Non-contiguity phrase-based SMT

1
Non-contiguity phrase-based SMT

• Hendra Setiawan

2
Outline

• A brief introduction to SMT
• Non-contiguity for SMT
• Near-future work

3
Introduction to SMT
f damoy Pat Riley rano pashol
e Pat Riley went home early
Alignment describes translation
4
Introduction to SMT

• Translation process
• Lexical mapping
• Reordering

(at word level)
(at word level)
5
Introduction to SMT

• Translation process (state-of-the-art)
• Lexical mapping
• (at phrase level)
• Reordering
• (at phrase level)

damoy Pat_Riley rano pashol
home Pat_Riley early went
damoy Pat_Riley rano pashol
Pat_Riley went home early
6
Outline

• A brief introduction to SMT
• Non-contiguity for SMT
• Near-future work

7
Contiguity to Non-contiguity (1)

• Lexical mapping (training)
• Can translate Pat_Riley to Pat_Riley
• but not not to ne pas
• Impose contiguity constraint when extracting
  Pat_Riley.
• Given automatic alignment only 40 coverage
• (max windows length 7)

e1 e2 ... el
fm ... . f2 f1
8
Contiguity to Non-contiguity (2)

• Relax the contiguity constraint
• Increase coverage from 40 to 53

e1 e2 ... el
fm ... . f2 f1
9
Contiguity to Non-contiguity (3)

• Reordering

10
Model Comparison
Pharaoh PLMM(ef) x PLMM(fe)
x Plm(e) x PD(ef)
Lexical Mapping
Reordering
11
Some close testing result
BLEU
Model combination
12
Outline

• A brief introduction to SMT
• Non-contiguity for SMT
• Near-future work

13
Near-future work

• Parameter estimation and backoff strategy for
  bigram model
• Perform open testing and analyze the result