge Graphs - GraphSummit Paris
 
GreenCode-A-VSCode-Plugin--Dario-Jurisic
GreenCode-A-VSCode-Plugin--Dario-JurisicGreenCode-A-VSCode-Plugin--Dario-Jurisic
GreenCode-A-VSCode-Plugin--Dario-Jurisic
 
May Marketo Masterclass, London MUG May 22 2024.pdf
May Marketo Masterclass, London MUG May 22 2024.pdfMay Marketo Masterclass, London MUG May 22 2024.pdf
May Marketo Masterclass, London MUG May 22 2024.pdf
 
原版定制美国纽约州立大学奥尔巴尼分校毕业证学位证书原版一模一样
原版定制美国纽约州立大学奥尔巴尼分校毕业证学位证书原版一模一样原版定制美国纽约州立大学奥尔巴尼分校毕业证学位证书原版一模一样
原版定制美国纽约州立大学奥尔巴尼分校毕业证学位证书原版一模一样
 
LORRAINE ANDREI_LEQUIGAN_HOW TO USE WHATSAPP.pptx
LORRAINE ANDREI_LEQUIGAN_HOW TO USE WHATSAPP.pptxLORRAINE ANDREI_LEQUIGAN_HOW TO USE WHATSAPP.pptx
LORRAINE ANDREI_LEQUIGAN_HOW TO USE WHATSAPP.pptx
 
E-commerce Application Development Company.pdf
E-commerce Application Development Company.pdfE-commerce Application Development Company.pdf
E-commerce Application Development Company.pdf
 
GOING AOT WITH GRAALVM FOR SPRING BOOT (SPRING IO)
GOING AOT WITH GRAALVM FOR  SPRING BOOT (SPRING IO)GOING AOT WITH GRAALVM FOR  SPRING BOOT (SPRING IO)
GOING AOT WITH GRAALVM FOR SPRING BOOT (SPRING IO)
 
Using Xen Hypervisor for Functional Safety
Using Xen Hypervisor for Functional SafetyUsing Xen Hypervisor for Functional Safety
Using Xen Hypervisor for Functional Safety
 
SWEBOK and Education at FUSE Okinawa 2024
SWEBOK and Education at FUSE Okinawa 2024SWEBOK and Education at FUSE Okinawa 2024
SWEBOK and Education at FUSE Okinawa 2024
 
GraphSummit Paris - The art of the possible with Graph Technology
GraphSummit Paris - The art of the possible with Graph TechnologyGraphSummit Paris - The art of the possible with Graph Technology
GraphSummit Paris - The art of the possible with Graph Technology
 
Transform Your Communication with Cloud-Based IVR Solutions
Transform Your Communication with Cloud-Based IVR SolutionsTransform Your Communication with Cloud-Based IVR Solutions
Transform Your Communication with Cloud-Based IVR Solutions
 
Energy consumption of Database Management - Florina Jonuzi
Energy consumption of Database Management - Florina JonuziEnergy consumption of Database Management - Florina Jonuzi
Energy consumption of Database Management - Florina Jonuzi
 
Revolutionizing Visual Effects Mastering AI Face Swaps.pdf
Revolutionizing Visual Effects Mastering AI Face Swaps.pdfRevolutionizing Visual Effects Mastering AI Face Swaps.pdf
Revolutionizing Visual Effects Mastering AI Face Swaps.pdf
 
A Study of Variable-Role-based Feature Enrichment in Neural Models of Code
A Study of Variable-Role-based Feature Enrichment in Neural Models of CodeA Study of Variable-Role-based Feature Enrichment in Neural Models of Code
A Study of Variable-Role-based Feature Enrichment in Neural Models of Code
 
socradar-q1-2024-aviation-industry-report.pdf
socradar-q1-2024-aviation-industry-report.pdfsocradar-q1-2024-aviation-industry-report.pdf
socradar-q1-2024-aviation-industry-report.pdf
 

MT Study SCFG

  • 1. Tree-­‐Based  Machine  Transla0on     Synchronous  Context-­‐Free  Grammar   Introduced  by  Akiva  Miura,  AHC-­‐Lab   2015/06/18   15/06/18 2015©Akiva  Miura      AHC-­‐Lab,  IS,  NAIST 1 MT  Study  Group  
  • 2. Contents   15/06/18 2015©Akiva  Miura      AHC-­‐Lab,  IS,  NAIST 2 6.2  Synchronous  Context-­‐Free  Grammar    6.2.1  Characteris0cs    6.2.2  Training    6.2.3  Syntac0c  Labels    6.2.4  Features    6.2.5  Decoding
  • 3. SCFG   15/06/18 2015©Akiva  Miura      AHC-­‐Lab,  IS,  NAIST 3 Synchronous  Context-­‐Free  Grammar  (SCFG):     •  bilingual  extension  of  CFG   •  can  be  applied  for  machine  transla0on  by  source   language  side  parsing  (transducing)
  • 4. Formalism   15/06/18 2015©Akiva  Miura      AHC-­‐Lab,  IS,  NAIST 4 SCFG  is  defined  as:     where:   G = N Σ Δ R A N   Σ Δ R A
  • 5. Rewrite  Rules   15/06/18 2015©Akiva  Miura      AHC-­‐Lab,  IS,  NAIST 5 * → α  β  φ ∈ R , α ( ) 1* 1 1 β ( ) 1* 1 φ : 1 1 1* 1 : α β → α  β  φ   → α  β , 1 1 1 : α β
  • 6. Rules  Example   15/06/18 2015©Akiva  Miura      AHC-­‐Lab,  IS,  NAIST 6 Example  of  rewrite  rules:     S  →  <NP1  が  VP2,  NP1  VP2>   VP  →  <NP1  を  V2,  V2  NP1>   VP  →  <PP1  V2,  V2  PP1>   VP  →  <NP1  V2,  V2  NP1>   PP  →  <NP1  の  P2,  P2  NP1>   NP  →  <NP1  の  NP2,  NP2  of  NP1> V  →  <開けた,  opened>                  |<座った,sat>   P  →  <上に,  on>   NP  →  <犬,  the  dog>                        |  <ドア,  the  door>                        |  <本,  the  book>                        |  <上に,  the  upper>
  • 7. Deriva0on  Example   15/06/18 2015©Akiva  Miura      AHC-­‐Lab,  IS,  NAIST 7 Example  of  deriva0on:     <S1,  S1>  ⇒  <NP2  が VP3,  NP2  VP3>        ⇒  <犬 が  VP3,  the  dog  VP3>        ⇒  <犬 が  NP4  を  V5,  the  dog  V5  NP4>        ⇒  <犬 が ドア を V5,  the  dog  V5  the  door>        ⇒  <犬 が ドア を 開けた,  the  dog  opened  the  door>
  • 8. Parse  Tree  Example   15/06/18 2015©Akiva  Miura      AHC-­‐Lab,  IS,  NAIST 8 Example  of  deriva0on  trees:     犬 NP2 が NP3 S1 NP4 を V5 ドア 開けた the dog NP2 VP3 S1 V5 NP4 opened the door
  • 9. Contents   15/06/18 2015©Akiva  Miura      AHC-­‐Lab,  IS,  NAIST 9 6.2  Synchronous  Context-­‐Free  Grammar    6.2.1  Characteris0cs    6.2.2  Training    6.2.3  Syntac0c  Labels    6.2.4  Features    6.2.5  Decoding
  • 10. Normal  Form   15/06/18 2015©Akiva  Miura      AHC-­‐Lab,  IS,  NAIST 10 •  SCFG  has  almost  the  same  characteris0cs  with  CFG,   but  does  not  have  normal  form   Explana0on:    rank  :  #  of  non-­‐terminals  in  the  right  part  of  rule    binariza0on  :  conversion  of  rules  with  rank  >=  3        to  rules  with  rank  <=  2      Any  CFG  can  be  converted  to  Chomsky  Normal  Form,    but  SCFG  can’t  
  • 11. Binariza0on  of  Rank-­‐3  Rules   15/06/18 2015©Akiva  Miura      AHC-­‐Lab,  IS,  NAIST 11 •  Any  Rank-­‐3  SCFG  rule  can  be  binarized:      e.g.  X  →  <A1  B2  C3,  C3  B2  A1>      introducing  new  non-­‐terminal  X’      X  →  <X’  1  C2,  C2  X’  1>      X’  →  <A1  B2,  B2  A1>  
  • 12. Binariza0on  of  Rank-­‐4  Rules   15/06/18 2015©Akiva  Miura      AHC-­‐Lab,  IS,  NAIST 12 •  Not  all  rank-­‐4  SCFG  rules  can  be  binarized:    e.g.  X  →  <A1  B2  C3  D4,  C3  A1  D4  B2>      X  →  <A1  B2  C3  D4,  B2  D4  A1  C3>       A1 X B2 C3 D4 C3 A1 D4 B2 X A1 X B2 C3 D4 B2 D4 A1 C3 X these are called “inside-out”
  • 13. Rela0on  of  Grammar  Ranks   15/06/18 2015©Akiva  Miura      AHC-­‐Lab,  IS,  NAIST 13 •  r-­‐CFG  is  set  of  languages  produced  by  rank-­‐r  rules   •  Any  r-­‐CFG  can  be  converted  to  equivalent  2-­‐CFG   Ø 1-­‐CFG  ⊊  2-­‐CFG  =  3-­‐CFG  =  4-­‐CFG  =  …  =  r-­‐CFG   •  r-­‐SCFG  is  set  of  language  pairs  produced  by  rank-­‐r  rules   •  3-­‐SCFG  can  be  converted  to  equivalent  2-­‐SCFG   •  r-­‐SCFG  (r  ≧  4)  can  not  be  banarized   Ø 1-­‐SCFG  ⊊  2-­‐SCFG  =  3-­‐SCFG  ⊊  4-­‐SCFG  ⊊  …  ⊊  r-­‐SCFG  
  • 14. Contents   15/06/18 2015©Akiva  Miura      AHC-­‐Lab,  IS,  NAIST 14 6.2  Synchronous  Context-­‐Free  Grammar    6.2.1  Characteris0cs    6.2.2  Training    6.2.3  Syntac0c  Labels    6.2.4  Features    6.2.5  Decoding
  • 15. Training   15/06/18 Automa0c  training  of  synchronous  rules: 彼1 は2 近   い3 う   ち4 に5   国   会6 を7   解   散8   す   る9   he1 ■ will2 disolve3 ■ ■ the4 ■ diet5 ■ in6 ■ the7 near8 ■ ■ future9 ■ ■ Word Alignment 近   い3 う   ち4 に5   国   会6 を7   解   散8   す   る9   disolve3 ■ ■ the4 ■ diet5 ■ in6 ■ the7 near8 ■ ■ future9 ■ ■ X1   に5 X2   解   散8   す   る9   dissolve3 ■ ■ X2 ■ in6 ■ the7 X1 ■ Phrase Extraction ↑ Synchronous Rule Extraction →
  • 16. Rule  Extrac0on   15/06/18 2015©Akiva  Miura      AHC-­‐Lab,  IS,  NAIST 16 These  rules  are  extracted  hierarchically,   then  called  “Hierarchical  Phrases/Rules”  (Hiero)   .21 2 ,. 1 1 1 .,1 1 .1. R ← ∅   2 . ∈ Φ R ← R ∪ →{ }   2 1: → α β . . ∈ Φ     α = α α β = β β R ← R ∪ → α α β β{ } 2 R = R ∈ ∪
  • 17. Rule  Restric0on   15/06/18 2015©Akiva  Miura      AHC-­‐Lab,  IS,  NAIST 17 •  Hierarchical  rule  extraction  method  is  exhaustive,   then  the  trained  grammar  will  be  oversized  and   very  ambiguous! Ø  need  to  limit  the  rules: •  minimal  phrase  pairs  for  the  same  alignment •  span  length  limitation  (e.g.  2  ≦  length  ≦  10) •  rule  length  limitation  (e.g.  length  ≦  5) •  rank  of  rules  (rank  ≦  2) •  prohibition  of  contiguous  non-‐‑‒terminals  (X1  X2) •  including  at  least  1  word  alignment
  • 18. Glue  Rules   15/06/18 2015©Akiva  Miura      AHC-­‐Lab,  IS,  NAIST 18 •  Because  of  the  span  length  limitation,  the   grammars  might  be  impossible  to  cover  long   sentences. Ø  introducing  heuristically  initial  synchronous  rules   called  “gleu  rules”: S  →  <S1  X2,  S1  X2> S  →  <X1,  X1> •  for  long  distance  reordering  (such  as  En↔Ja), we  can  introduce  also: S  →  <S1  X2,  X2  S1>
  • 19. Contents   15/06/18 2015©Akiva  Miura      AHC-­‐Lab,  IS,  NAIST 19 6.2  Synchronous  Context-­‐Free  Grammar    6.2.1  Characteris0cs    6.2.2  Training    6.2.3  Syntac0c  Labels    6.2.4  Features    6.2.5  Decoding
  • 20. Syntac0c  Labels   15/06/18 2015©Akiva  Miura      AHC-­‐Lab,  IS,  NAIST 20 •  In  standard  Hiero  rules,  using  only  2  non-­‐terminals:  S,  X   •  s0ll  very  ambiguous  (might  be  slow  and  inaccurate)   Ø introducing  syntac0c  labels  from  parse  tree 近   い3 う   ち4 に5   国   会6 を7   解   散8   す   る9   disolve3 ■ ■ the4 ■ diet5 ■ in6 ■ the7 near8 ■ ■ future9 ■ ■ NP PP NP VP IN+DT VP/PP VPVB
  • 21. Contents   15/06/18 2015©Akiva  Miura      AHC-­‐Lab,  IS,  NAIST 21 6.2  Synchronous  Context-­‐Free  Grammar    6.2.1  Characteris0cs    6.2.2  Training    6.2.3  Syntac0c  Labels    6.2.4  Features    6.2.5  Decoding
  • 22. Features   15/06/18 2015©Akiva  Miura      AHC-­‐Lab,  IS,  NAIST 22 •  Decoding  with  SCFG  also  uses  log  linear  model,  and  the   features  are  almost  the  same  with  PBMT   •  If  phrase  pairs  include  non-­‐terminals,    count  of  phrases  is   not  1  per  occurrence,  but  normalized  by  number  of   matched  rules   •  Addi0onal  penal0es:   •  rule  count  penalty:     •  glue  rule  count  penalty:       = −   = − ∈ ∧ ∈ R{ }
  • 23. Contents   15/06/18 2015©Akiva  Miura      AHC-­‐Lab,  IS,  NAIST 23 6.2  Synchronous  Context-­‐Free  Grammar    6.2.1  Characteris0cs    6.2.2  Training    6.2.3  Syntac0c  Labels    6.2.4  Features    6.2.5  Decoding
  • 24. Decoding   15/06/18 2015©Akiva  Miura      AHC-­‐Lab,  IS,  NAIST 24 •  SCFG  decoding  maximizes  the  viterbi  deriva0on  with   linear  combina0on  of  the  features: = () ,   = () , ',* ω ( )( )∑ ',* ω ( )( )∑ ≈ () , ∈D  G   =   = ω ( )
  • 25. Transla0on  Forest   15/06/18 2015©Akiva  Miura      AHC-­‐Lab,  IS,  NAIST 25 •  Example  of  decoding: ⽝犬0,1 が1,2 本2,3 の3,4 座った5,6 上に4,5 NP0,1 VP2,6 S0,6 PP2,5 NP2,5 NP2,3 P4,5 NP4,5 V5,6 the  dog sat the  upper on the  book NP0,1 V5,6 NP4,5 P4,5 NP2,3 PP2,5 of NP2,5 S0,6 ↑  Source  language  side  syntax  parsing Target  language  side  transla0on  forest  ↑
  • 26. End  Slide   15/06/18 2015©Akiva  Miura      AHC-­‐Lab,  IS,  NAIST 26